#6468 [Benchmark] RX 9070 XT on Windows, Gyro-S at 131K: app 0.2.17 / engine 0.1.63 with a tuned config (standard 12/12)
open · @ric20527 · 0 Kommentare · Auf GitHub
BenchmarksAMD / HIPDocumentationWindowsLinux
Beschreibung
## Summary A results-only report from the same PC and model as #6161 (RX 9070 XT 16 GB at stock clocks, Ryzen 5 7600X3D, 31 GiB RAM, Windows 11, Gyro-S TQ1_0 at 131K), now on app 0.2.17 / engine 0.1.63 with the configuration this PC uses every day. One quick suite (`tools/strata_bench.py --long`) and one standard run (`strata bench run`, complete, 12 of 12 sanity checks), each from a freshly started engine. | | #6161 tuned (0.1.60) | this run (0.1.63) | | --- | ---: | ---: | | standard: short decode tok/s | 83.7 | 79.8 (77.6-95.7) | | standard: 4K / 16K prompt tok/s | 1276.7 / 1608.3 | 1830.2 / 2446.1 | | standard: decode on the 16K case tok/s | 81.2 | 91.5 | | quick: 32K fresh prompt tok/s | 1568.9 | 2451.7 | | quick: chat decode tok/s | 71.3 | 86.9 | | quick: agent decode tok/s | 60.4 | 57.1 (39.4-87.6) | Prompt reads gain 41-56 %. Short-answer decode is within its own spread. The #6161 columns are from another day; the table does not split the gain between the new settings and the engine releases. ## Also in the report Earlier the same night, with a GPU overclock still active, the card reset the display driver in a synthetic agent test, with `--adapt-async 1` and also without it. At stock, with the option on, the same test ran clean. So the overclock is the likely cause (corrected in #5672). The resets logged no event 4101, only dumps in `LiveKernelReports`. The option is not in the measured configuration. ## Files `docs/bench-results/2026-11-21-community-rx-9070-xt-gyro-s-0.2.17/`: the human README, `quick/` (README, results.json, engine log excerpt) and `standard/` (README, results.json, outputs.jsonl, config.json, provenance). One row added to `docs/bench-results/README.md`. ## Validation Revision and environment for both runs: desktop app 0.2.17, control 0.1.63 (commit `12d491da`), official engine package `0.1.63-windows-x86_64-hip-rx9070` (not a source build), Windows 11 Pro 10.0.26300, AMD driver 32.0.32029.2010, RX 9070 XT at stock clocks, configuration in `standard/config.json`, 2026-11-21. **Passed:** - `python tools/strata_bench.py --config <config.json> --long` with the Python of the app's control, started 01:43 local time: every test completed, three runs each, no request failed. Results in `quick/results.json`, engine log excerpt in `quick/engine-log-excerpt.txt`. - `strata bench run --profile <same config.json> --track custom` (protocol `strata-qwen38-thinking-v1-draft.1`), started 01:46 from a freshly started engine: status complete, **12 of 12 sanity checks**. Three of nine performance requests reached the 512-token cap while reasoning, which the protocol allows. Files: `standard/results.json`, `standard/outputs.jsonl`, `standard/config.json`, `standard/provenance/`. - Engine log of both runs: every setting of the configuration in effect (full command line in `standard/`), DFlash2 drafter on, 17.00 GiB of experts resident in RAM, 8.00 GiB of it registered for the GPU. No engine error, and no new dump in `C:\Windows\LiveKernelReports` during either run. **Failed:** - None observed in the submitted runs. The earlier runs on the overclocked card (driver resets, see "Also in the report") are not part of these results. **Not run:** - The app's default configuration and the #6161 configuration on 0.1.63, so there is no same-day control and the gain is not split between settings and releases. - More than one run per suite, other models, Linux.
Mehr auf der Site
Links zu Install, Modellen, Releases.