#2144 bench: community report, 2x Tesla P40 (PCIe 3.0 x8), i5-6500, 39 GiB RAM, IQ3_S, 131K, 0.1.43

本文

Adds one community benchmark folder: `bench/results/2026-10-12-community-2x-tesla-p40-qwen3-8-flash-next-iq3-s/`. Results only; no engine, docs or setup changes.

## Machine and setup

- 2x Tesla P40 24 GB (sm_61), PCIe Gen 3 x8 each, 250 W power limit (default), layer split across both cards (auto)
- Intel Core i5-6500 (4 cores, 4 threads), 39.0 GiB RAM, USB SSD
- Debian 13, Linux 6.12.111, driver 580.178.04, CUDA 12.9 engine
- Strata 0.1.43, commit 7695859a; Qwen3.8-Flash-Next IQ3_S, 131,072-token context, `--kv int8`, `--spec 4 --mtp rt`, vision on
- The install's own configuration, unchanged. Run with `python tools/strata_bench.py --long` (suite 1, 3 runs per test).

## Results (median, with range)

| Test | Result |
| --- | --- |
| short chat decode | 36.0 tok/s (28.8-44.7, n=9), first token 0.89 s |
| 4K fresh prompt (4,019 tokens) | 344.1 tok/s (326.4-345.5), first token 11.72 s |
| 16K fresh prompt (16,058 tokens) | 509.0 tok/s (506.8-509.5), first token 31.60 s |
| 32K fresh prompt (32,049 tokens) | 634.8 tok/s (633.4-634.9), first token 50.55 s |
| agent session, follow-up turns | cache hit 68.7 %, first token 9.68 s, decode 37.4 tok/s (n=15) |

For context only (not a same-day A/B): the 2026-10-10 2x P40 IQ3_S report (#1799, Xeon E5-2699 v3, 125 GB RAM, 0.1.41) measured 35.0 tok/s decode and 350 tok/s prompt at 3.9K. This run, on a 4-core desktop CPU with 39 GiB of RAM, comes out about the same at the short lengths.

## What was not tested

- No correctness check (no needle or recall test); speed only
- No 128K prompt, no concurrency test
- Background load was not measured (nothing significant was running)

🤖 Generated with [Claude Code](https://claude.com/claude-code)