#3062 bench: 2x Quadro GV100 (sm_70) community report, dual vs single GPU

说明

Community benchmark results only, no engine changes.

2x Quadro GV100 (power-capped at 150 W, NVLink unbridged), Xeon E5-1650 v4, 125.7 GiB RAM, Unsloth UD-Q4_K_XL, 131K context, kv int8, spec 4.

Three strata_bench.py sessions per configuration (dual and single GPU):
- short-chat decode: 67.1 tok/s median (2x) vs 42.3 (1x) — 1.59x
- 16K fresh prompt: 1709 vs 1354 tok/s; 4K: 1071 vs 798
- agent session decode: 68.6 vs 54.9 tok/s, ~69% cache hit both

Warm learned expert cache in every session. Per-request JSON and tool-generated session reports included; details and limitations in README.md.