Qwen3.8 27B hits 280 tok/s on 2x R9700 with MXFP4, beating FP8 at hardware limits
whodoneit1 · reddit · 2026-09-02
- The author kept optimizing dual AMD Radeon R9700 local inference; the community has grown from 5 to 1,200 collaborating developers
- Built MXFP4 support on DeadCode's radiance image using W4A8 kernels, first matching then clearly beating FP8 throughput — now near the hardware limit of these cards
- BetterBench decode: 280 tok/s on JSON, 254 math, 226 code, 148 chat; prefill sustains 3,831 tok/s even at 94k prompt tokens
- KV cache reaches 940k tokens; the full image and repo are open source
More from Infra
- SGLang Summit sponsors grow from 11 to 36 in one week; Lip-Bu Tan, Lilian Weng to speak — BanghuaZ · 2026-10-18
- Yacine jokes his datacenter KPI is gigawatt-hours: get paged if not burning enough power — yacineMTB · 2026-10-18
- 1.68x inference speedup drops, directly accelerating GRPO post-training — yacineMTB · 2026-10-18
- OpenRouter is only a fraction of the total addressable market, poster argues — sloppenheimer · 2026-10-18
- Developer builds change alerts for LLM prices and retirements across Bedrock, Azure and Vertex — jsandhu0780 · 2026-10-18
- 12.3M tokens in 27 minutes for $0.14: agent session with 99.6% cache hit — teortaxesTex · 2026-10-18