← sources

gfx906-llama-cpp: New PP/TG gains for MI50/MI60/Radeon VII/AMD GCN

source: reddit (reddit:LocalLLaMA)published 2026-09-05 14:42ingested 2026-09-06 02:41

retained text

Time for another update! We have been busy and managed to improve the gains substantially (mostly from exploring existing llama cpp PRs and adopting relevant things).

Among other things the [README.md](http://README.md) was also appended to provide a better overall picture of what’s in the fork, why and from whom.

|metric|upstream t/s|fork t/s|gain|
|:-|:-|:-|:-|
|prefill PP16384|332.5|\~410|**+23%**|
|120k deep fill|231.4|\~264|**+14%**|
|TG |13.6|\~15.1|**+11%** (parity pre-mirror)|
|context|cannot fit|250k on 40 GB|tight-fit machinery|
|outputs|\-|\-|bit-identical (sha + token-for-token)|

[https://github.com/milpster/gfx906-llama-cpp/blob/master/README.md](https://github.com/milpster/gfx906-llama-cpp/blob/master/README.md)

(Yes i made this with AI)

  

document fetched 2026-09-06 00:36 · canonical https://reddit.com/r/LocalLLaMA/comments/1w82kpd/gfx906llamacpp_new_pptg_gains_for_mi50mi60radeon