Skip to content

Releases

Inference Performance Index v1

Choose a tag to compare

@jasonlizhengjian jasonlizhengjian released this 16 Oct 18:28

Inference Performance Index v1

Naming update: this project is now the Inference Performance Index (IPI). The v1.0.0 tag and existing download bytes are unchanged, so those historical bundles still use the previous ICI name. Use the default branch for the renamed documentation, figures and package metadata. New bundles built there use the inference-performance-index name.

Refreshed 29 October 2026. The v1.0.0 tag and downloadable bundles now include the accumulated report, figures, interactive explainers, and supporting-analysis updates. The original release commit is preserved at v1.0.0-original-2026-10-16.

The report uses the 15 October frozen base plus the 20 October Rubin revision: 19.2 TB/s HBM bandwidth and 1.5 TB/s outbound NVLink. VR200 scores 2.177 GB300 equivalents. The four-model basket, log 50–400 TPS band, workload, and pinned AISimulate 0.13.0 wheel are unchanged.

  • Reintegrated all 28 methodology-sensitivity scenarios across 23 accelerator profiles; all 616 non-Rubin scores are unchanged.
  • Refreshed model-side validation using the same 19 October InferenceX observations. Five Rubin recipes were retimed; the other 40 recipes are unchanged.
  • Preserved the Kimi/GB300 five-point comparison at 11.56% average absolute throughput difference under the study assumptions.
  • Archived the previous supporting analyses with hashes and added a verification record.

Reproduce

python3 reproduce.py rubin-update
python3 reproduce.py sensitivity
python3 studies/verify_analysis_refresh.py

The no-argument rebuild still reproduces the original 15 October score snapshot. See docs/SNAPSHOTS.md for the distinction.

Downloads

  • inference-capacity-index-v1.0.0.tar.gz: committed source, frozen inputs, documentation, visualizations, and verified AISim wheel.
  • visualizations.zip: all documentation and visualizations; open docs/visualizations/REPORT.html. Keep the folders together.
  • The separate AISimulate wheel is unchanged.

Current index: complete CPU reproduction

Choose a tag to compare

Four models, 23 accelerators; log 50-400 SOL TPS/user with workload-based prefill provisioning. Run python3 reproduce.py for an offline saved-result rebuild. Native setup, replay and full finite search are documented in README.md. Verified a clean environment, all 9,200 saved timing cases and a 718,289-candidate fresh search; all 161 score fields reproduce to floating-point precision. The archive includes the pinned unmodified AISim wheel. Python dependencies require package-index access for native installation. No GPUs, model weights, private traces or API keys are needed.