#3232 [Benchmark] 3x RTX 2080 Ti 22 GB (sm_75): v0.1.45 vs v0.2.0-beta, UD-Q4_K_XL at 262K, layer split 16,32

Beschreibung

## Problem and result

`docs/TEST_REQUESTS.md` asks for a report from "Three or four NVIDIA GPUs", and there is no sm_75 multi-GPU report in `bench/results`. This adds one: three RTX 2080 Ti 22 GB in a layer split on a 125B MoE at its native 262,144-token window, measured on 0.1.45 (2026-10-18) and re-measured on **0.2.0-beta** (2026-10-19) on the same rig and config.

Result on 0.2.0-beta: 65.2 tok/s short-chat decode, 586.8 / 1,345.2 / 1,758.7 tok/s prompt reading at 4K / 16K / 32K, 69.7 tok/s agent decode at a 69.3% conversation-cache hit. The 0.1.45 → 0.2.0-beta change is inside this box's run-to-run spread (+1.7%, +0.9%, +0.2%, −0.1%, −2.8%).

## Related discussion

`docs/TEST_REQUESTS.md`, row "Three or four NVIDIA GPUs". No issue tracks this report.

This supersedes #2976, which was closed in the 0.2.0-beta tracker reset. GitHub would not reopen it after the rebase (the branch had been force-pushed), so the same branch is opened again with the 0.2.0-beta re-measurement added.

## Approach

Two report folders, same hardware, same engine arguments, so they compare directly:

- `bench/results/2026-10-18-community-3x-2080ti/` — Strata 0.1.45
- `bench/results/2026-10-19-community-3x-2080ti-0.2.0-beta/` — Strata 0.2.0-beta (engine 0.1.45.3, `3d1fb480`), with the 0.1.45 comparison table in its README

`results.json` records a layer-split run as one card and leaves `quant` / `config` empty (#2887); those fields were filled in by hand from the server's start line and `nvidia-smi`, and the README says so. The measured numbers are untouched.

## Validation

**Passed:** `python tools/strata_bench.py --url http://127.0.0.1:8097 --long` against a server already running, three runs per prompt size, 263 s for the suite, `results.json` → `failures` empty. Build: source at the `v0.2.0-beta` tag (`3d1fb480`) with `-DSTRATA_ENABLE_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=75`; `cuobjdump --list-elf` lists sm_75 cubins only.

**Failed:** none observed.

**Not run:** correctness, needle or token-identity checks (this is a speed suite); an IQ-family pack on three cards (Q4_K_XL uses the K-quant expert rows, so the engine's AVX-2 path did the expert work); full-power cards (the power limits are 210 / 167 / 200 W).

## Limits

Speed only. The absolute numbers are below what the same rig does at the default 250 W per card. One model, one quantization, one context length (262,144 tokens). The agent rows are 15 requests, so their medians are the weakest in the report.