Warp sharing: tested, with live schedules preserved
Expert reuse, lane-zero reductions and attention broadcasts across both V100s. Read the matched controls, regressions, exact-output checks and retained exclusions.
Read the qualification ↗Performance / Evidence / Progression
Find the workload.
Read the result.
An inference number only makes sense beside its model, hardware and execution settings. Start with a matched study, or explore every measured configuration below.
98retained serving runs · November 2–3, 2026
16 + 2serving configurations + compatibility checks
2 GPUsV100 and Titan Xp · distinct environments
VNNI tiles: correctness verified, serving gain unresolved
Eight repetitions per model and input length before and after activation. Read the full result, including variation, unchanged outputs and excluded startup-policy attempts.
Read the live evaluation ↗Where does expXa stand against PXA4?
The latest matched test put long-prompt request medians at 24.27 seconds for expXa and 24.38 seconds for PXA4. Their observed ranges overlap. Open the study for short-prompt results, individual runs and the settings that make this comparison meaningful.
Read the engine comparison ↗Configuration explorer
This is a record browser, not an overall leaderboard. Historical phases, hardware and offload configurations differ. Compare engines within a matched study before interpreting differences as an optimization gain.
All 32 measured configuration/length combinations shown.
| Configuration / study phase | Input | Prefill tok/s | Decode tok/s | Request s | Range s | Runs |
|---|---|---|---|---|---|---|
| V100 · expXa0982 · parallel preparation | 512 | 203.97 | 37.28 | 6.00 | 5.77–6.07 | 3 |
| V100 · expXa0982 · parallel preparation | 4,096 | 201.53 | 32.96 | 24.27 | 24.16–27.10 | 3 |
| V100 · expXa0982 · old-preparation ablation | 512 | 190.34 | 38.16 | 6.00 | 5.99–6.05 | 3 |
| V100 · expXa0982 · old-preparation ablation | 4,096 | 189.38 | 30.93 | 25.83 | 25.53–25.89 | 3 |
| V100 · PXA4 · tuning control | 512 | 180.82 | 31.23 | 6.95 | 5.91–7.45 | 3 |
| V100 · PXA4 · tuning control | 4,096 | 202.00 | 30.41 | 24.38 | 24.01–26.76 | 3 |
| V100 · expXa a078 · before/after reference | 512 | 187.43 | 37.75 | 6.13 | 6.00–6.25 | 2 |
| V100 · expXa a078 · before/after reference | 4,096 | 179.34 | 32.46 | 26.81 | 26.17–27.44 | 2 |
| V100 · expXa a078 · Rust CPU port | 512 | 191.99 | 35.60 | 6.34 | 6.01–6.90 | 4 |
| V100 · expXa a078 · Rust CPU port | 4,096 | 186.92 | 33.18 | 25.82 | 25.68–25.89 | 4 |
| V100 · PXA4 · Nov 3 matched control | 512 | 209.54 | 37.62 | 5.84 | 5.77–6.46 | 4 |
| V100 · PXA4 · Nov 3 matched control | 4,096 | 202.79 | 33.65 | 24.01 | 23.66–26.26 | 4 |
| V100 · current live, refreshed | 512 | 123.24 | 32.94 | 8.10 | 7.71–8.52 | 3 |
| V100 · current live, refreshed | 4,096 | 125.61 | 31.11 | 36.56 | 36.51–39.68 | 3 |
| V100 · current engine clone | 512 | 108.57 | 33.13 | 8.95 | 8.53–13.57 | 3 |
| V100 · current engine clone | 4,096 | 128.13 | 30.10 | 36.77 | 34.53–40.92 | 3 |
| V100 · PXA4, same settings | 512 | 206.85 | 37.42 | 5.85 | 5.76–7.83 | 3 |
| V100 · PXA4, same settings | 4,096 | 199.57 | 29.30 | 25.16 | 23.59–27.93 | 3 |
| V100 · expXa, same settings | 512 | 79.85 | 31.94 | 10.42 | 9.05–12.16 | 3 |
| V100 · expXa, same settings | 4,096 | 90.73 | 29.21 | 49.42 | 41.64–49.83 | 3 |
| V100 · original live measurement | 512 | 135.89 | 36.06 | 7.34 | 7.19–7.38 | 3 |
| V100 · original live measurement | 4,096 | 135.54 | 32.66 | 34.10 | 33.96–34.26 | 3 |
| Current settings · Titan | 512 | 11.20 | 9.43 | 59.61 | 58.74–60.06 | 3 |
| Current settings · Titan | 4,096 | 11.29 | 8.46 | 378.47 | 376.86–380.06 | 3 |
| expXa · max tested offload · 19 CPU banks | 512 | 51.58 | 23.98 | 15.30 | 15.20–15.35 | 3 |
| expXa · max tested offload · 19 CPU banks | 4,096 | 51.23 | 22.25 | 85.65 | 85.61–86.25 | 3 |
| PXA4 · max tested offload · 19 CPU banks | 512 | 71.00 | 25.45 | 12.24 | 12.23–12.40 | 3 |
| PXA4 · max tested offload · 19 CPU banks | 4,096 | 70.75 | 22.75 | 63.51 | 63.46–63.53 | 3 |
| expXa · initial placement · 20 CPU banks | 512 | 49.09 | 24.25 | 15.71 | 15.52–15.89 | 3 |
| expXa · initial placement · 20 CPU banks | 4,096 | 49.26 | 22.41 | 88.87 | 88.83–89.24 | 3 |
| PXA4 · initial placement · 20 CPU banks | 512 | 67.92 | 23.31 | 13.03 | 12.89–13.43 | 3 |
| PXA4 · initial placement · 20 CPU banks | 4,096 | 67.56 | 21.59 | 66.46 | 66.31–80.33 | 3 |
The experiments that did not produce a score
PXA Overdrive
The requested switch did not activate an Overdrive recipe for the original Rust weights. Its probe is retained as an activation check, not a speed result.
Strata
The native loader required PLE tensors absent from this model. No replacement weights were substituted to manufacture a comparison.
Kairo’s Q6_K work is a separate track
Kairo runs a Qwen 35B-A3B Q6_K profile with tuned attention and native multi-token prediction. The retained checks cover text output, vision, caching, cancellation and long-context retrieval. They do not measure the same CPU-heavy workload as Rust, and the available before/after records are not a matched PXA4 ranking.
Read the scope of the Kairo checks →Keep the evidence within reach
Download the complete machine-readable record or the original evidence bundle. Follow dated publications to see what changed without losing the earlier results.