#2773 Community benchmark: RX 7900 XTX (gfx1100), Ryzen 7 9800X3D - Qwen3.8-Flash-Next IQ2_XS, 131K

Beschreibung

Community benchmark from an RX 7900 XTX (gfx1100) — a first report for this GPU in `bench/results/`, run with `tools/strata_bench.py` (suite 1) against an already-running server.

**Hardware** — AMD Radeon RX 7900 XTX 24 GiB; AMD Ryzen 7 9800X3D; 63.7 GiB RAM; Windows 11 (26100); driver 32.0.31041.1004; model and engine on a Samsung 990 PRO 2 TB NVMe.

**Model** — Qwen3.8-Flash-Next, IQ2_XS (68 GB across two GGUF shards, SHA-256 verified), context 131,072, KV int8 (32K resident), MTP drafts (`--spec 4 --spec-min-p 0.5`).

**Configuration** — install's own settings plus the community tuning switches: `--prefill auto:16384` and env `STRATA_PF_FUSED=1 STRATA_PF_GEMM=1 STRATA_PF_SWITCH_MIN_T=4096 STRATA_HIP_WMMA=1 STRATA_PA_FAST=1 STRATA_ROUTE_TAIL_SKIP=7` (full list in the README). On this machine these cut a 10.9K-token prefill from 14.1 s to 7.0 s.

**Results (medians)** — chat decode 75.7 tok/s, TTFT 0.39 s; 4K prompt 928.9 tok/s; 16K prompt 1,982.9 tok/s; agent session decode 82.3 tok/s with 68.6% conversation-cache reuse; suite time 140 s.

**Limitations** — speed only, no correctness check; background desktop load was present (expert cache filled to 9.0 GiB of VRAM during the run; ~14 GiB with an idle desktop); single machine, single sitting.

Reported by Neil Huang.