MacroPlacement

Our Progress: A Chronology

Table of Contents

• Introduction
• Our progress and major milestones
  • Publicly available commercial SP&R flow
  • Ariane133 macro placement using Circuit Training
  • Replication of proxy cost
  • NVDLA macro placement using Circuit Training
  • BlackParrot (Quad Core) macro placement using Circuit Training
  • Circuit Training stability study
  • Proxy cost correlation study
  • MemPool Group macro placement using Circuit Training
  • Macro placement on GLOBALFOUNDRIES 12 nm enablement
  • Macro placement using AutoDMP
  • Macro placement generated using Simulated Annealing
  • Macro placement by Human expert
  • Macro placement using Hier-RTLMP
• Pinned questions

Introduction

MacroPlacement is an open, transparent effort to provide a public, baseline implementation of Google Brain’s Circuit Training (Morpheus) deep RL-based placement method. In this repo, we aim to achieve the following.

• We want to enable anyone to perform RL-based macro placement on their own design, starting from design RTL files.
• We want to enable anyone to train their own RL models based on their own designs in any design enablements, starting from design RTL files.
• We want to demystify important aspects of the Google Nature paper, including aspects unavailable in Circuit Training and aspects where the Nature paper and Circuit Training clearly diverge, in order to help researchers and users better understand the methodology.
• We want to apply learnings from the community’s collective experiences with the Google Brain team’s arXiv result, Nature paper and Circuit Training repo – and demonstrate how communication of research results might be improved in our community going forward. A clear theme from the past months’ experience: “There is no substitute for source code.”

In order to achieve the above goals, our initial focus has been on the following efforts.

• Generating correct inputs and setup for Circuit Training. Since Circuit Training uses protocol buffer format to represent designs, we must translate standard LEF/DEF representation to the protocol buffer format. We must also determine how to correctly feed all necessary design information into the Google Brain’s Circuit Training flow, e.g., halo width, canvas size, and constraints. If we accomplish this, then we can run Google Brain’s Circuit Training to train our own RL models or perform RL-based macro placement for our own designs.
• Replicating important but missing parts of the Google Nature paper. Several aspects of Circuit Training are not clearly documented in the Nature paper, nor in the code and scripts that are visible in Circuit Training. Over time, these have included hypergraph-to-graph conversion; gridding, grouping and clustering; force-directed placement; various hyperparameter settings; and more. As we keep moving forward, based on our experiments and continued Q&A and feedback from Google, we will summarize the miscorrelations between the Google Nature paper and Google Brain’s Circuit Training, as well as corrective steps. In this way, the Circuit Training methodology and the results published in the Nature paper can be better understood by all.

Our Progress

June 6 - Aug 5: We have developed and made publicly available the SP&R flow using commercial tools Cadence Genus and Innovus, and open-source tools Yosys and OpenROAD, for Ariane (two variants – one with 136 SRAMs and another with 133 SRAMs), MemPool tile and NVDLA designs on NanGate45, ASAP7 and SKY130HD open enablement. We applaud and thank Cadence Design Systems for allowing their tool runscripts to be shared openly by researchers, enabling reproducibility of results obtained via use of Cadence tools. This was an important milestone for the EDA research community. Please see Dr. David Junkin’s presentation at the recent DAC-2022 “Open-Source EDA and Benchmarking Summit” birds-of-a-feather meeting.

The following describes our learning related to testcase generation and its implementation using different tools on different platforms.

1. The Google Nature paper uses the Ariane testcase (contains 133 256x16-bit SRAMs) for their experiment. Here we show that just instantiating 256x16 bit SRAMs results in 136 SRAMs in the synthesized netlist. Based on our investigations, we have provided the detailed steps to convert the Ariane design with 136 SRAMs to a Ariane design with 133 SRAMs.
2. We provide the required SRAM lef, lib along with the description to reproduce the provided SRAMs or generate a new SRAM for each enablement.
3. The SKY130HD enablement has only five metal layers, while SRAMs have routing up through the M4 layer. This causes P&R failure due to very high routing congestion. We therefore developed FakeStack-extended P&R enablement, where we replicate the first four metal layers to generate a nine metal layer enablement. We call this SKY130HD-FakeStack and have used it to implement our testcases. We also provide a script for researchers to generate FakeStack enablements with different configurations.
4. We provide power grid generation scripts for Cadence Innovus. During the power grid (PG) generation process we made sure the routing resource used by the PG is in the range of ~20%, matching the guidance given in Circuit Training.
5. Also we provide an Innovus Tcl script to extract the metrics reported in Table 1 of “A graph placement methodology for fast chip design”, at three stages of the post-floorplanning P&R flow, i.e., pre-CTS, post-CTSOpt, and post-RouteOpt (final). This script is included in the P&R flow. The extracted metrics for all of our designs, on different enablements, are available here.

June 10: grouper.py was released in CircuitTraining. This revealed that protobuf input to the hypergraph clustering into soft macros included the (x,y) locations of the nodes. (A grouper.py script had been shown to Prof. Kahng during a meeting at Google on May 19.) The use of (x,y) locations from a physical synthesis tool was very unexpected, since it is not mentioned in “Methods” or other descriptions given in the Nature paper. We raised issue #25 to get clarification about this. [July 10: The README added to the grouping area of CircuitTraining confirmed that the input netlist has all of its nodes already placed.]

We currently use the physical synthesis tool Cadence Genus iSpatial to obtain (x,y) placed locations per instance as part of the input to Grouping. The Genus iSpatial post-physical-synthesis netlist is the starting point for how we produce the clustered netlist and the *.plc file which we provide as open inputs to CircuitTraining. From post-physical-synthesis netlist to clustered netlist generation can be divided into the following steps, which we have implemented as open-source in our CodeElements area:

1. June 6: Gridding determines a dissection of the layout canvas into some number of rows and some number of columns of gridcells.
2. June 10: Grouping groups closely-related logic with the corresponding hard macros and clumps of IOs.
3. June 12: Clustering clusters of millions of standard cells into a few thousand clusters (soft macros).

June 22: We added our flow-scripts that run our gridding, grouping and clustering implementations to generate a final clustered netlist in protocol buffer format. Google’s netlist protocol buffer format documentation available in the CircuitTraining repo was very helpful to our understanding of how to convert a placed netlist to protobuf format. Our scripts enable clustered netlists in protobuf format to be produced from placed netlists in either LEF/DEF or Bookshelf format.

July 12: As stated in the “What is your timeline?” FAQ response [see also note [5] here], we presented progress to date in this MacroPlacement talk at the DAC-2022 “Open-Source EDA and Benchmarking Summit” birds-of-a-feather meeting.

July 26: Replication of the wirelength component of proxy cost. The wirelength is similar to HPWL where given a netlist, we take the width and height and sum them up for each net. One caveat is that for soft macro pins, there could be a weight factor which implies the total connections between the source and sink pins. If not defined, the default value is 1. This weight factor needs to be multiplied with the sum of width and height to replicate Google’s API. We provide the following table as a comparison between our implementations and Google’s API.

Testcase	Notes	Canvas width/height	Grid col/row	Google	Our
Ariane	Google’s Ariane	356.592 / 356.640	35 / 33	0.7500626080261634	0.7500626224300161
Ariane133	From MacroPlacement	1599.99 / 1598.8	50 / 50	0.6522555375409593	0.6522555172428797

July 31: The netlist protocol buffer format documentation also helped us to write this Innovus-based tcl script which converts physical synthesized netlist to protobuf format in Innovus. [This script was written and developed by ABKGroup students at UCSD. However, the underlying commands and reports are copyrighted by Cadence. We thank Cadence for granting permission to share our research to help promote and foster the next generation of innovators.] We use this post-physical-synthesis protobuf netlist as input to the grouping code to generate the clustered netlist. Fixes that we made while running Google’s grouping code resulted in this [08/01/2022] pull request. [08/05/2022: Google’s grouping code has been updated based on this PR.]

July 22-August 4: We shared with Google engineers our (flat) post-physical-synthesis-protobuf netlist (ariane.pb.txt) of our Ariane design with 133 SRAMs on the NanGate45 platform, along with the corresponding clustered netlist and the legalized.plc file (clustered netlist: netlist.pb.txt) generated using the CircuitTraining grouping code. The goal here was to verify our steps and setup up to this point. Also, we provide scripts (using both our CodeElements and CT-grouping) to integrate the clustered netlist generation with the SP&R flow.

August 5: The following table compares the clustering results for Ariane133-NG45 design generated by the Google engineer (internally to Google) and the clustering results generated by us using CT grouping code.

	Google Internal flow (from Google)	Our use of CT Grouping code
Number of grid rows x columns	21 x 24	21 x 24
Number of soft macros	736	738
HPWL	4171594.811	4179069.884
Wirelength cost	0.072595	0.072197
Congestion cost	0.727798	0.72853

August 11: We received information from Google that when a standard cell has multiple outputs, it merges all of them in the protobuf netlist (example: a full adder cell would have its outputs merged). The possible vertices of a hyperedge are macro pins, ports, and standard cells. Our Innovus-based protobuf netlist generation tcl script takes care of this.

August 15: We received information from Google engineers that in the proxy cost function, the density weight is set to 0.5 for their internal runs.

August 17: The proxy wirelength cost which is usually a value between 0 and 1, is related to the HPWL we computed earlier. We deduce the formulation as the following:

|netlist| is the total number of nets and it takes into account the weight factor defined on soft macro pins. Here is our proxy wirelength compared with Google’s API:

Testcase	Notes	Canvas width/height	Google	Our
Ariane	Google’s Ariane	356.592 / 356.640	0.05018661999974192	0.05018662006439473
Ariane133	From MacroPlacement	1599.99 / 1598.8	0.04456188308735019	0.04456188299072617

Replication of the density component of proxy cost. We now have a verified density cost computation. Density cost computation depends on gridcell density. Gridcell density is the ratio of the total area occupied by standard cells, soft macros and hard macros to the total area of the grid. If there are cell overlaps then it may result in grid density greater than one. To get the density cost, we take the average of the top 10% of the densest gridcells. Before outputting it, we multiply it by 0.5. Notice that this 0.5 is not the “weight” of this cost function, but simply another factor applied besides the weight factor from the cost function.

Testcase	Notes	Canvas width/height	Grid col/row	Google	Our
Ariane	Google’s Ariane	356.592 / 356.640	35 / 33	0.7500626080261634	0.7500626224300161
Ariane133	From MacroPlacement	1599.99 / 1598.8	50 / 50	0.6522555375409593	0.6522555172428797

August 18: The flat post-physical-synthesis protobuf netlist of Ariane133-NanGate45 design is used as input to CT grouping code to generate the clustered netlist. We then use this clustered netlist in Circuit Training. Coordinate Descent is (by default) not applied to any macro placement solution. Here is the link to our tensorboard. We ran Innovus P&R starting from the macro placement generated using CT, through the end of detailed routing (RouteOpt) and collection of final PPA / “Table 1” metrics. Following are the metrics and screen shots of the P&R database. Throughout the SP&R flow, the target clock period is 4ns. The power grid overhead is 18.46% in the actual P&R setup, matching the 18% mentioned in the Circuit Training repo. All results are for DRC-clean final routing produced by the Innovus tool.
[In the immediately-following content, we also show comparison results using other macro placement methods, collected since August 18.]
[As of August 24 onward, we refer to this testcase as “Our Ariane133-NanGate45_51” since it has 51% area utilization. A second testcase, “Our Ariane133-NanGate45_68”, has 68% area utilization which exactly matches that of the Ariane in Circuit Training.]

Circuit Training Baseline Result on “Our Ariane133-NanGate45_51”.

Macro placement generated by Circuit Training on Our Ariane-133 (NG45), with post-macro placement flow using Innovus21.1
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	2560080	214555	1018356	287.79	4343214	0.005	0	0.01%	0.02%
postCTS	2560080	216061	1018356	301.31	4345969	0.010	0	0.01%	0.02%
postRoute	2560080	216061	1018356	300.38	4463660	0.359	0

Comparison 1: “Human Gridded”. For comparison, a baseline “human, gridded” macro placement was generated by a human for the same canvas size, I/O placement and gridding, with results as follows.

Macro placement generated by a human on Our Ariane-133 (NG45), with post-macro placement flow using Innovus21.1
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	2560080	215188.9	1018356	285.96	4470832	-0.002	-0.005	0.00%	0.00%
postCTS	2560080	216322.9	1018356	299.62	4472866	0.001	0	0.00%	0.00%
postRoute	2560080	216322.9	1018356	298.60	4587141	0.284	0

Comparison 2: RePlAce. The standalone RePlAce placer was run on the same (flat) netlist with the same canvas size and I/O placement, with results as follows.

Macro placement generated by RePlAce (standalone, from HERE) on Our Ariane-133 (NG45), with post-macro placement flow using Innovus21.1
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	2560080	214910.71	1018356	288.654	4178509	0.003	0	0.03%	0.07%
postCTS	2560080	216006.63	1018356	302.013	4184690	0.007	0	0.05%	0.08%
postRoute	2560080	216006.63	1018356	301.260	4315157	-0.207	-0.41

Comparison 3: RTL-MP. The RTL-MP macro placer described in this ISPD-2022 paper and used as the default macro placer in OpenROAD was run on the same (flat) netlist with the same canvas size and I/O placement, with results as follows.

Macro placement generated using RTL-MP on Our Ariane-133 (NG45), with post-macro placement flow using Innovus21.1
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	2560080	216420.26	1018356	289.435	5164199	0.020	0	0.04%	0.05%
postCTS	2560080	217938.32	1018356	303.757	5185004	0.001	0	0.05%	0.07%
postRoute	2560080	217938.32	1018356	302.844	5306735	0.104	0

Comparison 4: The Hier-RTLMP macro placer was run on the same (flat) netlist with the same canvas size and I/O placement, with results as follows. [The Hier-RTLMP paper is in submission as of August 2022; availability in OpenROAD and OpenROAD-flow-scripts is planned by end of September 2022. Please email abk@eng.ucsd.edu if you would like a preprint, not for further redistribution.]

Macro placement generated using Hier-RTLMP on Our Ariane-133 (NG45), with post-macro placement flow using Innovus21.1
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	2560080	214783.83	1018356	288.356	4397005	0.005	0	0.02%	0.05%
postCTS	2560080	215911.67	1018356	302.176	4419305	0.009	0	0.04%	0.06%
postRoute	2560080	215911.67	1018356	301.468	4537458	0.311	0

August 20: Matching the area utilization. We revisited the area utilization of Our Ariane133 and realized that it (51%) is lower than that of Google’s Ariane (68%). So that this would not devalue our study, we created a second variant, “Our Ariane133-NanGate45_68”, which matches the area utilization of Google’s Ariane. Results are as given below.

Circuit Training Baseline Result on “Our Ariane133-NanGate45_68”.

Macro Placement generated Using CT (Ariane 68% Utilization)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	215575.444	1018355.73	288.762	4170253	0.002	0	0.01%	0.01%
postCTS	1814274	217114.520	1018355.73	302.607	4186888	0.001	0	0.00%	0.01%
postRoute	1814274	217114.520	1018355.73	301.722	4295572	0.336	0

Comparison 1: “Human Gridded”. For comparison, a baseline “human, gridded” macro placement was generated by a human for the same canvas size, I/O placement and gridding.

Macro Placement generated by human (Util: 68%)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	215779	1018355.73	289.999	4545632	-0.003	-0.004	0.09%	0.15%
postCTS	1814274	217192	1018355.73	303.786	4571293	0.001	0	0.13%	0.16%
postRoute	1814274	217192	1018355.73	302.725	4720776	0.206	0

Comparison 2: RePlAce. The standalone RePlAce placer was run on the same (flat) netlist with the same canvas size and I/O placement, with results as follows.

Macro Placement generated Using RePlAce (Util: 68%)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	217246	1018355.73	292.803	4646408	-0.007	-0.011	0.07%	0.13%
postCTS	1814274	218359	1018355.73	306.145	4657174	0.001	0	0.07%	0.17%
postRoute	1814274	218359	1018355.73	305.032	4809950	0.082	0

Comparison 3: RTL-MP. The RTL-MP macro placer was run on the same (flat) netlist with the same canvas size and I/O placement, with results as follows.

Macro Placement generated Using RTL-MP (Util: 68%)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	217057	1018355.73	292.800	4598656	-0.001	-0.001	0.00%	0.01%
postCTS	1814274	218045	1018355.73	306.475	4614827	0.007	0	0.00%	0.01%
postRoute	1814274	218045	1018355.73	303.380	4745004	0.294	0

Comparison 4: The Hier-RTLMP macro placer was run on the same (flat) netlist with the same canvas size and I/O placement, using two setups, with results as follows.

Macro Placement generated Using Hier-RTLMP (Util: 68%) [Setup 1]
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	218096	1018355.73	294.035	4967286	0.003	0	0.10%	0.12%
postCTS	1814274	219150	1018355.73	308.130	4984385	0.001	0	0.13%	0.13%
postRoute	1814274	219150	1018355.73	307.103	5137430	0.387	0

Macro Placement generated Using Hier-RTLMP (Util: 68%) [Setup 2]
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	216665	1018355.73	291.332	4917102	0.001	0	0.02%	0.06%
postCTS	1814274	217995	1018355.73	305.089	4931432	0.001	0	0.03%	0.05%
postRoute	1814274	217995	1018355.73	303.905	5048575	0.230	0

August 25: Replication of the congestion component of proxy cost. Reverse-engineering from the plc client API is finally completed, as described here. A review with Dr. Mustafa Yazgan was very helpful in confirming the case analysis and conventions identified during reverse-engineering. Replication results are shown below. With this, reproduction in open source code of the Circuit Training proxy cost has been completed. Note that the description here illustrates how the Nature paper, Circuit Training, and Google engineers’ versions can have minor discrepancies. (These minor discrepancies are not currently viewed as substantive, i.e., meaningfully affecting our ongoing assessment.) For example, to calculate the congestion component, the H- and V-routing congestion cost lists are concatenated, and the ABU5 (average of top 5% of the concatenated list) metric of this list is the congestion cost. By contrast, the Nature paper indicates use of an ABU10 metric. Recall: “There is no substitute for source code.”

Name	Description	Canvas Size	Col/Row	Congestion Smoothing	Google’s Congestion	Our Congestion
Ariane	Google’s Ariane	356.592 / 356.640	35 / 33	0	3.385729893179586	3.3857299314069733
Ariane133	Our Ariane	1599.99 / 1600.06	24 / 21	0	1.132108622298701	1.1321086382282062
Ariane	Google’s Ariane	356.592 / 356.640	35 / 33	1	2.812822828059799	2.81282287498789
Ariane133	Our Ariane	1599.99 / 1600.06	24 / 21	1	1.116203573147857	1.1162035989647672
Ariane	Google’s Ariane	356.592 / 356.640	35 / 33	2	2.656602005772668	2.6566020148393146
Ariane133	Our Ariane	1599.99 / 1600.06	24 / 21	2	1.109241385529823	1.1092414113467333

August 26: Moving on to understand benefits and limitations of the Circuit Training methodology itself. This next stage of study is enabled by confidence in the technical solidity of what has been accomplished so far – again, with the help of Google engineers.

Question 1. How does having an initial set of placement locations (from physical synthesis) affect the (relative) quality of the CT result?

A preliminary exercise has compared outcomes when the Genus iSpatial (x,y) coordinates are given, versus when vacuous (x,y) coordinates are given. The following CT result is for the “Our Ariane133-NanGate45_68” example where the input protobuf netlist to Circuit Training’s grouping code has all macro and standard cell locations set to (600, 600). This is just an exercise for now: other, carefully-designed experiments will be performed over the coming weeks and months.

Macro Placement generated using CT (Util: 68%) with a vacuous set of input (x,y) coordinates. The input protobuf netlist to Circuit Training’s grouping code has all macro and standard cell locations set to (600, 600).
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	216069	1018355.73	290.0818	4615961	-0.004	-0.021	0.01%	0.03%
postCTS	1814274	217118	1018355.73	303.7199	4619727	0	0	0.01%	0.02%
postRoute	1814274	217118	1018355.73	302.4018	4738717	0.171	0

Update to Question 1 on September 9: Two additional vacuous placements were run through the CT flow.

• Place all macros and standard cells at the lower left corner i.e., (0, 0).
• Place all macros and standard cells at the upper right corner, i.e., (max_x, max_y), where max_x = 1347.1 and max_y = 1346.8.
• (0, 0) gives us the best (by a small amount) result among the three vacuous placements. It has been requested that we report variances and p values. We are unsure how to resource such a request. Note that the original baseline result here, using the (x,y) information from physical synthesis, achieves a final routed wirelength of 4295572, around 7% better than the (0, 0) result.

The following table and screenshots show results for the (0, 0) vacuous placement.

Macro Placement generated using CT (Util: 68%) with a vacuous set of input (x,y) coordinates. The input protobuf netlist to Circuit Training’s grouping code has all macro and standard cell locations set to (0, 0).
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	215520	1018356	289.676	4489121	-0.006	-0.007	0.02%	0.09%
postCTS	1814274	216891	1018356	302.551	4495430	0.005	0	0.02%	0.10%
postRoute	1814274	216891	1018356	301.322	4606716	0.218	0

The following table and screenshots show results for (max_x, max_y), where max_x = 1347.1 and max_y = 1346.8.

Macro Placement generated using CT (Util: 68%) with a vacuous set of input (x,y) coordinates. The input protobuf netlist to Circuit Training’s grouping code has all macro and standard cell locations set to (max_x, max_y) = (1347.1, 1346.8)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	214817	1018356	288.454	4530507	0.002	0	0.01%	0.04%
postCTS	1814274	215844	1018356	301.719	4532853	0.007	0	0.03%	0.05%
postRoute	1814274	215844	1018356	300.763	4646396	0.228	0

Question 2. How does utilization affect the (relative) performance of CT?

Question 3. Is a testcase such as Ariane-133 “probative”, or do we need better testcases?

A preliminary exercise has examined Innovus P&R outcomes when the Circuit Training macro placement locations for Our Ariane133-NanGate45_68 are randomly shuffled. The results for four seed values used in the shuffle, and for the original Circuit Training result, are as follows. (We have extended this experiment here.)

Metric	Shuffle-1	Shuffle-2	Shuffle-3	Shuffle-4	CT_Result
Core_area (um^2)	1814274.28	1814274.28	1814274.28	1814274.28	1814274.28
Macro_area (um^2)	1018355.73	1018355.73	1018355.73	1018355.73	1018355.73
preCTS_std_cell_area (um^2)	217124.89	217168.25	217157.88	217020.09	215575.44
postCTS_std_cell_area (um^2)	218215.23	218231.19	218328.81	218073.45	217114.52
postRoute_std_cell_area (um^2)	218215.23	218231.19	218328.81	218073.45	217114.52
preCTS_total_power (mW)	292.032	292.692	292.676	292.764	288.762
postCTS_total_power (mW)	305.726	306.497	306.120	306.524	302.607
preRoute_total_power (mW)	304.394	304.996	304.711	305.093	301.722
preCTS_wirelength (um)	5057900	5069848	5092665	5119539	4170253
postCTS_wirelength (um)	5063278	5079451	5109801	5126540	4186888
postRoute_wirelength (um)	5186032	5194397	5227411	5247799	4295572
preCTS_WS (ns)	-0.006	0.001	0	-0.003	0.002
postCTS_WS (ns)	0.002	0.002	0.003	0.002	0.001
postRoute_WS (ns)	0.174	0.090	0.219	0.349	0.336
preCTS_TNS (ns)	-0.010	0	0	-0.019	0
postCTS_TNS (ns)	0	0	0	0	0
postRoute_TNS (ns)	0	0	0	0	0
preCTS_Congestion(H)	0.02%	0.02%	0.03%	0.02%	0.01%
postCTS_Congestion(H)	0.03%	0.04%	0.02%	0.06%	0.00%
postRoute_Congestion(H)
preCTS_Congestion(V)	0.06%	0.06%	0.07%	0.07%	0.01%
postCTS_Congestion(V)	0.07%	0.07%	0.08%	0.08%	0.01%
postRoute_Congestion(V)

September 9:

• We have added two more vacuous initial placements to the study of Question 1.
• We have added an initial study of impact from placement guidance to clustering. See Question 4.
• We have taken a look at the impact of Coordinate Descent on proxy cost and on Table 1 metrics. See Question 5.
• We have obtained a data point to compare two alternate Cadence flows for obtaining the initial macro placement. See Question 6.
• We have taken a look at a potential new baseline, which is simply to let the commercial physical synthesis / P&R tool flow run until the end of routing, without any involvement of CT. See Question 7.
• We have obtained an initial CT result on a second testcase, NVDLA, here.
• As this running log is becoming unwieldy, we propose to pin a summary of questions and conclusions to date at the bottom of this document. We will also add this into our GitHub, as planned. And, we request that questions and experimental requests be posed as GitHub issues, and that the limited bandwidth and resources of students be taken into account when making these requests.

Question 4. How much does the guidance to clustering that comes from (x,y) locations matter?

We answer this by using hMETIS to generate the same number of soft macros from the same netlist, but only via the npart (number of partitions) parameter. The value of npart in the call to hMETIS is chosen to match the number of standard-cell clusters (i.e., soft macros) obtained in the CT grouping process. Then, to preserve this number of soft macros, we skip the break up and merge stage in CT grouping.

[Brief overview of break up and merge: (A) Break up: During break up, if a standard cell cluster height or width is greater than sqrt(canvas area / 16), then it is broken into small clusters such that the height and width of each cluster is less than sqrt(canvas area / 16). (B) Merge: During merge, if the number of standard cells is less than the (average number of standard cells in a cluster / 4), then the standard cells of that cluster are moved to their neighboring clusters.]

We run hMETIS with npart = 810 (number of fixed groups is 153) to match the total number of standard cell clusters when CT’s break up and merge is run. The following table presents the results of this experiment. Outcomes are similar to the original Ariane133-NG45 with 68% utilization CT result. [The Question 1 study indicates that a vacuous placement harms the outcome of CT, i.e., “placement information matters”. But the Question 4 study suggests that a flow that does not bring in any placement coordinates (i.e., using pure hMETIS partitioning down to a similar number of stdcell clusters) does not affect results by much.]

Macro Placement generated using CT (Util: 68%) when the input clustered netlist is generated by running hMETIS npart = 810 and without running break up and merge
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	215552	1018356	288.642	4188406	-0.001	-0.001	0.02%	0.12%
postCTS	1814274	216618	1018356	302.086	4196172	0.002	0	0.02%	0.11%
postRoute	1814274	216618	1018356	300.899	4304113	0.264	0

Question 5. What is the impact of the Coordinate Descent (CD) placer on proxy cost and Table 1 metric?

In our August 18 notes, we mentioned that the default CT flow does NOT run coordinate descent. (Coordinate descent is not mentioned in the Nature paper.) The result in the CT repo shows the impact of Coordinate Descent (CD) on proxy cost for the Google Ariane design, but there is no data to show the impact of CD on Table 1 metrics.

We have taken the CT results generated for Ariane133-NG45 with 68% utilization through the CD placement step. The following table shows the effect of CD placer on proxy cost. The CD placer for this instance improves proxy wirelength and density at the cost of congestion, and overall proxy cost degrades slightly.

CD Placer effect on Proxy cost for Ariane133
Cost	CT w/o CD	+ Apply CD
Wirelength	0.0948	0.0861
Density	0.4845	0.4746
Congestion	0.7176	0.7574
Proxy	0.6959	0.7021

The following table shows the P&R result for the post-CD macro placement.

Macro placement generated by applying the Coordinate Descent placement step to Our Ariane-133 (NG45) 68% utilization when the input to the CD placer is the (default setup) CT macro placement. The post-macro placement flow uses Innovus21.1
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	215581	1018356	289.312	4238854	-0.001	-0.003	0.01%	0.06%
postCTS	1814274	217017	1018356	302.483	4249846	0.005	0	0.02%	0.07%
postRoute	1814274	217017	1018356	301.482	4358888	0.140	0

Even though CD improves proxy wirelength, the post-route wirelength worsens slightly (by ~1.47%) compared to the original CT macro placement.

Question 6. Are we using the industry tool in an “expert” manner? (We believe so.) We received an inquiry regarding the multiple ways in which macro placements could be obtained using Cadence tooling. To clarify:

• In our previous CT result shown here, the initial macro placement (which is fed into Genus iSpatial) is generated using Innovus Concurrent Macro Placer.
• It is also possible to use Genus iSpatial to perform both macro and standard-cell placement. In our experience, this worsens results, as shown below. I.e., based on our current understanding, the macro placement produced by Innovus Concurrent Macro Placer leads to the best results when fed to the CT flow.

Macro placement generated by Circuit Training on Our Ariane-133 (NG45) 68% utilization when the input macro and standard cell placement to CT grouping is generated by Genus iSpatial, and the post-macro placement flow is using Innovus21.1
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	215583	1018355.73	289.030	4476331	-0.002	-0.002	0.02%	0.03%
postCTS	1814274	216729	1018355.73	302.268	4483560	0.002	0	0.03%	0.09%
postRoute	1814274	216729	1018355.73	301.028	4590581	0.316	0

Question 7. What happens if we skip CT and continue directly to standard-cell P&R (i.e., the Innovus 21.1 flow) once we have a macro placement from the commercial tool?

At some point during the past weeks, we realized that this would also be a potential “baseline” for comparison. As can be seen below for both 68% and 51% variants of Ariane-133 in NG45, omitting the CT step can also produce good results by the Table 1 metrics. At this point, we do not have any diagnosis or interpretation of this data. One possible implication is that the Ariane-133 testcase is in some way not probative. The community’s suggestions (e.g., alternate testcases, constraints, floorplan setup, etc.) are always welcome.

Concurrent macro placement (Ariane 68%) continuing straight into the Innovus 21.1 P&R flow (no application of Circuit Training) [baseline CT result: here]
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	214050	1018355.73	286.117	3656436	0.007	0	0.02%	0.01%
postCTS	1814274	215096	1018355.73	299.438	3662225	0.01	0	0.01%	0.02%
postRoute	1814274	215096	1018355.73	298.934	3780153	0.285	0
Concurrent macro placement (Ariane 51%) continuing straight into the Innovus 21.1 P&R flow (no application of Circuit Training) [baseline CT result: here]
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	2560080	214060	1018355.73	285.509	3647997	0.047	0	0.00%	0.00%
postCTS	2560080	215117	1018355.73	298.362	3649940	0.011	0	0.00%	0.01%
postRoute	2560080	215117	1018355.73	297.849	3764148	0.210	0

Ariane 68%:

Question 8. How does the tightness of timing constraints affect the (relative) performance of CT?

[Comment: This is related to Question 2, and is part of the broad question of field of use / sweet spot. We still intend to work in the space of {design testcase} X {technology and design enablement} X {utilization} X {performance requirement}X experimental {questions, design/setup, execution} to reach conclusions that are above the bar of “satisfying readers”. Progress will continue to be reported here and in GitHub.]

Circuit Training Baseline Result on “Our NVDLA-NanGate45_68”.

We have trained CT to generate a macro placement for the NVDLA design. For this experiment we use the NanGate45 enablement; the initial canvas size is generated by setting utilization to 68%. We use the default hyperparameters used for Ariane to train CT for NVDLA design. The number of hard macros in NVDLA is 128, so we update max_sequnece_length to 129 in ppo_collect.py and sequence_length to 129 in train_ppo.py.

The following table and screenshots show the CT result.

Macro placement generated by Circuit Training on Our NVDLA (NG45) 68% utilization, post-macro placement flow using Innovus21.1
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	4002458	401713	2325683	2428.453	13601973	-0.003	-0.045	0.40%	1.22%
postCTS	4002458	404398	2325683	2514.685	13677780	-0.009	-0.027	0.44%	1.54%
postRoute	4002458	404398	2325683	2491.368	14317085	0.142	0

September 18:

• To address Question 8, we have performed a sweep of target clock period (TCP) constraint for Ariane133-68 in NG45. Experiments above were performed with a loose TCP of 4.0ns. According to our studies, the “hockey stick” ends at a TCP of 1.3ns, so we have generated netlists and run CT for TCP values of 1.3ns and 1.5ns. The results are shown below (post-physical synthesis summary results with TCP values of 4.0ns, 1.5ns, 1.3ns; CT + Innovus P&R results for 1.5ns, 1.3ns). We see that the wirelength numbers are worse for CT results compared to the CMP result, but the timing numbers for CT are better than CMP.
  • The following table shows the post-physical synthesis results of Ariane133-68-NG45 for different TCPs when the macro placement is generated using CMP.

Ariane133-NG45-68%-4.0ns CMP (Link to CT result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	215033	1018356	286.199	3535026	-0.001	-0.001	0.04%	0.01%
postCTS	1814274	216147	1018356	299.635	3544668	0.001	0	0.02%	0.01%
postRoute	1814274	216147	1018356	299.110	3649892	0.317	0
postRouteOpt	1814274	215738	1018356	295.127	3653200	0.397	0
Ariane133-NG45-68%-1.5ns CMP (Link to CT result]
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	232370	1018356	682.777	3635909	-0.008	-0.143	0.01%	0.01%
postCTS	1814274	234250	1018356	718.592	3663001	-0.002	-0.006	0.03%	0.10%
postRoute	1814274	234250	1018356	717.410	3777403	-0.221	-86.88
postRouteOpt	1814274	237178	1018356	718.866	3785973	-0.042	-6.311
Ariane133-NG45-68%-1.3ns CMP (Link to CT result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	251874	1018356	807.994	3885279	-0.15	-242.589	0.02%	0.02%
postCTS	1814274	254721	1018356	851.977	3923912	-0.127	-133.426	0.04%	0.10%
postRoute	1814274	254721	1018356	850.483	4049905	-0.239	-410.578
postRouteOpt	1814274	256230	1018356	851.546	4057140	-0.154	-196.527

• The following table shows the post-physical synthesis results of Ariane133-68-NG45 for different TCPs when the macro placement is generated using CT.

Ariane133-NG45-68%-1.5ns CT (Link to CMP result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	227917	1018356	673.158	4243883	-0.012	-0.648	0.03%	0.03%
postCTS	1814274	229836	1018356	708.797	4247346	-0.001	-0.007	0.07%	0.12%
postRoute	1814274	229836	1018356	707.522	4360419	-0.052	-9.218
postRouteOpt	1814274	230164	1018356	707.829	4364537	-0.009	-0.233

Ariane133-NG45-68%-1.3ns CT (Link to CMP result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
postSynth	1814274	244614	1018356	761.754	4884882	-0.764	-533.519
preCTS	1814274	244373	1018356	792.626	4732895	-0.123	-184.135	0.03%	0.11%
postCTS	1814274	247965	1018356	837.464	4762751	-0.084	-35.57	0.04%	0.15%
postRoute	1814274	247965	1018356	835.824	4887126	-0.123	-63.739
postRouteOpt	1814274	248448	1018356	836.399	4892431	-0.09	-57.448

September 19: We updated the detailed algorithm for gridding in Circuit Training. In contrast to the open-source grid_size_selection.py in Circuit Training repo, which still calls the wrapper functions of plc client, our python scripts implement the gridding from scratch and are easy to understand. The results of our scripts match exactly that of Circuit Training.

September 21: We updated the detailed algorithm for grouping and Clustering. Here we explicitly show how the netlist information such as net model is used during grouping and clustering, while the open-source Circuit Training implementation still calls the wrapper function of the plc client to get netlist information.

Among the more notable details that were not apparent from the Nature paper or the Circuit Training repo:

• For the gridding, we summarized the detailed algorithm for the entire gridding process. We also provided the details for macro packing and metric calculation.
• For the grouping, we identified how to translate the protocol buffer netlist into the hypergraph, which is the input to the hMETIS hypergraph partitioner when the gate-level netlist is clustered into soft macros.
• For the grouping, we also identified the details for each step: grouping the macro pins of the same macro into a cluster; grouping the IOs that are within close proximity of each other, boundary by boundary; grouping the closely-related standard cells, which connect to the same macro or the same IO cluster.
• For the clustering, we solved the following key issues: what exactly is the Hypergraph, and how is it partitioned? How to break up clusters that span a distance larger than breakup_threshold? And how to recursively merge small adjacent clusters?

September 30:

Circuit Training Baseline Result on “Our bp_quad-NanGate45_68”. We have trained CT to generate a macro placement for the bp_quad design. For this experiment we use the NanGate45 enablement; the initial canvas size is generated by setting utilization to 68%. We use the default hyperparameters used for Ariane to train CT for bp_quad design. The number of hard macros in bp_quad is 220, so we update max_sequence_length to 221 in ppo_collect.py and sequence_length to 221 in train_ppo.py.

bp_quad-NG45-68% CT result (Link to Tensorboard) (Link to corresponding CMP result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
postSynth	8449457	1828674	3917822	1903.716	36067460	0.325	0
preCTS	8449457	1827246	3917822	2042.610	35593805	-0.015	-0.64	0.12%	0.19%
postCTS	8449457	1836549	3917822	2214.398	35633384	0	0	0.14%	0.22%
postRoute	8449457	1836549	3917822	2197.750	36681437	-0.11	-63.817
postRouteOpt	8449457	1836148	3917822	2197.478	36718051	-0.003	-0.013

bp_quad-NG45-68% CMP result (Link to corresponding CT result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
postSynth	8449457	1808903	3917822	1875.440	20854975	0.327	0
preCTS	8449457	1814511	3917822	1990.066	20766279	-0.004	-0.041	0.02%	0.04%
postCTS	8449457	1824057	3917822	2160.034	20870489	0	0	0.03%	0.05%
postRoute	8449457	1824057	3917822	2159.687	21535697	-0.343	-307.935
postRouteOpt	8449457	1824031	3917822	2159.211	21556685	-0.003	-0.029

October 3:
We shared the Ariane133-NG45-68% protobuf netlist and clustered netlist with Google engineers. They ran training on the clustered netlist, and the following table shows the Table 1 metrics and proxy cost. Our training results resemble Google’s results.

Ariane-NG45-68%-4ns CMP result (Link to Our Result) (Link to tensorboard)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	215608	1018356	288.736	4260100	-0.001	-0.001	0.01%	0.01%
postCTS	1814274	216693	1018356	302.205	4268402	0.001	0	0.02%	0.02%
postRoute	1814274	216693	1018356	301.129	4377728	0.193	0

Cost	Ours	Google’s
Wirelength	0.0999	0.1023
Congestion	0.8906	0.9175
Density	0.4896	0.4773
Proxy	0.7900	0.7997

October 9:

Question 9. Are CT results stable? If not, how much does the outcome vary?

We see from the results in the CT repo that the outcomes of three runs with the same seed value are different. We ran six CT runs for Ariane133-NG45-68%-1.3ns design, and the following tables show the Table 1 metrics and the proxy cost details.

Metrics	Run1	Run2	Run3	Run4	Run5	Run6
core_area(um^2)	1814274	1814274	1814274	1814274	1814274	1814274
macro_area(um^2)	1018356	1018356	1018356	1018356	1018356	1018356
postSynth_std_cell_area(um^2)	245871	243223	242695	243382	246725	242711
preCTS_std_cell_area(um^2)	245235	244615	245921	243693	245426	241760
postCTS_std_cell_area(um^2)	247138	245862	246186	246099	247774	244237
postRoute_std_cell_area(um^2)	247138	245862	246186	246099	247774	244237
postRouteOpt_std_cell_area(um^2)	247725	246159	246776	246498	248151	244594
postSynth_total_power(mw)	757.853	751.37	755.971	769.154	760.549	759.477
preCTS_total_power(mw)	795.381	791.633	794.2	793.175	794.542	790.433
postCTS_total_power(mw)	837.759	833.972	833.019	837.791	837.733	833.350
postRoute_total_power(mw)	835.807	832.593	831.162	836.205	836.124	831.401
postRouteOpt_total_power(mw)	836.529	832.975	831.524	836.826	835.521	831.911
preCTS_wirelength(um)	4792929	4495121	4709296	4673400	4735851	4902798
postCTS_wirelength(um)	4833093	4529411	4749013	4690341	4777561	4929463
postRoute_wirelength(um)	4955517	4649621	4869873	4816827	4903796	5054361
postRouteOpt_wirelength(um)	4960472	4654146	4875070	4821225	4908694	5059042
postSynth_WS(ns)	-0.764	-0.764	-0.764	-0.764	-0.764	-0.764
preCTS_WS(ns)	-0.135	-0.104	-0.109	-0.1	-0.086	-0.091
postCTS_WS(ns)	-0.102	-0.056	-0.069	-0.106	-0.077	-0.08
postRoute_WS(ns)	-0.134	-0.077	-0.102	-0.13	-0.106	-0.089
postRouteOpt_WS(ns)	-0.133	-0.076	-0.105	-0.135	-0.081	-0.083
postSynth_TNS(ns)	-366.528	-592.301	-501.314	-363.351	-405.145	-342.59
preCTS_TNS(ns)	-196.114	-136.662	-151.307	-122.663	-104.413	-98.21
postCTS_TNS(ns)	-76.567	-13.883	-40.712	-60.272	-27.453	-21.711
postRoute_TNS(ns)	-167.965	-58.724	-110.496	-133.653	-45.42	-44.821
postRouteOpt_TNS(ns)	-123.027	-27.571	-79.826	-105.775	-33.286	-40.314
preCTS_Congestion (H)	0.06%	0.04%	0.03%	0.03%	0.03%	0.03%
postCTS_Congestion (H)	0.09%	0.03%	0.04%	0.03%	0.04%	0.05%
preCTS_Congestion (V)	0.11%	0.10%	0.13%	0.08%	0.16%	0.14%
postCTS_Congestion (V)	0.13%	0.13%	0.17%	0.12%	0.18%	0.18%

	Wirelength cost	Congestion cost	Density cost	Proxy cost
Run1	0.1052	0.97	0.5239	0.85215
Run2	0.1045	0.9417	0.5063	0.8285
Run3	0.1033	0.949	0.5193	0.83745
Run4	0.1034	0.9378	0.5185	0.8316
Run5	0.1056	0.9328	0.5418	0.8429
Run6	0.1104	0.96	0.5372	0.8590
Mean	0.1054	0.9486	0.5245	0.8419
STD	0.0026	0.0142	0.0131	0.0119

We further ran coordinate descent (CD) placer on the CT outcomes and the following tables show the Table 1 metrics and proxy cost details of the CD placer outcomes. Even though we see a significant improvement in the proxy cost, we do not see similar improvement in the Table 1 metric.

Metrics	Run1_CD	Run2_CD	Run3_CD	Run4_CD	Run5_CD	Run6_CD
core_area (um2)	1814274	1814274	1814274	1814274	1814274	1814274
macro_area (um2)	1018356	1018356	1018356	1018356	1018356	1018356
postSynth_std_cell_area (um2)	243566	244506	244016	244368	242548	247357
preCTS_std_cell_area (um2)	243267	241949	240051	245803	242336	245297
postCTS_std_cell_area (um2)	246719	244046	241932	247881	244474	247763
postRoute_std_cell_area (um2)	246719	244046	241932	247881	244474	247763
postRouteOpt_std_cell_area (um2)	247000	243860	241282	248055	245020	248377
postSynth_total_power (mW)	736.564	747.327	758.3497	749.487	752.643	750.437
preCTS_total_power (mW)	790.601	788.404	785.7521	797.216	789.500	794.160
postCTS_total_power (mW)	835.029	830.542	827.7217	839.145	832.896	836.920
postRoute_total_power (mW)	833.305	829.015	825.9415	837.320	830.757	835.113
postRouteOpt_total_power (mW)	833.109	828.801	824.8444	837.595	831.417	835.770
preCTS_wirelength (um)	4807227	4481988	4663403	4645833	4742585	4813011
postCTS_wirelength (um)	4830788	4501231	4680124	4683338	4779530	4839729
postRoute_wirelength (um)	4955395	4621695	4804536	4809309	4896653	4965139
postRouteOpt_wirelength (um)	4960842	4626687	4809650	4814381	4901760	4969937
postSynth_WS (ns)	-0.764	-0.764	-0.764	-0.764	-0.764	-0.764
preCTS_WS (ns)	-0.11	-0.092	-0.065	-0.115	-0.105	-0.143
postCTS_WS (ns)	-0.102	-0.058	-0.056	-0.101	-0.094	-0.11
postRoute_WS (ns)	-0.135	-0.076	-0.088	-0.107	-0.11	-0.14
postRouteOpt_WS (ns)	-0.129	-0.062	-0.055	-0.101	-0.109	-0.137
postSynth_TNS (ns)	-351.045	-331.782	-406.717	-431.986	-450.335	-444.635
preCTS_TNS (ns)	-133.192	-90.187	-57.052	-152.966	-139.133	-196.673
postCTS_TNS (ns)	-55.003	-19.074	-8.908	-47.75	-52.329	-101.123
postRoute_TNS (ns)	-145.14	-31.185	-15.033	-82.306	-96.749	-157.245
postRouteOpt_TNS (ns)	-109.739	-12.692	-8.418	-60.53	-66.632	-126.007
preCTS_Congestion (H)	0.03%	0.03%	0.07%	0.05%	0.04%	0.04%
postCTS_Congestion (H)	0.03%	0.03%	0.07%	0.05%	0.04%	0.05%
preCTS_Congestion (V)	0.16%	0.12%	0.10%	0.15%	0.17%	0.14%
postCTS_Congestion (V)	0.19%	0.16%	0.10%	0.18%	0.21%	0.15%

	Wirelength cost	Congestion cost	Density cost	Proxy cost
Run1_CD	0.0944	0.7942	0.4927	0.73785
Run2_CD	0.089	0.7829	0.4925	0.7267
Run3_CD	0.0928	0.796	0.4931	0.73735
Run4_CD	0.0957	0.8104	0.4951	0.7485
Run5_CD	0.0909	0.7799	0.4933	0.7275
Run6_CD	0.0922	0.7843	0.4934	0.7311
Mean	0.0925	0.7913	0.4934	0.7348
STD	0.0024	0.0114	0.0009	0.0082

October 15:
Question 10. What is the correlation between proxy cost and the post RouteOpt metrics?

We have collected macro placement generated by CT runs for Ariane133-NG45-68%-1.3ns that have proxy cost less than 0.9. There are ~40 such macro placements over four CT runs. From that 15 runs are chosen randomly, two runs from each bucket of proxy cost (0.9-i0.01, 0.9-(i+1)0.01] s.t. i ε [0, 6] and one run from (0.82, 0.83]. Table 1 metrics and proxy costs of these 15 runs are available in the following table.

	RUN1	RUN2	RUN3	RUN4	RUN5	RUN6	RUN7	RUN8	RUN9	RUN10	RUN11	RUN12	RUN13	RUN14	RUN15
core_area (um^2)	1814274	1814274	1814274	1814274	1814274	1814274	1814274	1814274	1814274	1814274	1814274	1814274	1814274	1814274	1814274
macro_area (um^2)	1018356	1018356	1018356	1018356	1018356	1018356	1018356	1018356	1018356	1018356	1018356	1018356	1018356	1018356	1018356
postSynth_std_cell_area (um^2)	242067	243116	243055	246488	243788	244004	244090	244844	245083	246072	240942	246725	242695	243643	243223
preCTS_std_cell_area (um^2)	243195	245232	242421	244504	244174	245232	241542	246361	243436	246115	244612	245426	245921	244513	244615
postCTS_std_cell_area (um^2)	246379	247012	243583	247185	246155	247948	244115	248349	247013	248156	246469	247774	246186	247138	245862
postRoute_std_cell_area (um^2)	246379	247012	243583	247185	246155	247948	244115	248349	247013	248156	246469	247774	246186	247138	245862
postRouteOpt_std_cell_area (um^2)	247121	247607	243894	247394	246878	248433	244274	248746	247320	248770	247390	248151	246776	247547	246159
postSynth_total_power (mw)	769.520	753.509	742.910	752.287	752.254	741.871	756.514	753.901	753.265	749.084	750.949	760.549	755.971	753.220	751.370
preCTS_total_power (mw)	791.074	793.708	787.915	792.428	791.913	792.947	787.022	791.689	790.387	795.202	791.286	794.542	794.200	791.590	791.633
postCTS_total_power (mw)	834.752	836.171	829.367	834.354	833.401	836.912	830.593	835.061	831.509	833.914	832.950	837.733	833.019	835.334	833.972
postRoute_total_power (mw)	833.184	834.695	828.029	833.086	831.875	835.325	828.821	833.941	830.484	832.671	831.772	836.124	831.162	833.983	832.593
postRouteOpt_total_power (mw)	833.961	835.436	828.254	833.318	832.649	835.803	829.066	834.304	831.652	833.287	832.768	835.521	831.524	834.484	832.975
preCTS_wirelength (um)	4728745	4717333	4642346	4628632	4659824	4873402	4882098	4543637	4649807	4709934	4486281	4735851	4709296	4585732	4495121
postCTS_wirelength (um)	4762085	4757761	4674012	4665159	4693884	4912764	4918705	4585918	4677979	4742407	4522423	4777561	4749013	4616680	4529411
postRoute_wirelength (um)	4885433	4888249	4797431	4795134	4817647	5042041	5043542	4716210	4807107	4869741	4650492	4903796	4869873	4742247	4649621
postRouteOpt_wirelength (um)	4890958	4893245	4802406	4800104	4822688	5047120	5048498	4720614	4811606	4874840	4655745	4908694	4875070	4746909	4654146
Wirelength_Cost	0.1042	0.1011	0.1032	0.1014	0.1032	0.1055	0.1064	0.1027	0.1048	0.1027	0.1023	0.1056	0.1033	0.1053	0.1045
postSynth_WS (ns)	-0.764	-0.764	-0.764	-0.79	-0.764	-0.764	-0.79	-0.764	-0.764	-0.764	-0.764	-0.764	-0.764	-0.764	-0.764
preCTS_WS (ns)	-0.114	-0.101	-0.08	-0.096	-0.116	-0.101	-0.066	-0.121	-0.117	-0.137	-0.124	-0.086	-0.109	-0.125	-0.104
postCTS_WS (ns)	-0.088	-0.08	-0.036	-0.066	-0.098	-0.076	-0.021	-0.098	-0.096	-0.053	-0.104	-0.077	-0.069	-0.109	-0.056
postRoute_WS (ns)	-0.121	-0.094	-0.072	-0.341	-0.118	-0.087	-0.088	-0.118	-0.123	-0.134	-0.137	-0.106	-0.102	-0.13	-0.077
postRouteOpt_WS (ns)	-0.125	-0.096	-0.063	-0.066	-0.089	-0.087	-0.041	-0.119	-0.13	-0.099	-0.126	-0.081	-0.105	-0.134	-0.076
postSynth_TNS (ns)	-326.535	-382.684	-477.484	-339.098	-401.614	-414.822	-367.119	-412.85	-422.819	-350.771	-313.919	-405.145	-501.314	-366.866	-592.301
preCTS_TNS (ns)	-147.905	-129.089	-92.977	-111.456	-141.654	-116.344	-62.661	-171.687	-156.067	-206.043	-169.834	-104.413	-151.307	-168.846	-136.662
postCTS_TNS (ns)	-69.386	-67.761	-4.902	-34.67	-60.302	-41.497	-2.514	-83.036	-62.184	-27.629	-122.576	-27.453	-40.712	-55.55	-13.883
postRoute_TNS (ns)	-172.018	-85.027	-48.269	-37.909	-85.811	-70.604	-15.213	-129.351	-128.868	-143.568	-199.374	-45.42	-110.496	-132.265	-58.724
postRouteOpt_TNS (ns)	-135.838	-70.139	-25.199	-33.755	-68.666	-47.43	-14.211	-118.13	-96.63	-105.577	-152.772	-33.286	-79.826	-94.025	-27.571
preCTS_Congestion (H)	0.04%	0.03%	0.04%	0.03%	0.02%	0.05%	0.03%	0.02%	0.03%	0.05%	0.04%	0.03%	0.03%	0.02%	0.04%
postCTS_Congestion (H)	0.05%	0.04%	0.05%	0.06%	0.04%	0.05%	0.04%	0.05%	0.04%	0.04%	0.06%	0.04%	0.04%	0.03%	0.03%
preCTS_Congestion (V)	0.17%	0.16%	0.11%	0.14%	0.16%	0.11%	0.16%	0.13%	0.15%	0.12%	0.14%	0.16%	0.13%	0.11%	0.10%
postCTS_Congestion (V)	0.16%	0.14%	0.13%	0.13%	0.15%	0.12%	0.16%	0.14%	0.18%	0.13%	0.15%	0.18%	0.17%	0.14%	0.13%
Congestion_Cost	1.0192	0.9983	1.0115	1.0062	0.9894	1.006	0.9813	0.9966	0.9932	0.9587	0.9672	0.9328	0.949	0.9439	0.9417
Wirelength_Cost	0.1042	0.1011	0.1032	0.1014	0.1032	0.1055	0.1064	0.1027	0.1048	0.1027	0.1023	0.1056	0.1033	0.1053	0.1045
Congestion_Cost	1.0192	0.9983	1.0115	1.0062	0.9894	1.006	0.9813	0.9966	0.9932	0.9587	0.9672	0.9328	0.949	0.9439	0.9417
Density_Cost	0.5622	0.5923	0.5543	0.5622	0.5523	0.5354	0.5409	0.53	0.5113	0.5439	0.5215	0.5418	0.5193	0.5136	0.5063
Proxy_Cost	0.8949	0.8964	0.8861	0.8856	0.87405	0.8762	0.8675	0.866	0.85705	0.854	0.84665	0.8429	0.83745	0.83405	0.8285

In the following table we report the Kendall rank correlation coefficient for proxy costs and postPlaceOpt metrics and for proxy costs and postRouteOpt metrics. Here values near +1, -1 indicate high correlation or anti-correlation and values near 0 indicate high miscorrelation.

Correlation between PostPlaceOpt metrics and proxy cost
Cost	Std Cell Area	Wirelength	Total Power	Worst Slack	TNS	Congestion (V)	Congestion (H)
Wirelength	-0.09662	0.33655	-0.12501	0.32851	0.29809	-0.06098	0.00000
Congestion	-0.30622	0.10476	-0.23810	0.17225	0.14286	0.18118	0.13093
Density	-0.08654	0.21053	0.15311	0.24038	0.19139	0.35399	0.03289
Proxy	-0.22967	0.23810	-0.06667	0.28708	0.23810	0.32210	0.06547

Correlation between PostRouteOpt metrics and proxy cost
Cost	Std Cell Area	Wirelength	Total Power	Worst Slack	TNS
Wirelength	-0.22116	0.31732	-0.14424	0.16347	0.31732
Congestion	-0.02857	0.08571	-0.00952	0.10476	-0.04762
Density	0.09569	0.22967	0.09569	0.26795	0.07656
Proxy	-0.00952	0.25714	0.04762	0.20000	0.04762

• Kendall rank correlation coefficients indicate poor correlation between proxy cost and postPlaceOpt metrics. Similarly, we see a poor correlation between proxy cost and postRouteOpt metrics.
  • We see the proxy costs of RUN3 and RUN7 are 0.8861 and 0.8675 respectively, which is much higher than the best proxy cost of 0.8285 (corresponding to RUN15), but the total power and TNS for RUN3 and RUN7 are better than RUN15.

Circuit Training Baseline Result on “Our MemPool_Group-NanGate45_68”.
We have trained CT to generate a macro placement for the MemPool Group design. For this experiment we use the NanGate45 enablement; the initial canvas size is generated by setting utilization to 68%. We use the default hyperparameters used for Ariane to train CT for bp_quad design. The number of hard macros in MemPool Group is 324, so we update max_sequence_length to 325 in ppo_collect.py and sequence_length to 325 in train_ppo.py.

MemPool group-NG45-68%-4ns CT result (Flow2. Final DRC Count: 19367) (Link to Tensorboard)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
postSynth	11371934	4976373	3078071	3149.187	113753318	0	0
preCTS	11371934	4916168	3078071	2528.429	113557846	-0.033	-42.949	3.03%	1.51%
postCTS	11371934	4867885	3078071	2707.906	113908550	-0.001	-0.018	3.55%	1.76%
postRoute	11371934	4867885	3078071	2742.635	123398335	-0.749	-13254.6
postRouteOpt	11371934	4861749	3078071	2742.982	123578279	-0.206	-26.811

MemPool group-NG45-68%-4ns CMP result (Flow2. Final DRC Count: 26)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
postSynth	11371934	4947251	3078071	2938.815	94419498	0	0
preCTS	11371934	4891095	3078071	2402.835	96594902	-0.018	-150.478	1.72%	0.78%
postCTS	11371934	4846216	3078071	2584.086	97108227	-0.003	-0.043	1.85%	0.87%
postRoute	11371934	4846216	3078071	2589.973	102792205	-0.241	-4400.6
postRouteOpt	11371934	4837150	3078071	2586.602	102907484	-0.02	-1.029

November 25:
We document two variant Evaluation Flows (taking macro placements through Innovus place-and-route) that we use, in this Evaluation Flow document. Posted results up to now have been obtained with Evaluation Flow 2. The Evaluation Flow document shows that results and conclusions are nearly identical between Evaluation Flow 1 and Evaluation Flow 2. However, going forward we will report our macro placement assessments using Evaluation Flow 1.

CT Results with a Commercial (GLOBALFOUNDRIES 12nm) Design Enablement
We have run CT to generate macro placements for Ariane133, BlackParrot and MemPool Group designs on GLOBALFOUNDRIES 12nm (GF12) enablement. The following tables present the normalized design metrics. Core area, standard cell area and macro area are normalized with respect to the core area. Total power is normalized with respect to the reported preCTS total power when CMP is used. Similarly, we normalize the wirelength and congestion based on the reported preCTS wirelength and congestion when CMP is used. The timing numbers are normalized with respect to the target clock period.

• The following table and screenshots provide details of the Ariane133 GF12 implementation when CMP is used to generate the macro placement.

Ariane133-GF12-68% CMP (results are normalized as described here )
Physical Design Stage	Core Area	Standard Cell Area	Macro Area	Total Power	Wirelength	WS	TNS	Congestion (H)	Congestion (V)
preCTS	1	0.137	0.555	1.0000	1.0000	-0.130	-259.985	0.00	1.00
postCTS	1	0.139	0.555	1.1442	1.0112	-0.145	-114.783	0.00	1.00
postRoute	1	0.139	0.555	1.1356	1.0432	-0.185	-142.688
postRouteOpt	1	0.139	0.555	1.1352	1.0443	-0.159	-142.274

• The following table and screenshots provide details of Ariane133 GF12 implementation when CT is used to generate the macro placement.

Ariane133-GF12-68% CT (results are normalized as described here) (Link to Tensorboard)
Physical Design Stage	Core Area	Standard Cell Area	Macro Area	Total Power	Wirelength	WS	TNS	Congestion (H)	Congestion (V)
preCTS	1	0.138	0.555	1.0120	1.1652	-0.130	-239.531	0.00	0.50
postCTS	1	0.140	0.555	1.1623	1.1828	-0.138	-140.220	0.00	1.00
postRoute	1	0.140	0.555	1.1530	1.2151	-0.138	-145.883
postRouteOpt	1	0.140	0.555	1.1519	1.2161	-0.145	-115.805

• (Updated on December 20) The following table and screenshots provide details of Ariane133 GF12 implementation when AutoDMP is used to generate the macro placement.

Ariane-GF12-68% AutoDMP (results are normalized as described here)
Physical Design Stage	Core Area	Standard Cell Area	Macro Area	Total Power	Wirelength	WS	TNS	Congestion (H)	Congestion (V)
preCTS	1	0.136	0.555	0.9941	1.0214	-0.116	-204.181	0.00	0.50
postCTS	1	0.138	0.555	1.1406	1.0337	-0.126	-114.774	0.00	1.00
postRoute	1	0.138	0.555	1.1318	1.0670	-0.180	-187.204
postRouteOpt	1	0.137	0.555	1.1296	1.0681	-0.130	-90.493

• (Updated on April 30, 2023) The following table and screenshots provide details of Ariane133-GF12 implementation when Hier-RTLMP is used to generate the macro placement.

Ariane133-GF12-68% Hier-RTLMP (results are normalized as described here)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1	0.138	0.555	1.0218	1.3219	-0.144	-307.690	0.00	3.5
postCTS	1	0.140	0.555	1.1657	1.3389	-0.169	-190.458	0.00	3.5
postRoute	1	0.140	0.555	1.1557	1.3772	-0.270	-289.089
postRouteOpt	1	0.139	0.555	1.1541	1.3785	-0.181	-178.470

• The following table and screenshots provide details of BlackParrot (Quad Core) GF12 implementation when CMP is used to generate the macro placement.

BlackParrot-GF12-68% CMP (results are normalized as described here)
Physical Design Stage	Core Area	Standard Cell Area	Macro Area	Total Power	Wirelength	WS	TNS	Congestion(H)	Congestion(V)
preCTS	1	0.176	0.501	1.0000	1.0000	0.001	0.000	1.00	1.00
postCTS	1	0.178	0.501	1.1526	1.0079	0.000	0.000	1.00	1.00
postRoute	1	0.178	0.501	1.1436	1.0304	-0.014	-2.629
postRouteOpt	1	0.178	0.501	1.1437	1.0306	0.001	0.000

• The following table and screenshots provide details of BlackParrot (Quad Core) GF12 implementation when CT is used to generate the macro placement.

BlackParrot-GF12-68% CT [results are normalized as described here] (Link to Tensorboard)
Physical Design Stage	Core Area	Standard Cell Area	Macro Area	Total Power	Wirelength	WS	TNS	Congestion(H)	Congestion(V)
preCTS	1	0.178	0.501	1.1068	1.6993	0.001	0.000	3.00	2.00
postCTS	1	0.179	0.501	1.2621	1.7058	0.000	0.000	2.00	2.20
postRoute	1	0.179	0.501	1.2469	1.7372	-0.028	-11.492
postRouteOpt	1	0.179	0.501	1.2462	1.7379	0.001	0.000

• (Updated on December 20) The following table and screenshots provide details of BlackParrot (Quad-Core) GF12 implementation when AutoDMP is used to generate the macro placement.

BlackParrot-GF12-68% AutoDMP [results are normalized as described here]
Physical Design Stage	Core Area	Standard Cell Area	Macro Area	Total Power	Wirelength	WS	TNS	Congestion (H)	Congestion (V)
preCTS	1	0.176	0.501	1.0012	0.9891	0.001	0.000	1.0	1.0
postCTS	1	0.178	0.501	1.1519	0.9967	0.000	0.000	1.0	1.2
postRoute	1	0.178	0.501	1.1433	1.0199	-0.045	-12.419
postRouteOpt	1	0.178	0.501	1.1433	1.0202	0.000	0.000

• The following table and screenshots provide details of MemPool Group GF12 implementation when CMP is used to generate the macro placement.

MemPool Group-GF12-68% CMP [results are normalized as described here ]
Physical Design Stage	Core Area	Standard Cell Area	Macro Area	Total Power	Wirelength	WS	TNS	Congestion(H)	Congestion(V)
preCTS	1	0.415	0.308	1.0000	1.0000	-0.154	-12479.05	1.00	1.00
postCTS	1	0.406	0.308	1.0663	1.0109	-0.134	-1828.60	1.07	1.26
postRoute	1	0.406	0.308	1.0631	1.0507	-0.213	-5882.00
postRouteOpt	1	0.405	0.308	1.0601	1.0521	-0.197	-1961.25

• The following table and screenshots provide details of MemPool Group GF12 implementation when CT is used to generate the macro placement.

MemPool Group-GF12-68% CT [results are normalized as described here ] (Link to Tensorboard)
Physical Design Stage	Core Area	Standard Cell Area	Macro Area	Total Power	Wirelength	WS	TNS	Congestion(H)	Congestion(V)
preCTS	1	0.419	0.308	1.1094	1.222	-0.170	-13620.25	1	1.22
postCTS	1	0.414	0.308	1.1966	1.2331	-0.179	-3615.65	1.27	1.57
postRoute	1	0.414	0.308	1.1987	1.2798	-0.178	-6350.95
postRouteOpt	1	0.410	0.308	1.1847	1.282	-0.195	-1849.40

• (Updated on December 21) The following macro placement is generated by Sayak Kundu based on the tile configuration received from Matheus Cavalcante, ETH Zürich and Jiantao Liu.

MemPool Group-GF12-68% human macro placement [results are normalized as described here]
Physical Design Stage	Core Area	Standard Cell Area	Macro Area	Total Power	Wirelength	WS	TNS	Congestion (H)	Congestion (V)
preCTS	1	0.418	0.308	1.033	1.084	-0.157	-12888.500	0.73	1.09
postCTS	1	0.409	0.308	1.105	1.093	-0.142	-2663.800	0.80	1.30
postRoute	1	0.409	0.308	1.103	1.136	-0.200	-4989.700
postRouteOpt	1	0.406	0.308	1.091	1.138	-0.149	-1766.450

(Updated on May 1, 2023)

We have tuned the timing constraints for the BlackParrot (Quad-Core) and MemPool Group designs on GF12. The results of different MacroPlacer solutions for the tuned designs are as follows:

• BlackParrot (Quad-Core)-GF12-68% CMP: The subsequent table and screenshots presents the post P\&R details of BlackParrot (Quad-Core) design on GF12 enablement when the macro placement is generated by CMP.

BlackParot-GF12-68% Innovus CMP [results are normalized as described here]
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1	0.188	0.498	1.000	1.000	-0.099	-230.148	1.00	1.00
postCTS	1	0.190	0.498	1.148	1.009	-0.080	-93.367	1.00	1.00
postRoute	1	0.190	0.498	1.138	1.033	-0.171	-1033.653
postRouteOpt	1	0.190	0.498	1.138	1.034	-0.087	-138.918

• BlackParrot (Quad-Core)-GF12-68% CT: The subsequent table and screenshots presents the post P\&R details of BlackParrot (Quad-Core) design on GF12 enablement when the macro placement is generated by CT.

BlackParrot-GF12-68% CT (wirelength cost: 0.0756, congestion cost: 0.7329, density cost: 0.6526, proxy cost: 0.7684) (Link to tensorboard)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1	0.190	0.498	1.083	1.568	-0.108	-244.624	2.00	1.80
postCTS	1	0.192	0.498	1.238	1.572	-0.087	-115.327	2.00	2.00
postRoute	1	0.192	0.498	1.223	1.605	-0.209	-270.951
postRouteOpt	1	0.191	0.498	1.219	1.606	-0.089	-66.473

• BlackParrot (Quad-Core)-GF12-68% SA: The subsequent table and screenshots presents the post P\&R details of BlackParrot (Quad-Core) design on GF12 enablement when the macro placement is generated by SA.

BlackParrot-GF12-68% SA (wirelength cost: 0.0576, congestion cost: 0.6619, density cost: 0.5971, proxy cost: 0.6871) [results are normalized as described here]
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1	0.189	0.498	1.030	1.239	-0.119	-234.785	1.00	1.40
postCTS	1	0.191	0.498	1.183	1.246	-0.111	-159.242	1.00	1.80
postRoute	1	0.191	0.498	1.171	1.274	-0.296	-4161.765
postRouteOpt	1	0.191	0.498	1.175	1.275	-0.160	-325.995

• BlackParrot (Quad-Core)-GF12-68% Human Expert: The subsequent table and screenshots presents the post P\&R details of BlackParrot (Quad-Core) design on GF12 enablement when the macro placement is generated by Huamn Expert.

BlackParot-GF12-68% Human Expert [results are normalized as described here]
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1	0.189	0.498	1.010	1.065	-0.107	-264.618	1.00	2.60
postCTS	1	0.190	0.498	1.157	1.074	-0.048	-40.525	2.00	3.20
postRoute	1	0.190	0.498	1.148	1.106	-0.266	-340.181
postRouteOpt	1	0.189	0.498	1.144	1.107	-0.049	-15.400

• BlackParrot (Quad-Core)-GF12-68% AutoDMP: The subsequent table and screenshots presents the post P\&R details of BlackParrot (Quad-Core) design on GF12 enablement when the macro placement is generated by AutoDMP (Nvidia).

BlackParot-GF12-68% AutoDMP [results are normalized as described here]
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1	0.189	0.498	1.005	1.008	-0.136	-254.904	1.00	1.00
postCTS	1	0.191	0.498	1.153	1.017	-0.076	-99.649	1.00	1.20
postRoute	1	0.191	0.498	1.143	1.043	-0.253	-361.892
postRouteOpt	1	0.190	0.498	1.140	1.043	-0.062	-61.772

• BlackParrot (Quad-Core)-GF12-68% Hier-RTLMP: The subsequent table and screenshots presents the post P\&R details of BlackParrot (Quad-Core) design on GF12 enablement when the macro placement is generated by Hier-RTLMP.

BlackParrot-GF12-68% Hier-RTLMP [results are normalized as described here]
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1	0.188	0.498	1.035	1.249	-0.100	-214.208	2.00	1.60
postCTS	1	0.190	0.498	1.188	1.257	-0.079	-102.866	1.00	1.80
postRoute	1	0.190	0.498	1.177	1.288	-0.213	-339.322
postRouteOpt	1	0.190	0.498	1.173	1.289	-0.082	-54.313

• MemPool Group-GF12-68% CMP: The subsequent table and screenshots presents the post P\&R details of MemPool Group design on GF12 enablement when the macro placement is generated by CMP.

MemPool Group-GF12-68% Innovus CMP [results are normalized as described here]
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1	0.412	0.312	1.000	1.000	-0.073	-4486.957	1.00	1.00
postCTS	1	0.403	0.312	1.056	1.007	-0.058	-196.767	1.00	1.00
postRoute	1	0.403	0.312	1.055	1.048	-0.126	-2495.000
postRouteOpt	1	0.393	0.312	1.025	1.051	-0.101	-167.530

• MemPool Group-GF12-68% CT: The subsequent table and screenshots presents the post P\&R details of MemPool Group design on GF12 enablement when the macro placement is generated by CT.

MemPool Group-GF12-68% CT (Wirelength cost: 0.069, Congestion cost: 0.810, Density Cost: 1.039, Proxy Cost: 0.994) (Link to tensorboard) [results are normalized as described here]
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1	0.416	0.312	1.085	1.189	-0.085	-5086.783	0.76	1.25
postCTS	1	0.409	0.312	1.153	1.196	-0.090	-578.565	0.73	1.33
postRoute	1	0.409	0.312	1.154	1.244	-0.196	-5010.696
postRouteOpt	1	0.400	0.312	1.124	1.247	-0.087	-124.331

• MemPool Group-GF12-68% SA: The subsequent table and screenshots presents the post P\&R details of MemPool Group design on GF12 enablement when the macro placement is generated by SA.

MemPool Group-GF12-68% SA (Wirelength cost: 0.064, Congestion cost: 0.940, Density Cost: 1.325, Proxy Cost: 1.196) [results are normalized as described here]
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1	0.415	0.312	1.081	1.187	-0.083	-5070.000	1.29	1.42
postCTS	1	0.408	0.312	1.138	1.197	-0.094	-415.182	1.32	1.52
postRoute	1	0.408	0.312	1.145	1.248	-0.149	-4161.478
postRouteOpt	1	0.403	0.312	1.130	1.250	-0.077	-262.988

• MemPool Group-GF12-68% Human Expert: The subsequent table and screenshots presents the post P\&R details of MemPool Group design on GF12 enablement when the macro placement is generated by Human Expert.

MemPool Group-GF12-68% Human Expert [results are normalized as described here]
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1	0.414	0.312	1.027	1.065	-0.081	-4820.478	0.48	1.00
postCTS	1	0.407	0.312	1.092	1.070	-0.062	-357.957	0.55	1.04
postRoute	1	0.407	0.312	1.091	1.113	-0.142	-3350.652
postRouteOpt	1	0.398	0.312	1.059	1.116	-0.075	-105.913

• MemPool Group-GF12-68% AutoDMP: The subsequent table and screenshots presents the post P\&R details of MemPool Group design on GF12 enablement when the macro placement is generated by AutoDMP (Nvidia).

MemPool Group-GF12-68% AutoDMP [results are normalized as described here]
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1	0.415	0.312	1.015	1.037	-0.105	-5260.304	1.00	1.13
postCTS	1	0.407	0.312	1.078	1.044	-0.104	-517.435	1.00	1.22
postRoute	1	0.407	0.312	1.077	1.089	-0.116	-3304.174
postRouteOpt	1	0.400	0.312	1.054	1.091	-0.103	-267.739

• MemPool Group-GF12-68% Hier-RTLMP: The subsequent table and screenshots presents the post P\&R details of MemPool Group design on GF12 enablement when the macro placement is generated by Hier-RTLMP.

MemPool Group-GF12-68% Hier-RTLMP [results are normalized as described here]
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1	0.411	0.312	1.031	1.086	-0.076	-4525.696	0.62	0.92
postCTS	1	0.405	0.312	1.100	1.095	-0.072	-394.957	0.68	1.04
postRoute	1	0.405	0.312	1.101	1.138	-0.139	-3301.739
postRouteOpt	1	0.397	0.312	1.074	1.140	-0.068	-94.348

An Observation regarding “Pure Commercial Flow”. The Evaluation Flow document also sheds light on the relative strength of a “Pure Commercial Flow”, as follows. CT uses the placement information generated by physical synthesis (Genus iSpatial). Observe that if we go straight into Evaluation Flow 1 from physical synthesis (without running CT), this will produce a “pure commercial flow” (i.e., CMP) outcome without any use of Circuit Training. From the data in the Evaluation Flow document, we see that with the “pure commercial flow”, CMP macro placements produce similar timing and power numbers compared to CT macro placements. However, the postRouteOpt wirelength of CT macro placements is at least 18% larger than the postRouteOpt wirelength of CMP macro placements.
Please note that we report this data as part of our study of Circuit Training. It is not intended to “benchmark” any commercial EDA tool in any sense, and the data should not be interpreted as providing any sort of “benchmarking” comparison or value judgment regarding the commercial tool.

November 27:
We have extended the experiment of Question 3 to assess the difficulty of our testcases. As mentioned here, we take the CT-generated macro placement and then randomly swap the same-size macros. We use the shuffle_macro.tcl script for this experiment. The following items provide details of the macro shuffling experiments for different testcases.

• Ariane: The target clock period of the shuffling experiment for Ariane133-NG45-68% shown here is 4ns, which is very relaxed (see here for clock period sweep results). Hence, we ran the same macro shuffling experiment for a tighter target clock period of 1.3ns. The following table shows the preCTS / postPlaceOpt and postRouteOpt metrics. We shuffled the macros using six different seed values of 111, 222, 333, 444, 555 and 666.
  • For the shuffled designs, the total power increases by 1.4%, the wirelength increases by 16%, and the runtime increases by 9% on average.

Ariane133-NG45-68%-1.3ns
Metrics	CT	Shuffle-111	Shuffle-222	Shuffle-333	Shuffle-444	Shuffle-555	Shuffle-666
Core_area (um^2)	1814274	1814274	1814274	1814274	1814274	1814274	1814274
Macro_area (um^2)	1018356	1018356	1018356	1018356	1018356	1018356	1018356
preCTS_std_cell_area (um^2)	243264	246309	243426	246181	247134	243731	246412
postRouteOpt_std_cell_area (um^2)	244002	250080	246325	249506	249494	246242	247918
preCTS_total_power (mw)	789.871	802.369	796.562	803.034	801.677	794.323	802.673
postRouteOpt_total_power (mw)	828.747	845.726	836.735	844.61	843.227	837.434	838.833
preCTS_wirelength (um)	4727728	5515599	5547501	5489654	5508653	5448399	5549232
postRouteOpt_wirelength (um)	4893776	5690000	5712986	5667587	5687840	5628320	5724530
preCTS_WS (ns)	-0.091	-0.112	-0.109	-0.141	-0.144	-0.095	-0.151
postRouteOpt_WS (ns)	-0.079	-0.091	-0.099	-0.106	-0.157	-0.048	-0.108
preCTS_TNS (ns)	-110.373	-136.145	-136.781	-197.545	-196.557	-96.462	-210.187
postRouteOpt_TNS (ns)	-25.762	-66.855	-86.119	-81.177	-159.035	-16.386	-75.133
preCTS_Congestion (H)	0.03%	0.04%	0.05%	0.05%	0.04%	0.04%	0.05%
preCTS_Congestion (V)	0.12%	0.12%	0.15%	0.12%	0.12%	0.10%	0.10%
Runtime (second)	3451	3786	3427	3591	3748	3851	3994

• BlackParrot (Quad-Core): We have performed a similar macro shuffling experiment for the BlackParrot (Quad-Core) design. The following table shows the preCTS / postPlaceOpt and postRouteOpt metrics. We shuffled the macros using six different seed values of 111, 222, 333, 444, 555 and 666.
  • For the shuffled designs, the total power increases by 6%, the wirelength increases by 33%, and the runtime increases by 16% on average.

BlackParrot (Quad-Core)-NG45-68%-1.3ns (bp_clk)
Metrics	CT	Shuffle-111	Shuffle-222	Shuffle-333	Shuffle-444	Shuffle-555	Shuffle-666
core_area (um^2)	8449457	8449457	8449457	8449457	8449457	8449457	8449457
macro_area (um^2)	3917822	3917822	3917822	3917822	3917822	3917822	3917822
preCTS_std_cell_area (um^2)	1954954	1985365	1986378	1985226	1984435	1988719	1991871
postRouteOpt_std_cell_area (um^2)	1978731	2008143	2037502	2033273	2014517	2027724	2016049
preCTS_total_power (mw)	4329.795	4604.961	4619.481	4608.242	4591.569	4632.783	4620.598
postRouteOpt_total_power (mw)	4685.509	4959.629	5004.988	4998.899	4959.435	5005.635	4977.157
preCTS_wirelength (um)	39101445	51131110	51444279	52030185	52035717	53176682	51997133
postRouteOpt_wirelength (um)	40467467	53098209	53425737	54070974	54030437	55365255	54171082
preCTS_WS (ns)	-0.220	-0.228	-0.193	-0.205	-0.199	-0.217	-0.222
postRouteOpt_WS (ns)	-0.260	-0.179	-0.305	-0.342	-0.211	-0.289	-0.251
preCTS_TNS (ns)	-1385.900	-1105.900	-826.103	-912.903	-1116.400	-944.540	-1065.400
postRouteOpt_TNS (ns)	-3657.000	-835.927	-6542.400	-8738.100	-1816.000	-3548.600	-1322.200
preCTS_Congestion (H)	0.21%	0.52%	0.71%	0.64%	0.62%	0.53%	0.66%
preCTS_Congestion (V)	0.29%	0.54%	0.44%	0.50%	0.45%	0.68%	0.57%
Runtime (second)	22367	26089	25940	25293	24745	32431	31591

• MemPool Group: We have tried a similar macro shuffling experiment for MemPool Group, but none of our runs completed (i.e., flow failure).

December 20:
We thank NVIDIA Research for access to AutoDMP, an autotuned DREAMPlace-based macro placer that will be reported at ISPD-2023. We have generated macro placements of Ariane and BlackParrot using AutoDMP, in both NG45 and GF12 enablements. The results are as follows:

• Ariane133-NG45-68%-1.3ns: Following table and screenshots show the macro placement result of Ariane133 on NG45, generated using AutoDMP.

Ariane133-NG45-68%-1.3ns AutoDMP (Link to CT result) (Link to CMP result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	243431	1018356	783.810	3604121	-0.105	-140.503	0.00%	0.01%
postCTS	1814274	243612	1018356	821.621	3630937	-0.097	-47.167	0.03%	0.15%
postRoute	1814274	243612	1018356	821.558	3759529	-0.102	-75.677
postRouteOpt	1814274	243720	1018356	821.654	3763817	-0.095	-37.496

• Ariane133-GF12-68%: Link to AutoDMP macro placement details of Ariane on GF12 enablement.

• BlackParrot-NG45-68%-(bp clock)1.3ns: Following table and screenshots show the macro placement result of BlackParrot (Quad-Core) on NG45, generated using AutoDMP.

BlackParrot Quad-Core-NG45-68%-1.3ns AutoDMP (Link to CT result) (Link to CMP result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	8449457	1903521	3917822	4069.801	22483473	-0.183	-584.774	0.02%	0.07%
postCTS	8449457	1916465	3917822	4438.356	22616243	-0.145	-288.267	0.05%	0.09%
postRoute	8449457	1916465	3917822	4434.782	23349968	-0.195	-2164.900
postRouteOpt	8449457	1920024	3917822	4438.571	23376406	-0.190	-1183.100

• BlackParrot-GF12-68%: Link to AutoDMP macro placement details of BlackParrot on GF12 enablement.

December 21:
Question 11. How does the initial placement generated by different physical synthesis tools affect the CT solution?

We observe that whether the initial placement solution is generated using Flow-2 (CMP-Genus iSpatial) or the initial placement is generated by DC-Topo (links to scripts), the final CT outcomes are similar.

The following table and screenshots provide details of Ariane133-NG45-68%-1.3ns CT macro placement when DC-Topo is used to generate the initial placement solution.

Ariane133-NG45-68%-1.3ns CT result when the initial placement information is generated by Synopsys DC-Topo physical synthesis.
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	284197	1018356	815.500	4544323	-0.155	-261.254	0.02%	0.17%
postCTS	1814274	286795	1018356	858.088	4599954	-0.146	-118.845	0.02%	0.20%
postRoute	1814274	286795	1018356	857.217	4705640	-0.203	-302.019
postRouteOpt	1814274	287151	1018356	857.755	4710065	-0.206	-255.818

Link to result of Ariane133-NG45-68%-1.3ns CT macro placement when Flow-2 (CMP-Genus iSpatial physical synthesis) is used to generate the initial placement information.

Question 12. How well does Simulated Annealing (SA) optimize the proxy cost?
Details of our SA implementation, which we denote as SA-UCSD, are here. We have used SA-UCSD to generate macro placements for Ariane and BlackParrot (Quad-Core). We find that SA-UCSD produces better proxy costs than CT.

• Ariane133-NG45-68%-1.3ns: The configuration that results best proxy cost (wirelength cost: 0.0881, congestion cost: 0.8257, density cost: 0.5084, proxy cost: 0.75515): action_probs: [0.2, 0.2, 0.2, 0.2, 0.2], num_actions: 3, max_temperature: 7e-5, num_iters: 50000, seed: 1, spiral_flag: True
  • The following table and screenshots provide details of Ariane133-NG45-68%-1.3ns SA-UCSD macro placement.

Ariane133-NG45-68%-1.3ns SA-UCSD result (Link to CT result) (Link to CMP result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	243604	1018356	786.182	3825529	-0.130	-187.073	0.01%	0.03%
postCTS	1814274	245443	1018356	827.698	3868208	-0.099	-52.565	0.02%	0.06%
postRoute	1814274	245443	1018356	827.546	3982401	-0.125	-114.924
postRouteOpt	1814274	245804	1018356	828.053	3986262	-0.112	-75.338

• BlackParrot (Quad-Core)-NG45-68%-1.3ns: The configuration that results best proxy cost (wirelength cost: 0.0604, congestion cost: 0.9581, density cost: 0.7383, proxy cost: 0.90860): action_probs: [0.2, 0.2, 0.2, 0.2, 0.2], num_actions: 1, max_temperature: 10e-5, num_iters: 20000, seed: 1, spiral_flag: False
  • The following table and screenshots provide details of BlackParrot (Quad-Core)-NG45-68%-1.3ns SA-UCSD macro placement.

BlackParrot Quad-Core-NG45-68%-(bp clock)1.3ns SA-UCSD (Link to CT result) (Link to CMP result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	8449457	1921810	3917822	4185.031	30470310	-0.209	-863.535	0.08%	0.32%
postCTS	8449457	1934844	3917822	4560.519	30568687	-0.107	-267.191	0.09%	0.36%
postRoute	8449457	1934844	3917822	4539.416	31510301	-0.239	-6022.700
postRouteOpt	8449457	1943841	3917822	4547.886	31550599	-0.222	-3263.800

Question 13. How good are human macro placements relative to Circuit Training?
We observe that human macro placements can achieve smaller wirelength than CT, with similar timing and power numbers. Details of human macro placements for BlackParrot (Quad-Core) and MemPool Group on NG45 enablement are as follows:

• BalckParrot-NG45-68%-1.3ns: We thank Dr. Jinwook Jung of IBM Research for providing his human macro placement of BlackParrot Quad-Core design as an alternative baseline. The following table and screenshots provide details of BlackParrot (Quad-Core)-NG45-68%-1.3ns human macro placement. Link to the script.
  • Dr. Jung informed us that he spent about 0.5 hours learning about the design, 2.5 hours coming up with initial floorplan scripts, and an additional 2.5 hours refining the initial version, for a total of 5.5 hours of effort. Dr. Jung also informed us that his floorplan design includes 4 identical tiles, and that these are arranged so as to create more free space.

BlackParrot Quad-Core-NG45-68%-1.3ns Human macro placement (not a gridded placement) (Link to CT result) (Link to CMP result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	8449457	1907164	3917822	4107.931	24814112	-0.195	-530.552	0.08%	0.12%
postCTS	8449457	1918983	3917822	4475.523	24944903	-0.097	-209.587	0.09%	0.13%
postRoute	8449457	1918983	3917822	4468.904	25888999	-0.120	-454.561
postRouteOpt	8449457	1919928	3917822	4469.552	25915520	-0.097	-321.918

• MemPool Group-NG45-68%-4ns: The following macro placement is generated by Sayak Kundu based on the tile configuration received from Matheus Cavalcante, ETH Zürich and Jiantao Liu. Link to the MemPool Group macro placement script. The following table and screenshots provide details of MemPool Group-NG45-68%-4ns human macro placement.

MemPool Group-NG45-68%-4ns human macro placement (not a gridded placement) (Link to CT result) (Link to CMP result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	11371934	4930345	3078071	2459.392	101645170	-0.021	-141.801	0.39%	0.86%
postCTS	11371934	4883741	3078071	2640.242	102110339	-0.003	-0.055	0.58%	0.96%
postRoute	11371934	4883741	3078071	2642.017	107463344	-0.246	-2941.400
postRouteOpt	11371934	4873872	3078071	2639.916	107597894	-0.049	-11.897

We have also added

• Ariane133-NG45-68%-1.3ns: Link to the human macro placement details of Ariane on NG45 enablement.
• MemPool Group-GF12-68%: Link to the human macro placement details of MemPool Group on GF12 enablement.

March 5:
Question 14. What is the impact on CT results when DREAMPlace is used instead of force-directed placement?

We have integrated DREAMPlace in Circuit Training (commit hash: 91e14fd1caa5b15d9bb1b58b6d5e47042ab244f3) and trained CT to generate macro placement solutions for Ariane, BlackParrot and MemPool Group designs. We referer to CT with DREAMPlace as CT+DREAMPlace and CT with FD as CT+FD. The training results are as follows:

• Ariane133-NG45-68%-1.3ns: Following table and screenshots presents the macro placement solution generated by CT+DREAMPlace for Ariane133 design with 68% floorplan utilization, 1.3ns target clock period on NG45 enablement. (Wirelength Cost:0.0678, Congestion cost: 0.8320, Density cost: 0.5239)

Ariane133-NG45-68%-1.3ns CT+DREAMPlace result (Link to tensorboard) (Link to CT+FD result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	244313	1018356	791.482	4669338	-0.135	-176.306	0.05%	0.12%
postCTS	1814274	244976	1018356	830.645	4693972	-0.106	-75.708	0.05%	0.15%
postRoute	1814274	244976	1018356	828.923	4822561	-0.124	-109.91
postRouteOpt	1814274	245438	1018356	829.353	4827641	-0.126	-93.752

• BlackParrot(Quad-Core)-NG45-68%-1.3ns: Following table and screenshots presents the macro placement solution generated by CT+DREAMPlace for BlackParrot design with 68% floorplan utilization, 1.3ns target clock period on NG45 enablement. (Wirelength cost: 0.0878, Density cost: 0.5687, Congestion cost: 1.1420)

BP(Quad-Core)-NG45-68%-1.3ns CT+DREAMPlace (Link to tensor board) (Link to CT+FD result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	8449457	1959789	3917822	4396.086	42267061	-0.209	-1132.2	0.28%	0.57%
postCTS	8449457	1978100	3917822	4783.785	42346079	-0.163	-680.8	0.29%	0.63%
postRoute	8449457	1978100	3917822	4751.075	43883402	-0.201	-1406.3
postRouteOpt	8449457	1979794	3917822	4753.696	43931174	-0.178	-850.8

• MemPool Group-NG45-68%-1.3ns: Following table and screenshots presents the macro placement solution generated by CT+DREAMPlace for MemPool Group design with 68% floorplan utilization, 4ns target clock period on NG45 enablement. (Wirelength cost: 0.0728, Density cost: 0.6617, Congestion cost: 1.2714) DRC Count: 14779.

MemPool Group-NG45-68%-4ns CT+DREAMPlace (Link to tensorboard) (Link to CT+FD Result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	11371934	4990302	3078071	2659.403	121635791	-0.015	-71.824	3.33%	3.26%
postCTS	11371934	4969651	3078071	2839.139	122062712	-0.004	-0.104	3.49%	3.19%
postRoute	11371934	4969651	3078071	2893.588	132078512	-1.137	-29243.4
postRouteOpt	11371934	4995348	3078071	2908.959	132299696	-0.072	-97.892

Question 15. Should we factor in density cost while using DREAMPlace for CT?

We update the density weight from 0.5 to 0.0, then rerun CT-DREAMPlace for Ariane, BlackParrot and MemPool Group designs. The training results are as follows:

• Ariane133-NG45-68%-1.3ns: Following table and screenshots presents the macro placement solution generated by CT+DREAMPlace for Ariane133 design with 68% floorplan utilization, 1.3ns target clock period on NG45 enablement when density weight is 0. (Wirelength Cost: 0.0715, Congestion cost: 0.8111, Density cost: 0.5251)

Ariane133-NG45-68%-1.3ns CT+DREAMPlace result (Density Weight = 0.0) (Link to tensorboard) (Link to CT+FD result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	245097	1018356	793.171	4959656	-0.137	-202.147	0.04%	0.17%
postCTS	1814274	248172	1018356	839.062	4993255	-0.117	-108.074	0.04%	0.15%
postRoute	1814274	248172	1018356	836.985	5114089	-0.164	-243.834
postRouteOpt	1814274	248775	1018356	837.655	5119513	-0.16	-152.043

• BlackParrot(Quad-Core)-NG45-68%-1.3ns: Following table and screenshots presents the macro placement solution generated by CT+DREAMPlace for BlackParrot design with 68% floorplan utilization, 1.3ns target clock period on NG45 enablement when density weight is 0. (Wirelength cost: 0.0791, Density cost: 0.5770, Congestion cost: 1.0964)

BP(Quad-Core)-NG45-68%-1.3ns CT+DREAMPlace (Density weight = 0.0) (Link to tensorboard) (Link to CT+FD result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	8449457	1947589	3917822	4323.518	38208933	-0.233	-1177.6	0.33%	0.46%
postCTS	8449457	1961564	3917822	4703.800	38314312	-0.153	-468.3	0.37%	0.49%
postRoute	8449457	1961564	3917822	4674.250	39753854	-0.200	-1995.5
postRouteOpt	8449457	1964239	3917822	4677.048	39800843	-0.180	-809.0

• MemPool Group-NG45-68%-1.3ns: Following table and screenshots presents the macro placement solution generated by CT+DREAMPlace for MemPool Group design with 68% floorplan utilization, 4ns target clock period on NG45 enablement when density weight is 0. (Wirelength cost: 0.0711, Density cost: 0.6666, Congestion cost: 1.2605 ) DRC Count: 3260

MemPool Group-NG45-68%-4ns CT+DREAMPlace (Density weight = 0.0) (Link to tensorboard) (Link to CT+FD Result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	11371934	4934839	3078071	2613.613	119923841	-0.027	-146.5	2.56%	2.51%
postCTS	11371934	4928559	3078071	2802.851	120508367	-0.003	-0.1	2.87%	2.66%
postRoute	11371934	4928559	3078071	2848.873	130024068	-0.803	-19920.7
postRouteOpt	11371934	4953483	3078071	2858.071	130243153	-0.050	-33.5

We observe from the above results that CT+DREAMPlace achieves similar result for density weight 0 and 0.5.

Question 16. Why does your study (and, ISPD-2023 paper) use Cadence CMP 21.1, which was not available to Google engineers when they wrote the Nature paper?

We used Innovus version 21.1 since it was the latest version of our place-and-route evaluator of macro placement solutions. CMP 21.1 is part of Innovus 21.1. Using the latest version of CMP was also natural, given our starting assumption that RL from Nature would outperform the commercial state-of-the-art.

We have now run further experiments using older versions of CMP and Innovus. We find that the macro placements produced by CMP across versions 19.1, 20.1 and 21.1 lead to the same qualitative conclusions. Additional details:

• The Concurrent Macro Placer (CMP) was available in both the 19.1 and 20.1 versions of Cadence Innovus. Our published flow scripts can also be used to run Innovus 19.1 and 20.1 with a few lines commented out: lines1 and lines2.
• For the Ariane133-NG45-68%-1.3ns testcase, we have run CMP + Innovus in two additional Cadence releases (19.1, 20.1). This corresponds to Steps “4” and “5” of the industrial evaluation flow in Figure 2 of our paper, and a “pure commercial tool flow”.
• We assess the CT macro placement that is reported in Table 1 of our ISPD-2023 paper, using all three Innovus P&R versions. The CT post-P&R results are inferior to those obtained with corresponding CMP versions.
• This new study reinforces the conclusion obtained using CMP + Innovus (21.1) in our paper. This can be independently verified using provided scripts. We do not provide additional numbers, in order to avoid benchmarking of the Cadence tool versions.

Below are screenshots of Ariane-NG45-68%-1.3ns for (in order, top-down) CMP + P&R outcomes in Innovus 19.1, 20.1 and 21.1 versions.

• Ariane133-NG45-68%-1.3ns (CMP + Innovus 19.1)

• Ariane133-NG45-68%-1.3ns (CMP + Innovus 20.1)

• Ariane133-NG45-68%-1.3ns (CMP + Innovus 21.1 is the same as in Figure 3 of our paper)

• Left to right: CT macro placement from the ISPD-2023 paper, with P&R using Innovus 19.1, 20.1 and 21.1. (21.1 is the same as in Figure 3 of our paper.)

Question 17. What are the outcomes of CT when the training is continued until convergence?

To put this question in perspective, training “until convergence” is not described in any of the guidelines provided by the CT GitHub repo for reproducing the results in the Nature paper. For the ISPD 2023 paper, we adhere to the guidelines given in the CT GitHub repo, use the same number of iterations for Ariane as Google engineers demonstrate in the CT GitHub repo, and obtain results that closely align with Google’s outcomes for Ariane. (See FAQs #4 and #13.)

We run CT training for an extended number (=600) of iterations, for each of Ariane, BlackParrot and MemPool Group on NG45, and make the following observations.

• For Ariane the proxy cost improves from 0.857 to 0.809 (link to the new tensorboard). However, the Nature Table 1 metrics are very similar: routed wirelength improves from 4,894mm to 4,739mm; Total power degrades from 828.7 mW to 829.4 mW; worst negative slack and total negative slack respectively degrade from -79ps to -85ps, and from -25.8ns to -62.7ns. The final proxy cost and the Nature Table 1 metrics achieved through training until convergence are still not better than those achieved by SA.
• For BlackParrot, the proxy cost improves significantly from 1.021 to 0.889 (link to new tensorboard). Routed wirelength improves significantly from 36,845mm to 30,929mm. Also total power improves from 4627.4mW to 4547.8mW. However, the worst negative slack and total negative slack respectively degrade from -185ps to -199ps, and from -1040.8ns to -1263.4ns. The final proxy cost achieved by CT is better than that achieved by SA. The Nature Table 1 metrics are still similar to those achieved by SA.
• For MemPool Group, CT diverges, and it never converges. Thus, the final proxy cost is unchanged. Here is the link to tensorboard. So, the CT code does not guarantee full convergence.
• Note 1: We have not studied what happens if SA is given triple the runtime used in our previously-reported experiments.
• Note 2: Our new data underscore the poor correlation between proxy cost and ground-truth metrics noted in Section 5.2.3 of the ISPD-2023 paper.

Our new data from using triple the CT training budget indicate that training until convergence, compared to the configurations explored in the ISPD-2023 paper, improves proxy cost but does not significantly improve chip metrics on Ariane and MemPool Group. Among chip metrics for BlackParrot, routed wirelength improves significantly while other metrics are similar to what we previously reported. Overall, training until convergence does not qualitatively change comparisons to results of Simulated Annealing and human macro placements reported in the ISPD 2023 paper.

The subsequent tables and figures present the Nature Table 1 metrics of Ariane and BlackParrot on NG45, for macro placement solutions generated by CT training until convergence. (For MemPool Group, using triple the default number of CT iterations did not change the final proxy cost.)

Ariane133-NG45-68%-1.3ns CT result (Link to tensorboard)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion(H)	Congestion (V)
preCTS	1814274	242539	1018356	787.798	4577259	-0.095	-121.911	0.04%	0.11%
postCTS	1814274	244220	1018356	830.273	4610696	-0.07	-41.635	0.05%	0.13%
postRoute	1814274	244220	1018356	828.935	4734768	-0.095	-90.160
postRouteOpt	1814274	244666	1018356	829.419	4739136	-0.085	-62.685

BlackParrot (Quad-Core)-NG45-68%-1.3ns CT result (Link to tensorboard)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	8449457	1922798	3917822	4185.939	29820259	-0.179	-648.911	0.10%	0.26%
postCTS	8449457	1935706	3917822	4563.875	29956480	-0.138	-355.347	0.12%	0.28%
postRoute	8449457	1935706	3917822	4542.299	30893195	-0.188	-2280.100
postRouteOpt	8449457	1940957	3917822	4547.832	30928844	-0.199	-1263.400

Question 18. To study the benefit that CT derives from use of a commercial placement solution, why do you compare with giving CT “impossible” initial placements, where all instances are placed at the same location?

• Section 5.2.1 of our ISPD-2023 paper discusses the advantage that CT derives from its use of initial placement information from a commercial EDA tool. To measure this advantage, we study what happens when CT is deprived of this placement information.
• In Question 1, August 2022 we used “vacuous” placements where the same (x,y) location is given for all instances. This corresponds to the use of placements that have as little information content as possible. However, after publication of our ISPD-2023 paper, comments were made that such placements are “impossible”.
• We have now performed a second study that gradually perturbs the EDA tool’s placement and measures the effect on CT outcomes. In this second study, we always maintain legal placements: every placement that is fed to CT is “possible”. Our new study directly assesses how CT’s performance changes as the commercial EDA tool’s placement is degraded.
  • Note 1: CT’s grouping flow requires (x,y) coordinates in the input.
  • Note 2: We cannot use a “random, but possible” placement as input to CT. This leads to a blowup of the numbers of clusters and edges in the adjacency matrix. [E.g.: “IndexError: index 3500 is out of bounds for axis 0 with size 3500” from CT. There is also a default limit of 42000 edges in CT.]
• The gen_perturbed_placement procedure below randomly perturbs the original placement solution from commercial physical synthesis, by shuffling the placed locations of a prescribed fraction of instances in the design. (E.g., when the parameter x = 0.05, the locations of 5% of the netlist will be shuffled.)

Procedure gen_perturbed_placement
Input: seed, x

# x indicates the fraction of instances to be moved 0 < x < 1.0
1. For w, h in {unique list of instance (width, height)}
  a. instance_list = {list of instances with width = w and height = h}
  b. instance_list = shuffle(instance_list, seed)
  c. instance_count = length(shuffled_instance_list)
  d. shuffled_instance_list = instance_list[:instance_count*x]
  e. shuffle_placement(shuffled_instance_list, seed)

Procedure shuffle_placement
Input: instance_list, seed

1. X, Y, Orient = {list of lower left coordinate and orientation of instances in the instance_list}
2. shuffled_instance_list = shuffle(instance_list, seed)
3. For i in range(length(instance_list)):
  a. Update location and orientation of shuffled_instance_list[i] with (X[i], Y[i]) and Orient[i]

• The table below shows what happens as the commercial EDA tool’s “possible” initial placement is degraded into other “possible” initial placements, for all combinations of x = {0.01, 0.05, 0.15} and seed = {21, 42, 63}. The value x = 0.0 corresponds to the CT outcome that we report in Table 1 of our ISPD-2023 paper. We include the “Human” and “SA” rows from our Table 1 for ease of reference.
• From the data, we observe that degrading the commercial placement information worsens all CT outcomes except for routed wirelength across all seed values. Runtime is also worsened, e.g., with x = 0.15 the CT runtime in our environment was 52.0 hours which is 1.6 times longer than when x = 0.0 (See #13 of our FAQs.). This is at least in part because having more moving elements (soft macros) increases CT’s runtime in force-directed placement and proxy cost evaluation.
• For the nine perturbed placements, SA yields better proxy cost and chip metrics compared to CT in most cases.
  • Note 3: We have not studied what happens if SA is given 1.6 times the runtime used in our previously-reported experiments.

April 27, 2023:
We have run Hier-RTLMP macro placer, as described in the arXiv paper, on our modern benchmarks. The code for Hier-RTLMP is open-sourced here. We use the default settings to generate the macro placement solutions. The results are as follows:

• Ariane133-NG45-68%-1.3ns: Following table and screenshots show the macro placement result of Ariane133 on NG45, generated using Hier-RTLMP.

Ariane133-NG45-68%-1.3ns Hier-RTLMP (Link to CT result) (Link to CMP result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	1814274	246916	1018356	796.781	5087055	-0.149	-192.7	0.11%	0.08%
postCTS	1814274	247403	1018356	836.595	5136058	-0.110	-104.2	0.15%	0.10%
postRoute	1814274	247403	1018356	835.096	5291106	-0.178	-356.0
postRouteOpt	1814274	248296	1018356	836.002	5296879	-0.165	-223.4

• Ariane133-GF12-68%: Link to the HierRTLMP macro placement details of Ariane on GF12 enablement.

• BlackParrot (Quad-Core)-NG45-68%-1.3ns: Following table and screenshots show the macro placement result of BlackParrot (Quad-Core) on NG45, generated using Hier-RTLMP.

BlackParrot-NG45-68%-1.3ns Hier-RTLMP (Link to CT result) (Link to CMP result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	8449457	1908372	3917822	4148.534	27687847	-0.169	-455.5	0.13%	0.17%
postCTS	8449457	1923367	3917822	4522.966	27810361	-0.123	-181.5	0.15%	0.20%
postRoute	8449457	1923367	3917822	4509.596	28835670	-0.166	-906.8
postRouteOpt	8449457	1925012	3917822	4511.780	28865504	-0.150	-456.6

• BlackParrot (Quad-Core)-GF12-68%: Link to the HierRTLMP macro placement details of BlackParrot (Quad-Core) on GF12 enablement.

• MemPool Group-NG45-68%-4ns: Following table and screenshots show the macro placement result of MemPool Group on NG45, generated using Hier-RTLMP.

MemPool Group-NG45-68%-4ns Hier-RTLMP (62 DRCs) (Link to CT result) (Link to CMP result)
Physical Design Stage	Core Area (um^2)	Standard Cell Area (um^2)	Macro Area (um^2)	Total Power (mW)	Wirelength (um)	WS (ns)	TNS (ns)	Congestion (H)	Congestion (V)
preCTS	11371934	4939447	3078071	2489.1	105739299	-0.016	-50.5	2.05%	1.03%
postCTS	11371934	4895581	3078071	2671.4	106267958	-0.002	-0.1	2.31%	1.18%
postRoute	11371934	4895581	3078071	2696.2	113924593	-0.503	-4743.7
postRouteOpt	11371934	4889459	3078071	2695.3	114073113	-0.062	-4.9

• MemPool Group-GF12-68%: Link to the HierRTLMP macro placement details of MemPool Group on GF12 enablement.

Pinned (to bottom) question list:

Question 1. How does having an initial set of placement locations (from physical synthesis) affect the (relative) quality of the CT result?
Question 2. How does utilization affect the (relative) performance of CT?
Question 3. Is a testcase such as Ariane-133 “probative”, or do we need better testcases?
Question 4. How much does the guidance to clustering that comes from (x,y) locations matter?
Question 5. What is the impact of the Coordinate Descent (CD) placer on proxy cost and Table 1 metric?
Question 6. Are we using the industry tool in an “expert” manner? (We believe so.)
Question 7. What happens if we skip CT and continue directly to standard-cell P&R (i.e., the Innovus 21.1 flow) once we have a macro placement from the commercial tool?
Question 8. How does the tightness of timing constraints affect the (relative) performance of CT?
Question 9. Are CT results stable? If not, how much does the outcome vary?
Question 10. What is the correlation between proxy cost and the postRouteOpt Table 1 metrics?
Question 11. How does the initial placement generated by different physical synthesis tools affect the CT solution?
Question 12. How well does Simulated Annealing (SA) optimize Circuit Training’s proxy cost?
Question 13. How good are human macro placements relative to Circuit Training?
Question 14. What is the impact on CT results when DREAMPlace is used instead of force-directed placement?
Question 15. Should we factor in density cost while using DREAMPlace for CT?
Question 16. Why does your study (and, ISPD-2023 paper) use Cadence CMP 21.1, which was not available to Google engineers when they wrote the Nature paper?
Question 17. What are the outcomes of CT when the training is continued until convergence?
Question 18. To study the benefit that CT derives from use of a commercial placement solution, why do you compare with giving CT “impossible” initial placements, where all instances are placed at the same location?

This site is open source. Improve this page.