Skip to content

BENCHMARKS.md

Latest commit

 

History

History
37 lines (26 loc) · 1.04 KB

BENCHMARKS.md

File metadata and controls

37 lines (26 loc) · 1.04 KB

Benchmarks

SparseFlow is presented publicly as a benchmark-first inference product.

Public Benchmark Summary

  • 1.4x average speedup on validated production benchmark shapes
  • 1.6x-1.7x peak gains on FFN-heavy inference paths
  • 30-40% potential inference cost reduction on a good workload fit
  • Zero model changes required to start evaluating

Validation Status

  • Validated platforms: A100, RTX 3090
  • In active validation: RTX 4090
  • Not yet publicly validated in this repository: H100, additional RTX 30/40 variants

How To Read These Numbers

These are public benchmark summaries intended for product evaluation, not a full public dump of benchmark internals.

Publicly visible material should stay at the level of:

  • Supported workload categories
  • Validated hardware status
  • Reported speedup ranges
  • Evaluation scope and current limits

Keep the following private:

  • Kernel-level traces
  • Internal tuning thresholds
  • Deployment-specific operating points
  • Proprietary validation fixtures
  • Customer-specific workload details