# Benchmark Data

These are the raw `llm-inference-bench` JSON artifacts cited by the blog posts.
Checksums are SHA-256.

| File | SHA-256 |
| --- | --- |
| `ai01-dsv4f-v9-tp3-b12x-a16-450w-confirm-20260710-212947.json` | `5caddfbc5d0bd67115c30ca69a2d1b7b51059f5cdc13630a64f4e9badb902ff6` |
| `ai01-dsv4f-v9-tp3-b12x-oneshot-min16k-c1-30s.json` | `7d54bf4690fa2c3d6091692830b7823fb44a1fc7087d30f1cc5865f33245df68` |
| `ai01-dsv4f-v9-tp3-b12x-oneshot-min32k-c1-30s.json` | `dc032f6d1367d83cab994aa032c9dff495c23824347c2d9497adc478b5299656` |
| `ai01-dsv4f-v9-tp3-b12x-oneshot-min64k-c1-30s.json` | `abe779832850d0e171954c1f52aa291d642ab6b204250954da77c63d34e116d7` |
| `ai01-dsv4f-v9-tp3-b12x-oneshot-min64k-confirm.json` | `b6fd1a8209eb100c3cda90da85ac944cee31d5ec0979ffd68debcc764fdadffb` |
| `ai01-dsv4f-v9-tp3-b12x-mtp2-matched-450w-confirm.json` | `f88464b6d95c55d63a9d21269c9d4be49363781f8f96ae195b1dc069160c03cb` |
| `ai01-dsv4f-v9-tp3-dspark-n5-ropefix-450w-confirm.json` | `503378b85454a0b1d8f3e2c8facd4455f8339643bf717b751c2737a99be9ad1b` |
| `ai01-dsv4f-v9-b12x-a16-450w-20260710-183109.json` | `105e2f5a150193301952e25babd5c47da9573119f050114bd824121744b3da56` |
| `ai01-dsv4f-v9-lucifer-cutlass-450w-20260710-192953.json` | `c92926b35f5214a75691c68e924f1f91d329a820a2db13671731a8fdf31eed50` |
| `ai01-ds4-v10-dspark-1m-graph192-20260713-132554.json` | `8737058dd6290d50c18958ff409ae7b325b7527d17eece4ab8284b12f1ed0360` |
| `ai01-ds4-v10-mtp2-1m-prefill-gpu912-20260712-224154.json` | `39e17dd9bcb5b54a5c394415121d54aa4232571aae0fd6e269939292f0cce8a5` |
| `ai01-ds4-v10-dspark-1m-gpu95-graph192-20260713-125526.json` | `ad8481e3112b98d8b0004efaa752a5b89970fd751366cb5fbcdca467cdbf4bc9` |

The following two artifacts are retained as rejected measurements. Their
reported near-1M prefill rates failed token-count, TTFT, and server-validation
sanity checks; part five explains why they are not benchmark results.

| Rejected file | SHA-256 |
| --- | --- |
| `ai01-ds4-v10-mtp2-1m-prefill-20260712-223749.json` | `b788bab94519d20f804ae33bd87c0c7b46240bb3a99f8fafda90ed3f3f90e9b8` |
| `ai01-ds4-v10-dspark-1m-gpu95-graphauto-20260713-131957.json` | `0462badac486baf09d97e0a0ef5e400058b9dc1f2f675954a51a8c529800c413` |

The benchmark version is recorded inside each artifact. TP3 matched results
use `0.4.29`; later TP2 v10 results use `0.4.30`. Results from different
versions are not treated as matched A/B measurements unless their methodology
and launch configuration are independently shown to agree.