#1811 Community benchmark: RX 9070 16 GB (gfx1201) on Windows 11, IQ3_XXS, engine 0.1.41

Descrição

## Summary

Results-only contribution: a community benchmark report for an **AMD Radeon RX 9070 16 GB (gfx1201) on Windows 11**, Flash-Next **IQ3_XXS**, prebuilt HIP engine **0.1.41**, 131,072-token context. No engine, setup or docs-of-record changes beyond the two index rows below.

## What changed

- `bench/results/2026-10-10-community-rx-9070/` — new folder:
  - `README.md` — hardware/software, model and configuration (with SHA-256 for both GGUF shards and the expert profile), method, results, failures and limitations, following `docs/COMMUNITY_BENCHMARKS.md`.
  - `bench-output.json`, `bench-output-128k.json` (+ the harness's `.md` summaries) — per-run data for all 15 measured rows, the 6 rejected rows and the 6 needle checks.
  - `benchmark.py` — the harness (standard library only; adapted from the 2026-10-03 R9700 report, with a context-fit guard added after the nominal 128K prompt was rejected).
  - `strata-iq3_xxs-session.log` — the whole measured engine session (startup lines, every prompt/decode/cache timing line).
  - `strata-iq3_xxs.json` — the launch configuration used (no credentials).
- `bench/results/COMMUNITY.md` — index row for the folder (PR cell `#1811`).
- `docs/COMMUNITY_BENCHMARKS.md` — bullet in the Community reports list.

## Extra Notes

**Measured:** 2026-10-10, one engine session, three runs each at 4,195 / 33,467 / 129,970–129,973 fresh prompt tokens (nonce per run, `0 reused` confirmed in the log), 256-token cap, greedy, `reasoning_effort: none`, warm expert cache, calibration and warm-up excluded.

| Label | Prompt tok/s median (range) | Decode tok/s median (range) | TTFT s median (range) |
| --- | --- | --- | --- |
| 4,096 | 586.2 (584.0–590.7) | 58.7 (57.9–60.5) | 7.208 (7.127–7.225) |
| 32,768 | 804.5 (803.4–804.8) | 57.9 (55.8–59.8) | 41.669 (41.640–41.716) |
| 126,640 (fit from 128,000) | 832.9 (831.8–844.1) | 57.9 (54.6–61.0) | 156.180 (154.094–156.386) |

Recall: 6/6 exact-string needles found (33,512 and 130,018 prompt tokens, depths 10/50/90 %), one run each.

**Failures kept as failures:** the nominal 128,000-token label produced 131,428 prompt tokens and was rejected with HTTP 400 (prompt + max_tokens > 131,072); those 6 rows sit in `bench-output.json` outside every median, and the leg was re-run at a fitted size.

**Limitations:** one machine, one engine session, desktop in normal use, no clock/power capture, no answer-quality check, opt-in `STRATA_HIP_WMMA` untested, 131,072-token context (below the model's native window), single quant.

**How it was checked:** every row's log line is paired to its request by prompt-token count (`pairing_ok` true on all 15 measured rows); client TTFT brackets the engine's read time on every row (e.g. 7,127 ms vs 7,102 ms), so the engine's timings are corroborated as lower bounds; prompt/decode figures are quoted from the engine log, TTFT/total from the client — no "tokens / total time" figures.

**Not changed:** README headline claims, engine sources, defaults, or any behavior — the default path stays byte-identical (results-only PR).