Six models. Sixty tasks. Methodology included. Why Vulkan (and not ROCm) is the path forward on gfx1151 this generation.
ROCm on gfx1151 has a hang we could not work around — long-context generations lock the compute queue and only a full GPU reset recovers. Vulkan does not. So every number we publish is a Vulkan number, and we document the backend on every row.
The 60-task suite is the same one harness-go ships with: bug fixes, small features, refactors, test authoring. Scored on pass@1 with a deterministic test runner, not vibes. Models include the heavy-weights we run locally and frontier baselines pulled via API so the numbers can be directly compared.
Throughput is the interesting part. On 128GB unified memory we fit weights that do not fit on a single consumer GPU, and on a Vulkan backend we are getting respectable tokens/sec on the 100B-class weights. For a regulated customer who cannot ship prompts to a third party, these numbers change the shape of what is possible.
Full table, methodology, and the harness code live on me.nerds.run — refreshed monthly as new weights land.