← sources

swapped 5600x to 5600g and got 4x slower inference on vulkan

source: reddit (reddit:LocalLLaMA)published 2026-10-29 22:24ingested 2026-10-30 02:40

retained text

\- qwen/qwen3.8-27b fully in main GPU VRAM   
\- rx 7900 xtx 24gb  
\- ddr4 32gb 3200  
\- installed 5600g instead of 5600x to free some VRAM on main gpu  
\- disabled iGPU in LM Studio

got 15-20t\\s instead of 65-75. iGPU show 0% load. Main GPU - 100% load.   
ROCm performance didn't changed.  
What is wrong with Vulkan?

document fetched 2026-10-30 00:44 · canonical https://reddit.com/r/LocalLLaMA/comments/1xiy7of/swapped_5600x_to_5600g_and_got_4x_slower