Skip to main content
Attar: GPU-accelerated VLSI physical design in PyTorch

Diverse P&R benchmarks, distilled into one plug-and-play PyTorch environment on the GPU.
Evolve VLSI placement, timing-optimization and routing algorithms in minutes, not days.

Python PyTorch PyPI Downloads Validated Tests License

Quick start  ·  Why Attar  ·  What's inside  ·  How it works  ·  Gallery

Quick start

pip install attar
import attar
d = attar.load('after_fp/orfs_asap7/aes', device='cuda')   # from the zoo, as tensors on the GPU
d.place()                                                   # or set d.pos from your own algorithm
print(attar.Evaluator(d).hpwl(), d.timing().wns)            # HPWL, and WNS as OpenSTA reports it
d.export('aes.def')                                         # back to OpenROAD, Innovus, ICC2, ...

Designs download from the Attar zoo on first use. The first timing run on a GPU compiles its kernels (about two minutes, once). No GPU? Everything also runs on the CPU, or try the demo in Colab: Open In Colab

Timing-driven placement written in fifteen lines of Python, run on cva6 and swerv: place for wirelength, time it, weight the nets by their slack as OpenROAD's global placement does, place again, time it

Timing-driven placement in 15 lines (examples/timing_driven.py): cva6 WNS -195 → -75 ps, in 2.2 s on an RTX 3080 Ti.

Why Attar

  • Fast. Exact HPWL of 8.9M pins in 3 ms; a 2.2M-cell design placed and legalized in 49 s. An idea, from your code or an AI agent's, is tested in seconds, not in an overnight P&R run.
  • One environment. 262 designs from 24 benchmark suites (ISPD 2005-2024, ICCAD 2015 and 2025, MCNC, GSRC, IBM-HB+, OpenROAD-flow-scripts) load with one call, in one format, from LEF/DEF, Bookshelf or YAL. An algorithm is a function from tensors to positions: no database API, no TCL.
  • Faithful. Against OpenROAD and OpenSTA on the same placement: HPWL within 0.00002%, timing (WNS) within 0.34%, power within 0.06%, global-routing wirelength within 3.6%.

What's inside

Roadmap in three cards. Data, 262 designs in the zoo: contest benchmarks ISPD 2005 (8), ISPD 2015 (20), ISPD 2018 (10), ISPD 2019 (10), ISPD 2024 (5), ICCAD 2015 (8), ICCAD 2025 (13), GSRC placement (4) for placement; MCNC (10), GSRC blocks (12), IBM-HB+ (28) for floorplan; ASAP7 (15 placement, 5 floorplan), Nangate45 (8 and 8) and NS3K (98 placement). Engines, done: die sizing, packing, macro placement, IO placement, global placer, legalizer, PDN-aware refinement, detailed placement, clock tree synthesis, global router, gate sizing, buffering and cloning, debuffering; not yet: delay-balanced clock trees. Evaluation, done: HPWL, Steiner wirelength, legality, displacement, density overflow, RUDY, global routing overflow, setup and hold timing, GR parasitics, propagated clocks and CRPR, slew, capacitance and fanout limits, power; not yet: blocked pin checking. Left to your P&R tool: PDN synthesis, detailed routing, routing DRC, sign-off

Every engine and evaluator, with its accuracy against OpenROAD and OpenSTA: docs/features.md.

How it works

Sources (LEF/DEF, Bookshelf, YAL, GSRC blocks) are distilled into the Attar IR and loaded as a Design on the GPU; attar.naive moves it and adds facts, the Evaluator measures it; export writes a DEF for your own flow (OpenROAD, Innovus, ICC2, ...)

Attar splits physical design at the power grid, as production flows do:

  • before_fp: given a netlist, find the die, IO pins and macros (d.floorplan()).
  • after_fp: given a fixed floorplan with its PDN, place the cells, then size, buffer, route and time them (d.place()).

Every design states the facts it has and where each came from, and every engine says which facts it needs:

>>> d.status()
  ✅ netlist           def             12268 nets, 42885 pins
  ✅ cell_geometry     lef             12420 cells (12009 movable, 0 macros), 56 masters
  ✅ pdn               def             1815 wire and 4785 via shapes on M1, M2, M3
  ...
  ❌ placement         12009 movable cells unplaced
  -> only a placement is missing: place the design, then d.export("out.def")

Docs: concepts · engines and evaluators · benchmarks · API · validation

Floorplan mode: GSRC n100 at 85% block area, floorplanned twice: GP plus LP legalization leaves 5 macros outside the core (not legal), B*-tree packing fits them all (legal)
Floorplanning GSRC n100: B*-tree packing fits all 100 blocks into the core in 5.1 s.
cva6 after placement, colored by global-routing load on OpenROAD's gcells ariane136 after global placement: cells with a signal pin on the PDN, in stripes along the power grid
Routing congestion cva6, placed in 2.4 s. Pin access ariane136: cells with a signal pin on the PDN's vias.
ariane136 legalized: legal, rails and PDN included GSRC n300 floorplanned by B*-tree packing, with the strongest connections between blocks and IO pins
Legal placement ariane136, rails and PDN included, in 2.4 s. Floorplan GSRC n300: 300 blocks, 569 IO pins, packed in 43 s.

Acknowledgements

Attar's placer follows ePlace, DREAMPlace and Xplace; its checks follow OpenROAD. It is validated against OpenROAD-flow-scripts on ASAP7 and Nangate45. Benchmarks are from the ISPD, ICCAD and GSRC contests, MCNC and ORFS.

License

MIT License

Copyright © 2026 RayChang987  ·  free to use, modify and distribute; see LICENSE

Attar

Metadata

Release files for attar 1.0.4

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for attar 1.0.4
File Size Uploaded Provenance Actions
attar-1.0.4.tar.gz 365.1 kB Without provenance Details for attar-1.0.4.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for attar 1.0.4
File Interpreter ABI Platform Actions
attar-1.0.4-py3-none-any.whl Python 3 none any Details for attar-1.0.4-py3-none-any.whl

Total release size: 721.8 kB

Release files / attar-1.0.4.tar.gz

Download URL attar-1.0.4.tar.gz
Size 365.1 kB
Tags Source
SHA-256 checksum
How to use checksums
74990c811dd432e642b90ea86491192feef99823eb1e1f9e393a0a4c881b3eb1
BLAKE2b-256 checksum
How to use checksums
20f107dd6c2c9eb460d78ac356da07f469e109074d08ce05138fbdd40aaa175e
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/6.1.0 CPython/3.8.10

Release files / attar-1.0.4-py3-none-any.whl

Download URL attar-1.0.4-py3-none-any.whl
Size 356.7 kB
Tags Python 3
SHA-256 checksum
How to use checksums
8c7161bf4b7c11a4934a5cf06b1fb30bc1dae7c0420a0827109e2ed5ecc5f202
BLAKE2b-256 checksum
How to use checksums
f0e5ec397c24ca62cc7ab1075cf7f96a1f679a007bc8ffa4feefa7ea2ee22cd4
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/6.1.0 CPython/3.8.10

Security for attar 1.0.4

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page