ChiPFormer

Transferable Chip Placement via Offline Decision Transformer

Yao Lai, Jinxin Liu, Zhentao Tang, Bin Wang, Jianye Hao, Ping Luo

The University of Hong Kong / Shanghai AI Lab / Huawei Noah's Lab / Zhejiang University / Tianjin University

ICML 2023

Introdution

ChiPFormer casts the chip placement as an offline RL formulation and enables learning a transferable placement policy from fixed offline data, which can improve the placement efficiency and quality (up to 97% time and 65% HPWL decrease).

Video

Model Architecture

Placement Process Animation

Phase 1

Place macros by ChiPFormer.

Chip: Adaptec3

Chip: Adaptec4

Phase 2

Fix macros, place standard cells in coarse.

Phase 3

Set macros and standard cells movable,

place all modules.

Placement Result Comparison

Chip: Adaptec3

Human

HPWL = 19.41e7Overlap = 0%

DeepPR (NeurIPS 21')

HPWL = 24.11e7Overlap = 24.35%

PRNet (NeurIPS 22')

HPWL = 23.24e7Overlap = 10.71%

GraphPlace (Nature 21')

HPWL = 25.80e7Overlap = 1.24%

MaskPlace (NeurIPS 22')

HPWL = 21.47e7Overlap = 0%

DREAMPlace (DAC 19')

HPWL = 15.63e7Overlap = 0%

Flora (DAC 22')

HPWL = 15.65e7Overlap = 0%

ChiPFormer

HPWL = 13.97e7Overlap = 0%

ChiPFormer: Adaptec4

Human

HPWL = 17.44e7Overlap = 0%

DeepPR

HPWL = 23.40e7Overlap = 32.11%

PRNet

HPWL = 23.64e7Overlap = 13.36%

GraphPlace

HPWL = 25.58e7Overlap = 7.43%

MaskPlace

HPWL = 22.97e7Overlap = 0%

DREAMPlace

HPWL = 14.41e7Overlap = 0%

Flora

HPWL = 14.30e7Overlap = 0%

ChiPFormer

HPWL = 12.97e7Overlap = 0%

Offline Dataset Release

We release the offline training data in Google Drive. (500 placement results for 12 chip circuits.)

Documents can be seen in Google Docs.

Code Release

We have released our code on github.

Page updated

Google Sites

Report abuse