Choosing inference hardware, with the disagreements left in.

Every claim on this site is one JSON record in data/, and every record cites the sources it was checked against. Where two sources disagree, the record is marked contested and both numbers are kept — this repo does not pick a favourite. Adraft record has not been verified yet. Neither is a verified record, and on this site they never look like one.

Start from a search, a side-by-side comparison, or the decision docs below.

verifiedmedium · 0.778,892 verified · 143 contested · 445 draft · 25 deprecated · mean confidence 0.77

236Accelerators 165FLOP classes 119Inference engines 109Quantization formats 75Interconnect 4,667Benchmarks 475Gotchas 219Where to buy / source hardware 15Hosted inference APIs 191Models 294Papers 23Compilers & kernel DSLs 316Metric exposures 39KV-cache efficiency 2,562Sources (provenance)

Answer a question

These pages join records across types to answer questions no single record can. Every number traces to the records it was derived from.

Cost per Tokensupply × benchmark → USD/Mtok Model × Engine Compatibility14 families × 8 engines Roofline Exploreraccelerator specs × flop classes

Decision docs

Written from the records, citing record ids inline. Every claim in these docs is backed by a record you can open.

Hardware selection by workload01-hardware-selection FLOP map: which operations dominate, and what that means for hardware02-flop-map Engine selection: which engine for which hardware03-engine-selection Quantization: weight-only-4bit vs W4A4 are different regimes04-quant-selection Known traps, grouped by severity05-known-traps

All records

9 cross-references in data/ name a record that does not exist yet; those render as plain text rather than as broken links. Sources (2562) are kept out of the main navigation on purpose: they are provenance, not findings. Compare any two or four records side by side on thecompare page.

Using the repo directly

This site is generated. The repo is the source of truth, and the same answers are available without a browser:

python tools/query.py list accelerators
python tools/query.py get accelerators/nvidia-h100-sxm
python tools/query.py refs gotchas/aiter-gate-skips-rdna3-gfx1100
python tools/query.py search "prefix caching"
python tools/cost_per_token.py

python tools/build_site_data.py    # regenerate site/src/data/
cd site && npm run build     # rebuild this site