llamaperf
Submit

Unknown family

on AMD MI50 32GB · llama.cpp

Tone: positiveAMD hardware
Sep 9, 2026
Throughput
15.1 t/s gen · 410.0 tokens/s prompt processing (prefill)Prompt processing measures input prefill speed; it does not tell you the time to first token.
VRAM reported
32 GB

Summary

User benchmarks the gfx906-llama-cpp fork on an MI50 32GB, reaching 410 t/s prefill at PP16384 against 332.5 t/s upstream, 264 t/s on a 120k deep fill against 231.4 t/s upstream, and 15.1 t/s TG against 13.6 t/s upstream. Outputs are bit-identical to upstream. The user attributes the gains to exploring existing llama.cpp PRs.
View source ↗