inference/research

Working notes

Scratch space for research in progress. Nothing here is authoritative; once a fact is verified it moves into data/ as a record.

Conventions

  • Record ids are kebab-case slugs, globally unique.
  • Source ids are bare slugs; sources in any record must resolve in data/sources/.
  • Today's date is used for accessed and updated.
  • Vendor spec numbers are vendor_claim: true unless independently measured.
  • Dense vs 2:4-sparse must be explicit on every FLOPS figure.
  • Link bandwidth is recorded per-direction; if the vendor quoted bidirectional aggregate, the conversion goes in bandwidth_basis.

Open questions

  • Which parts of the consumer-GPU "AI TOPS" figures can be traced back to a documented dense fp8/fp16 number, and which are int4+sparse marketing only?
  • Is there a defensible roofline-balance TFLOPS figure for the FLOP classes, or does it have to stay qualitative per part?
  • How much of the ROCm-on-consumer-Radeon story has actually been fixed upstream?