Discussion on "I Tested ROCm vs Vulkan for AMD Local LLMs — One Backend Silently Tanked by 6x"

Rost Glukhov · 2026-09-15T11:40:35.372Z

The same model, the same AMD GPU, two different compute stacks. One hit 143 tokens per second. The other — on a slightly bigger model — quietly collapsed to a sixth of its neighbour's speed, and the s

Discussion on "I Tested ROCm vs Vulkan for AMD Local LLMs — One Backend Silently Tanked by 6x" | Hashnode

Search Hashnode

I Tested ROCm vs Vulkan for AMD Local LLMs — One Backend Silently Tanked by 6x

Responses