tapeout-labs/aplace-transitions

aPlace Transitions Dataset

Placement optimization transitions from the aPlace chip placement system. Each record captures a single placement step (gravity solve, swap, or legalization) with before/after state, enabling offline reinforcement learning and imitation learning for EDA.

Dataset Summary

Split	Circuits	Records	Source Benchmarks
balanced_tilos	15	356,767	TILOS MacroPlacement (ariane, mempool, nvdla) + IBM-ISPD04 (ibm01–ibm12)
balanced_ispd_iccad	13	428,007	ISPD 2005 (adaptec1–4, bigblue1) + ICCAD 2015 (superblue1–18)
Total	28	784,774

All transitions were generated using RUDY-guided simulated annealing with balanced acceptance criteria.

Circuit Metadata

Circuit	Macros	Source
ariane133	133	TILOS
ariane136	136	TILOS
ibm01	246	IBM-ISPD04
ibm02	271	IBM-ISPD04
ibm03	290	IBM-ISPD04
ibm04	295	IBM-ISPD04
ibm06	178	IBM-ISPD04
ibm07	291	IBM-ISPD04
ibm08	301	IBM-ISPD04
ibm09	253	IBM-ISPD04
ibm10	786	IBM-ISPD04
ibm11	373	IBM-ISPD04
ibm12	651	IBM-ISPD04
mempool_tile	20	TILOS
nvdla	128	TILOS
iccad15_superblue1	175	ICCAD15
iccad15_superblue3	195	ICCAD15
iccad15_superblue4	177	ICCAD15
iccad15_superblue5	294	ICCAD15
iccad15_superblue7	253	ICCAD15
iccad15_superblue10	84	ICCAD15
iccad15_superblue16	101	ICCAD15
iccad15_superblue18	84	ICCAD15
ispd05_adaptec1	543	ISPD05
ispd05_adaptec2	566	ISPD05
ispd05_adaptec3	723	ISPD05
ispd05_adaptec4	1329	ISPD05
ispd05_bigblue1	560	ISPD05

Features

Field	Type	Description
circuit_id	string	Circuit identifier
step_type	string	One of: gravity_solve, swap, legalization
step_number	int	Sequential step index within the circuit's optimization trajectory
action	string (JSON)	Action parameters (e.g. {"type": "swap", "cell_i": 5, "cell_j": 12})
hpwl_before	float	Half-perimeter wirelength before the step
hpwl_after	float	Half-perimeter wirelength after the step
delta_hpwl	float	Change in HPWL (after − before). Present for swap steps
accepted	bool	Whether this move was accepted (SA acceptance criterion). Present for swap steps
overlap_before	float	Cell overlap metric before the step
overlap_after	float	Cell overlap metric after the step
is_legal_before	bool	Whether placement was legal before the step
is_legal_after	bool	Whether placement was legal after the step
rudy_before	float	RUDY congestion estimate before the step. Present for swap steps
rudy_after	float	RUDY congestion estimate after the step. Present for swap steps
delta_rudy	float	Change in RUDY (after − before). Present for swap steps
positions_before	string (JSON)	Array of [x, y] macro positions before the step
positions_after	string (JSON)	Array of [x, y] macro positions after the step
config	string (JSON)	Optimization hyperparameters for this run
timestamp	string	ISO timestamp of when the transition was generated

Usage

from datasets import load_dataset

# Load a specific config
ds = load_dataset("tapeout-labs/aplace-transitions", "balanced_tilos")

# Load everything
ds_full = load_dataset("tapeout-labs/aplace-transitions", "full")

# Filter to accepted swaps only
swaps = ds["train"].filter(lambda x: x["step_type"] == "swap" and x["accepted"] == True)

# Parse positions for a record
import json
record = ds["train"][0]
positions = json.loads(record["positions_before"])  # List of [x, y] pairs

Generation Process

Transitions were generated by running aPlace's RUDY-guided simulated annealing optimizer on each circuit:

1. Gravity solve: Initial quadratic placement using a gravity-based objective
2. Swap: Pairwise cell swaps evaluated by a combined HPWL + RUDY objective, accepted/rejected via Metropolis criterion
3. Legalization: Greedy legalization to remove overlaps

The "balanced" prefix indicates that the acceptance temperature was calibrated to produce a ~40-60% acceptance rate, yielding a mix of improving and exploratory moves suitable for offline RL.

Data Sparsity

To reduce file size, full position arrays (positions_before, positions_after) and hpwl_before/overlap_before are stored at ~10% sampling rate (every 10th swap step). The remaining 90% of swap rows have these fields as NULL.

Fields available on every row: circuit_id, step_type, action, delta_hpwl, hpwl_after, accepted, rudy_before, rudy_after, delta_rudy.

For training, the PlacementTransitionDataset in the aPlace codebase reconstructs positions for NULL rows by tracking running state sequentially.

Citation

@misc{aplace_transitions_2025,
  title={aPlace Transitions: Placement Optimization Trajectories for Offline Reinforcement Learning},
  author={Tapeout Labs},
  year={2025},
  url={https://huggingface.co/datasets/tapeout-labs/aplace-transitions}
}

License

CC-BY-4.0