Releases·Tool Release
153 tok/s on 1x AMD Radeon R9700 running Qwen3.8 27b NVFP4, 470 tok/s @ 8 conc requests, Prefill @ 3,619 tok/s
A developer released optimizations for single AMD Radeon R9700 cards in their vLLM fork (radiance-vllm-mxfp4), roughly doubling performance across the board. Benchmarks with Unsloth's Qwen3.8 27B NVFP4 show up to 153 tok/s single-stream decode, 471 tok/s at 8 concurrent requests, and ~3,619 tok/s prefill.
272 engagement1 sourceThu, Sep 17, 2026, 03:14 PM
AI Overview
- Optimizations doubled single-R9700 performance across the board, peaking at 153.1 tok/s decode () and 3,619 tok/s prefill at 16k depth.
- Concurrency testing scaled to 471 tok/s at 8 requests vs 120 tok/s single-stream, using Unsloth's Qwen3.8 27B NVFP4.
- Update latency held at 42.3–42.5 ms across most categories, dropping to 34.2–34.3 ms for reasoning and summarization.
- The poster released the code as BetterBench plus vllm-mxfp4 repos on Codeberg and GitHub.
Related
Releases·Tool Release
Dev's vLLM MXFP4 build runs NVFP4 Qwen3.8-27B on AMD R9700 at 5,809 tok/s prefill, 276 tok/s decode
A developer extended their MXFP4 vLLM build on the Radiance image with online NVFP4-to-MXFP4 conversion, letting NVFP4-quantized models use the MXFP4 fast path on AMD R9700 GPUs. Running Unsloth's Qwen3.8-27B NVFP4 quant, they measured 5,809 tok/s prefill and 276 tok/s decode via their BetterBench harness.
A developer extended their MXFP4 vLLM build on the Radiance image with online NVFP4-to-MXFP4 conversion, letting NVFP4-quantized models use the MXFP4 fast path on AMD R9700 GPUs. Running Unsloth's Qwen3.8-27B NVFP4 quant, they measured 5,809 tok/s prefill and 276 tok/s decode via their BetterBench harness.
69 engagement1 sourceredditTue, Sep 15, 2026, 08:51 PM
Social·Community
Users seek real-world R9700 Qwen3.8-27B token rates, cite 51.8 tok/s AMD claim
A Reddit user asks for sustained token-per-second benchmarks of the PowerColor R9700 running Qwen3.8-27B at Q4_K_XL with llama.cpp/Vulkan, noting AMD's 51.8 tok/s figure lacks context-length details. They also cite 5090 benchmarks showing Qwen3.8 performance degrades sharply with context length, and question MTP speculative decoding stability.
A Reddit user asks for sustained token-per-second benchmarks of the PowerColor R9700 running Qwen3.8-27B at Q4_K_XL with llama.cpp/Vulkan, noting AMD's 51.8 tok/s figure lacks context-length details. They also cite 5090 benchmarks showing Qwen3.8 performance degrades sharply with context length, and question MTP speculative decoding stability.
24 engagement1 sourceredditSun, Aug 23, 2026, 09:05 PM