Quad R9700 with vLLM-Radiance hits 17.6k tok/s prefill on EPYC
im_EDEN · reddit · 2026-09-03
A Redditor benchmarked four Radeon AI Pro R9700s running vLLM-Radiance: official vLLM was painfully slow, but the Radiance fork reached 17,636 tok/s prefill and 36.6 tok/s decode, peaking at 106 tok/s with an 80% MTP acceptance rate. Setup: Gigabyte MZ32-AR0 + EPYC 7282, Qwen 3.8 27B fp8 at 262k context, one card on PCIe 4x8. Notably all four cards run at 100% utilization even though only dual-GPU setups are officially supported.
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