#2818 bench: community report — 2x RTX 3060 12 GB, Unsloth UD-IQ4_XS, 131K context, agent-calibrated (engine 0.1.44.1)

说明

Community speed report from a dedicated 2x RTX 3060 machine: the UD-IQ4_XS pack at the full 131,072-token window, the engine at a `--layer-split auto` split across both cards, with the expert pool tiered in 94 GiB of system RAM.

Adds `bench/results/2026-10-16-community-2x-rtx-3060-ud-iq4-xs-0.1.44.1/` (`README.md` + `results.json`), produced by `tools/strata_bench.py` suite 1 against the running server.

## Machine

- MEG Z490 ACE, i7-11700 (8C/16T), 94.2 GiB RAM; Linux, driver 595.99.02, CUDA 13.2
- 2x RTX 3060 12 GB, both CPU-direct at Gen4 x8 (the CPU's 20 lanes split x8/x8, the NVMe takes the last x4); power limit 170 W per card — decode plateaus near 90 W, only prefill bursts reach the limit
- Strata 0.1.44.1 (commit 296b59d2, release build); `layer split auto`: layers 0-21 on GPU 0, layers 22-47 plus a 2939-slot (6.62 GiB) expert cache on GPU 1
- Calibrated with `./setup.sh --calibrate`: `--pcie-frac 0.20 --spec-min-p 0.70 --adapt-every 1 --adapt-swaps 80 --adapt-decay 0.97`; `STRATA_STAGE_TRIM=1 STRATA_IO_PREFETCH=1`

## Headline medians

| Test | Result |
| --- | --- |
| short chat decode | 45.2 tok/s (TTFT 0.62 s) |
| agent follow-up turns decode | 52.1 tok/s |
| fresh prompt 4K / 16K | 799 / 1216 tok/s |
| agent turn TTFT | 5.22 s |
| conversation cache hit | 68.6 % |

## Notes

- Reproducibility: two independent suites on the same running server a few hours apart; every headline median agrees within 3 % (the README records the earlier run).
- Read alongside #2608 (same GPU pair, IQ3_S at 65K): the smaller quant decodes slightly faster on short chats there, while this report has 131K context and ~28 % faster 16K prefill — complementary points for the 3060 pair.
- Speed only: no recall/HumanEval run, no concurrency row; labelled as such in the README.

- [x] README read once before sharing; scrubber output checked (no host names, paths, or addresses)
- [x] report added under `bench/results/` in the `docs/COMMUNITY_BENCHMARKS.md` format
- [x] engine version, hardware, power limits and settings recorded in the README