R9V: A designer set of kernels I've been working on for R9700s/RDNA4. Qwen3.8-Flash-Next Unsloth IQ4_XS (w/ TP on 2 R9700s, MTP, SSD n-gram, 128k ctx, vision): TG256 of *78 tok/s* (~3x increase), PP8192 of *1510 tok/s* (~30x increase).

A developer has created R9V, a set of kernels for R9700s/RDNA4, achieving significant performance increases. Using Qwen3.8-Flash-Next Unsloth IQ4_XS with specific configurations, the developer reported a TG256 speed of 78 tok/s (a ~3x increase) and a PP8192 speed of 1510 tok/s (a ~30x increase). While demonstrating substantial advantages over llama.cpp ROCm and Vulkan, the R9V kernels also caused system crashes.