← back

Macro Placement

Chip macro placement that beat RePlAce by 30%

Python · PyTorchcode ↗

NOTE

The placer module is located in submissions/idk/placer.py.

Partcl/HRT Macro Placement Challenge

Hudson River Trading        Partcl

Win $20,000 by developing better macro placement algorithms!

Partcl and Hudson River Trading are excited to co-host a competition to solve the macro placement problem.

About Macro Placement

Macro placement is the problem of positioning large fixed-size blocks (SRAMs, IPs, analog macros, etc.) on a chip floorplan so that routing congestion, timing, power delivery, and area constraints are balanced. Unlike standard-cell placement, macros have strong geometric and connectivity constraints, so the challenge is to explore a highly discrete design space while minimizing wirelength, avoiding blockages, and preserving downstream routability and timing quality.

For example, the ibm01 benchmark has:

Simulated annealing on ibm01
Force-directed placement on ibm01

About HRT Hardware

Hudson River Trading (HRT) is a leading quantitative trading firm at the forefront of technical innovation in global financial markets.

HRT’s Hardware team builds the high-performance compute systems at the core of our trading infrastructure. We use FPGAs and ASICs to drive low-latency decision-making and power custom solutions across the trading stack, from bespoke circuits to machine learning accelerators.

We’re proud to sponsor this competition because advancing macro placement and low-level hardware optimization directly aligns with the kinds of hard, performance-critical engineering challenges our teams tackle every day.

Joining Hudson River Trading’s hardware team means working alongside leading engineers in one of the most advanced computing environments in global finance. Learn more about open roles at hudsonrivertrading.com.

About Partcl

Partcl is rebuilding chip design infrastructure from the ground up for the GPU era.

Modern chip design is slow, fragmented, and fundamentally constrained by tools built decades ago. Critical workflows like timing analysis and placement still take hours to days - limiting how much engineers can explore and optimize.

We’re changing that.

Partcl develops GPU-accelerated systems for physical design that run orders of magnitude faster than legacy tools. Our goal is simple: make iteration cheap enough that design space exploration becomes the default, not the exception.

Background Papers

[1] An Updated Assessment of Reinforcement Learning for Macro Placement

[2] Assessment of Reinforcement Learning for Macro Placement

[3] Reevaluating Google's Reinforcement Learning for IC Macro Placement

[4] A graph placement methodology for fast chip design

🏆 Prizes

For full Grand Prize scoring rules, feasibility gate, tie-breaking, and ORFS-failure handling, see SCORING.md.

Submission Format

Additional Rules

Allowed

Not Allowed

Evaluation Details

Evaluation is two-tiered:

Tier 1: Proxy Cost Ranking (All Submissions)

All submissions are ranked by proxy cost across the 17 IBM benchmarks (ibm01–ibm04, ibm06–ibm18). This is the primary qualifying metric. Proxy cost is computed using the TILOS MacroPlacement evaluator:

Proxy Cost = 1.0 × Wirelength + 0.5 × Density + 0.5 × Congestion

Baseline numbers are from: An Updated Assessment of Reinforcement Learning for Macro Placement

Tier 2: OpenROAD Flow Validation (Top Submissions)

The top 7 submissions by proxy score will be evaluated through the full OpenROAD flow on NG45 designs to measure real PnR outcomes: WNS, TNS, and Area.

🚀 Quick Start

Installation

# Clone the repository
git clone https://github.com/partcleda/partcl-macro-place-challenge.git
cd partcl-macro-place-challenge

# Initialize TILOS MacroPlacement submodule (required for evaluation)
git submodule update --init external/MacroPlacement

# Install the package and all dependencies
uv sync

# Verify the setup
uv run evaluate submissions/examples/greedy_row_placer.py -b ibm01

Run Your First Example

# Run the greedy row placer on ibm01
uv run evaluate submissions/examples/greedy_row_placer.py

# Run on all 17 IBM benchmarks
uv run evaluate submissions/examples/greedy_row_placer.py --all

# Run on NG45 commercial designs (ariane133, ariane136, mempool_tile, nvdla)
uv run evaluate submissions/examples/greedy_row_placer.py --ng45

# Visualize the result
uv run evaluate submissions/examples/greedy_row_placer.py --vis
uv run evaluate submissions/examples/greedy_row_placer.py --all --vis

Running on all benchmarks produces a summary like:

Benchmark     Proxy        SA   RePlAce     vs SA  vs RePlAce  Overlaps
   ibm01    2.0463    1.3166    0.9976    -55.4%     -105.1%         0
   ibm02    2.0431    1.9072    1.8370     -7.1%      -11.2%         0
   ...
     AVG    2.2109    2.1251    1.4578     -4.0%      -51.7%         0

The greedy placer achieves zero overlaps but makes no attempt to optimize wirelength or connectivity — your job is to do better! See SETUP.md for the full API reference and submissions/examples/ for working examples.

🎯 IBM Benchmark Suite (ICCAD04)

We evaluate on the complete ICCAD04 IBM benchmark suite:

BenchmarkMacrosNetsCanvas (μm)Area Util.SA BaselineRePlAce Baseline
ibm012467,26922.9×23.042.8%1.31660.9976 ⭐
ibm022547,53823.2×23.543.1%1.90721.8370 ⭐
ibm032698,04524.1×24.344.2%1.74011.3222 ⭐
ibm042858,65424.8×25.144.8%1.50371.3024 ⭐
ibm063189,74526.1×26.546.1%2.50571.6187 ⭐
ibm0733510,32826.8×27.246.8%2.02291.4633 ⭐
ibm0835210,90127.5×27.947.4%1.92391.4285 ⭐
ibm0936911,46328.1×28.548.0%1.38751.1194 ⭐
ibm1038712,01828.8×29.248.6%2.11081.5009 ⭐
ibm1140512,56829.4×29.849.2%1.71111.1774 ⭐
ibm1242313,11130.1×30.549.8%2.82611.7261 ⭐
ibm1344113,64730.7×31.150.4%1.91411.3355 ⭐
ibm1446014,17831.4×31.851.0%2.27501.5436 ⭐
ibm1547914,70432.0×32.451.6%2.30001.5159 ⭐
ibm1649815,22532.7×33.152.2%2.23371.4780 ⭐
ibm1751715,74133.3×33.752.8%3.67261.6446 ⭐
ibm1853716,25334.0×34.453.4%2.77551.7722 ⭐

Each benchmark includes:

Baseline Analysis:

💡 Why This Is Hard

Despite "only" 246-537 macros, this problem is extremely challenging:

  1. Massive search space: ~10^800 possible placements (even with constraints)
  2. Conflicting objectives: Wirelength wants clustering, density wants spreading, congestion wants routing space
  3. Non-convex landscape: Millions of local minima, discontinuities, plateaus
  4. Long-range dependencies: Moving one macro affects costs globally through thousands of nets
  5. Hard constraints: No overlaps between heterogeneous sizes (33× size variation)
  6. Tight packing: 43-53% area utilization leaves little slack
  7. Runtime matters: Must be fast enough to be practical (< 5 minutes ideal)

Classical methods (SA, RePlAce) have been refined for decades but still have room for improvement!

📖 Documentation

📚 References

🏅 Leaderboard

Submissions are ranked by average proxy cost across all 17 IBM benchmarks (lower is better). Zero overlaps required on all benchmarks. Scores are unverified until confirmed by judges.

RankTeamAvg Proxy CostBestWorstOverlapsRuntimeVerifiedNotes
1"Carrotato"0.9671——03.8min/benchNew 5/12.
2"Shoom"0.978——055min/benchResubmitted 5/12. Previous variant verified 1.4901.
3"vmallela"1.01090.76441.2921015.5h total:white_check_mark:Verified 1.0109 (self-reported 1.1).
4"DREAMPlaceProMaxUltra"1.01210.79551.216706h total:white_check_mark:Verified 1.0121 (self-reported 1.0467). Built and ran from team-provided Dockerfile.
5"QuantSC"1.0285——055min/benchNew 5/16. (Vincent Xianrui Zhang, USC)
6"QED"1.031——030min/benchNew 5/16. (Shubham Shukla)
7"Cezar"1.037——04h totalResubmitted 5/11. Previous variant verified 1.1893 (same self-report 1.037). Now ships team-provided Dockerfile.
8"Place, Route, Roll"1.0594——0~17min/benchNew 5/17. (Mike Gao & Amanda Yin, Cerebras) Team's prior submission was DQ'd for overlaps; resubmitted with 0 overlaps.
9"thinkorplace"1.0771——052min/benchNew 5/13. (Brendan Murrell & Xiaofei Zhang)
10"KLA MACH"1.0817——030min/benchResubmitted 5/16 (phase39, team "Chuanqi Chen"). Previous: phase38 (1.1335), phase36 verified 1.2121. Consolidates UTDA / Chuanqi Chen / KLA MACH submissions (one algo per team).
11"Archgen"1.09480.79311.4345015.25h total:white_check_mark:Verified 1.0948 (self-reported 1.16511 — 5.6% BETTER than self-reported).
12"Vibe"1.1443——013851s total:white_check_mark:Verified 1.1443 (self-reported 1.1477).
13"JonaU"1.1524——055min/benchNew 5/13. (Jona Uhe)
14"ArzunPD"1.1883——055min/benchResubmitted 5/8 (was 1.2478).
15"Talyxion"1.2075——07.2min/benchNew 5/10.
16"Hoop Dreams"1.22070.89721.507205h total:white_check_mark:Verified 1.2207 (self-reported 1.2206). Built and ran from team-provided Dockerfile.
17"MakerCode"1.2282——0~55min/benchResubmitted 5/16 (v2). Was 1.26. (Wei Yet Ng)
18"KKPlace"1.2451——055min/benchNew 5/13. (Kenneth Hou)
19"Two-IIITK-Kids"1.2648——0~34min/benchResubmitted 5/15. Was 1.2868. (Amruth Ayaan Gulawani & Joel Dan Philip, IIIT Kottayam)
20"Adam_A"1.2655——0682s/benchNew 5/10.
21"The Basin Jumpers"1.2748——0900s/benchResubmitted 5/13 (William Zhang & Leon Do). Was "William Zhang" / "Convex Optimization", verified 1.4556.
22"RoRa"1.27880.95771.622202.6h total:white_check_mark:Verified 1.2788 (self-reported 1.2723). Resubmitted 5/1.
23"MTK"1.28180.90731.6529037s/bench (GPU):white_check_mark:Verified 1.2818 (self-reported 1.317).
24"Electric Beatle"1.3253——02000s/bench (GPU)Resubmitted 4/30 (was verified 1.3913).
25"UToronto Analytical"1.33230.93711.6545024min total:white_check_mark:Verified 1.3323 (self-reported 1.3325).
26"V5"1.3382——0850s/benchNew 4/23.
27"jrslbenn"1.353——0750s/benchNew 5/4.
28"Barsat Khadka"1.38——01000-1800s/benchNew 5/5.
29"Varun's Parallel Worlds"1.40171.03621.7298027s/bench:white_check_mark:
30"UT Austin - AS"1.4076——017s/bench
31"ByteDancer"1.41511.02361.7792038min/bench:white_check_mark:
32"TAISPlAce"1.4321——028min/bench
33"Pragnay"1.4427——0632s/benchBlocked on compute_proxy_cost(..., plc=None) in fallback path.
34"No Man's Sky"1.4445——08.8min/benchNew 5/4, resubmitted 5/6.
35"Aegir"1.4553——0104s/benchNew 5/9.
36"another Waterloo kid"1.4568——0118s/benchBlocked on Modal cloud dispatch — can't run air-gapped.
—RePlAce (baseline)1.45780.99761.83700—:white_check_mark:
37"W3 Solutions"1.4824——090s/benchRuntime exceeds 1h/bench cap.
38"Captain.Rhinoceros"1.4906——0<26.6min/benchNew 5/13. (John Saleeby)
39"Jiangban Ya"1.49431.08911.8099049s/bench:white_check_mark:
40"UTAUSTIN-CT"1.50621.13631.794106s/bench:white_check_mark:
41"oracleX"1.51301.13401.7937011s/bench:white_check_mark:
42"SEVmakers"1.5200——0200s/benchPrivate repo — pending judge access.
43"CA"1.52471.22261.794502s/bench:white_check_mark:Verified 1.5247 (self-reported 1.5238).
44"ZeroLatency"1.5286——017s totalNew 5/9.
45"#5 ubc cpen student"1.53371.14111.8084013s/bench:white_check_mark:
46Will Seed (Partcl)1.53381.16251.7965035s total:white_check_mark:
47"6ummy"1.5361——0221s/benchNew 5/17. (Jungwoo Shin)
48"RUDY Can't Fail"1.53971.19271.888106min total:white_check_mark:Verified 1.5397 (self-reported 1.3605).
49"UT Austin - RH"1.6037——04.5s/bench
50"Mr_Chonk"1.7031——0~23min/benchNew 5/16. (Kush Voorakkara)
51"Binghamton"1.7621——02min/benchNew 5/10. Team also reports 1.5375 from an alternate non-legalized run.
52"UT Austin - CT"1.8706——0187s/bench
53"rpocevi"1.8894——022.5s/benchNew 5/9.
54"AS"1.91211.46142.350800.16s total:white_check_mark:
55"Adi's Team"2.0025——03726s/benchBlocked on compute_proxy_cost(skip_congestion=True) kwarg.
56"Sharc #1"2.04331.51432.4336096s/bench:white_check_mark:
—SA (baseline)2.12511.31663.67260—:white_check_mark:
57"Satisficing"2.2109——0<1s/benchNew 5/17. (Noah Cencini)
—Greedy Row (demo)2.21091.67282.769600.3s total:white_check_mark:
—"MacroBio"pending————
DQ"Mike Gao"self-reported 1.3255——193916min/bench1939 overlaps (old submission). Resubmitted 5/17 as "Place, Route, Roll" (rank 8).
DQ"BakaBobo"self-reported 1.4044———282s/benchMissing import — code won't run.

Submit your results via the Submission Link!

🤔 FAQ

Q: What benchmarks are used? A: Tier 1 (proxy cost) uses 17 IBM ICCAD04 benchmarks — the standard academic suite with well-established baselines. Tier 2 (OpenROAD flow) uses NG45 commercial designs (ariane133, ariane136, mempool_tile, nvdla) plus 1-2 hidden designs. You can evaluate on both with --all (IBM) and --ng45 (NG45).

Q: What if I beat one baseline but not the other? A: You must beat BOTH SA and RePlAce baselines on WNS, TNS, and Area to qualify for the Grand Prize. You can still win the Proxy or Innovation prizes regardless.

Q: Are there hidden test cases? A: All 17 IBM benchmarks for proxy cost ranking are public. The 4 NG45 designs are also public. For the OpenROAD flow evaluation (Tier 2), we will additionally test on 1-2 hidden NG45 designs to ensure generalization.

Q: What counts as "beating" the baseline? A: For proxy cost (Tier 1), your aggregate score across all IBM benchmarks must be lower than the baselines. For the Grand Prize (Tier 2), your OpenROAD results for WNS, TNS, and Area must surpass both SA and RePlAce baselines on NG45 designs.

📧 Contact

📄 License

This project is licensed under the Apache License 2.0 - see LICENSE.md for details.

Competition Updates

The organizers may update or clarify rules, evaluation details, timelines, prizes, or infrastructure as needed to ensure fairness, technical accuracy, and smooth operation of the competition. Any updates will be communicated through official channels and will apply going forward.

Participation in the competition constitutes acceptance of the current rules and any subsequent updates. The organizers’ decisions regarding scoring, eligibility, and interpretation of these rules are final.

Submissions & contact information may be shared with sponsors.