Macro Placement — AlphaChip, DREAMPlace, and OpenROAD Baseline

Scope

This document records the e1 macro-placement strategy across three placers, the post-route PPA validator that disciplines all three, and the "False Dawn" controversy that frames the entire effort. It is the human-readable companion to docs/evidence/pd/macro-placement-evidence.yaml and research/alpha_chip_macro_placement/07_post_route_ppa/.

Three placers under evaluation

Placer	Class	License	RL training cost	What it produces
OpenROAD (RTL-MP + macro placement step)	analytical + simulated annealing	open	none	baseline .plc per release run
AlphaChip (Google Circuit Training r0.0.4)	reinforcement-learning policy	open (Apache 2.0)	8x V100 or 1x H200 multi-day	candidate .plc from a trained policy
DREAMPlace 4.0	GPU-accelerated analytical	open	none	candidate .plc; macro_place_flag=1

All three placers emit a Circuit Training .plc (or DEF that we convert to .plc via scripts/alphachip/convert_lefdef_to_pb.sh). That gives us one common interchange format and one validator.

The validator: scripts/run_post_route_ppa.py

AlphaChip publishes a proxy cost (`wirelength_cost + 0.5 * congestion_cost

• 0.5 * density_cost`). The arXiv 2302.11014 "False Dawn" critique demonstrated that proxy improvements do not always translate to routed-PPA improvements after detailed route. Our discipline is:

1. Run every placer; capture proxy cost (compare_proxy_costs.sh).
2. Re-run OpenROAD detailed route on each candidate .plc (run_post_route_ppa.py).
3. Compare routed wirelength, DRC count, congestion histogram, hold/setup TNS, max-slew/cap violations, and post-route power.
4. Reject any candidate whose post-route PPA is worse than the OpenROAD baseline, regardless of proxy delta.

compare_proxy_costs.sh is now wired to call the validator after the proxy step so the two deltas land in the same output directory.

Current e1 numbers

Captured under research/alpha_chip_macro_placement/05_experiments/ and research/alpha_chip_macro_placement/06_e1_notes/:

16-macro smoke (toy benchmark)

Placer	Proxy cost
OpenROAD baseline	0.499
AlphaChip (toy training)	0.761

Interpretation: AlphaChip is worse here. The toy training run does not converge in a handful of iterations; this is consistent with the upstream Ariane recipe needing many-hour PPO runs.

256-macro full release (e1 softmacro benchmark)

Placer	Proxy cost
OpenROAD baseline	0.2379
AlphaChip (longer run)	0.2308
Delta	-3.01 % proxy

Interpretation: AlphaChip wins on proxy by 3 %. That is not PPA yet. The post-route PPA validator is BLOCKED until we have hard macros in the floorplan; see pd/macros/manifest.yaml. Without hard macros the AlphaChip proxy win has zero PPA leverage because there is nothing to place.

The "False Dawn" controversy — honest summary

The Mirhoseini et al. Nature 2021 paper claimed RL placement beat human designers on Google TPU blocks. Cheng et al. 2023 (arXiv 2302.11014) re-ran the experiment with reproducibility-strict protocol and found:

• The published proxy gains were not always reproducible.
• When reproduced, proxy gains did not consistently translate to routed wirelength or post-route timing improvements.
• The original paper compared to simulated annealing, not the state-of-the-art commercial placer (Innovus / Genus).

Google responded with the AlphaChip rebrand and updated benchmarks; the debate is not settled. Our position:

• Treat AlphaChip as one candidate placer, not a guaranteed win.
• Always validate routed PPA, never proxy alone.
• Compare against OpenROAD's macro placer and DREAMPlace's macro placement, not a hand-tuned annealer.
• Track the PPO training cost (8x V100 days or 1x H200 days) in the scorecard. If a placer requires GPU-days of training to match an analytical placer's output in seconds, the analytical placer wins on the cost dimension that matters for the 2028 schedule.

Blocked steps

• PPO RL training for an e1-specific policy is BLOCKED on H200 GPU access. The Nebius H200 runbook is concrete; the actual run has not executed.
• Post-route PPA truth is BLOCKED on hard macros existing (OpenRAM Sky130 macros not yet built). See pd/macros/manifest.yaml.
• Commercial-EDA placer comparison (Innovus + Genus) is BLOCKED on the commercial EDA gate (docs/evidence/pd/commercial-eda-gate.yaml).

What unblocks the macro-placement evidence gate

1. At least one OpenRAM Sky130 macro produces verified LVS-clean GDS and is placed in the OpenLane Sky130 release.
2. compare_proxy_costs.sh runs both proxy AND post-route PPA for the OpenROAD baseline and at least one alternative placer (AlphaChip or DREAMPlace).
3. The PPA JSON for both placers is archived under research/alpha_chip_macro_placement/07_post_route_ppa/.
4. The macro-placement-evidence.yaml gate transitions to complete_local_evidence.

Until then the gate fails closed.