Pull requests / #1173

#1173 bench: community report, RX 7900 XTX on Windows 11, IQ3_S, engine 0.1.40

closed · @nytasuk-commits · 0 コメント · GitHub で見る

BenchmarksSetup & installAMD / HIPModels & quantsDocumentationWindows

本文

Results-only submission, no engine changes. Adds bench/results/2026-10-06-community-rx-7900-xtx/ following docs/COMMUNITY_BENCHMARKS.md.

Hardware: RX 7900 XTX 24 GB, Ryzen 9 7950X3D, 192 GB RAM, Windows 11 Pro (build 26200), AMD driver 32.0.31041.1004.
Software: Strata commit 1735d6471df29b42c26170efaac1f1446a58640f, engine 0.1.40, ready-made AMD engine installed by START-HERE.bat.
Model: original Flash-Next IQ3_S, 262,144 context, int8 KV, --prefill auto:32768, --conversation-cache-mib 8192.

Configurations tested: 3 fresh runs each at ~100 and ~38K prompt tokens (cache_n 0 on all), plus 6 needle checks at 32K and 128K (6 of 6 found).
Headline: about 1,612 tok/s prompt reading and about 78-80 tok/s decode (medians).

Limitations: one machine, 3 runs per configuration, non-streaming (no TTFT), VRAM/PCIe link width/GPU power not recorded. Windows HIP is listed as outside the validated scope in AMD_HIP.md; this is a data point on that gap. Raw data (runs.csv, needles.json) and the run config are in the folder.

関連リンク

インストール・モデル・リリースへの站内リンク。