Skip to content

Improve MJ97 placer#86

Open
Mummanajagadeesh wants to merge 27 commits into
partcleda:mainfrom
Mummanajagadeesh:main
Open

Improve MJ97 placer#86
Mummanajagadeesh wants to merge 27 commits into
partcleda:mainfrom
Mummanajagadeesh:main

Conversation

@Mummanajagadeesh
Copy link
Copy Markdown

@Mummanajagadeesh Mummanajagadeesh commented May 6, 2026

MJ97 Placer Update — ProxOpt v2

Proxy: 1.3241 | Overlaps: 0 | Team: MJ97

Multi-phase pipeline per benchmark:

  • Legalize from initial placement
  • WL thermostat (SA on sparse pair graph, overlap-safe moves)
  • Coarse-to-fine coordinate descent with exact proxy eval
  • LNS: destroy/repair patches of connected macros, accept if proxy improves
  • Final soft-macro FD pass on best hard placement
  • Multi-start (3x) with gaussian perturbation, keep global best

Mummanajagadeesh and others added 27 commits April 22, 2026 18:31
…leda#7

Verified avg proxy 1.0948 across all 17 IBM benchmarks, 0 overlaps,
~15.25h total runtime (3229s/bench avg). Self-reported 1.16511 — 5.6%
BETTER than self-reported.

Best ibm01=0.7931, worst ibm17=1.4345. AutoDMP++ analytical placement
+ multi-stage soft-CD / exact local search (submission_4.py).

Ran inside `eval_docker/Dockerfile` (pytorch 2.5.1 / cuda 12.4 / py3.11) —
local uv (pytorch 2.10 / py3.12) hit CUBLAS_STATUS_INVALID_VALUE in the
team's bundled placer.

Moves Archgen from rank 9 (self-reported 1.16511) to rank 7 (verified),
displacing Vibe → 8 and JonaU → 9.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add 6 new entries: QuantSC (1.0285), QED (1.031), Place Route Roll
(1.0594), 6ummy (1.5361), Mr_Chonk (1.7031), Satisficing (2.2109).

Update 3 existing entries: KLA MACH → 1.0817 (phase39), MakerCode →
1.2282 (v2), Two-IIITK-Kids → 1.2648. Update Mike Gao DQ note to
reference resubmission as Place Route Roll.

Strip implementation details from all Notes cells.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The Tier 1 section said "18 IBM benchmarks" while every other reference
in the same README (and SETUP.md, the evaluator, the FAQ) correctly
says 17. The ICCAD04 suite skips ibm05; the actual set is
ibm01–ibm04 and ibm06–ibm18.

Reported by Shreyash.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants