Swapping 5600X for 5600G slashes Vulkan LLM inference 4x while ROCm stays unaffected
Jebbyk1 · reddit · 2026-10-30
A local-LLM user replaced a 5600X with a 5600G to free VRAM and saw Vulkan inference on an RX 7900 XTX (qwen3.8-27b) drop from 65-75 t/s to 15-20 t/s, with the iGPU at 0% load and the main GPU pegged — while ROCm performance was unchanged. The thread asks what's wrong with Vulkan, likely an APU/memory-topology or driver scheduling issue.
More from Infra
- SGLang RFC decouples model code from parallelism so new models skip reimplementation — BanghuaZ · 2026-10-30
- Jassy says AWS will be a $1T revenue business — after years of heavy AI capex — TechCrunch AI · 2026-10-30
- Even 2020-era GPUs keep gaining rental value, undercutting the AI bubble thesis — JFPuget · 2026-10-30
- Intel earnings call: server CPU TAM seen 5x to $200B+, chip shortage into 2028 — ryanshrout · 2026-10-30
- Lightweight FastAPI proxy gateway adds auth for locally hosted LLMs in OpenCode — Frizzy-MacDrizzle · 2026-10-30
- Nvidia reportedly in talks to backstop >$5B on Nscale's 460 MW site, locking in Anthropic's $45B Vera Rubin deal — ns123abc · 2026-10-30