llama.cpp PR #28398: hardware v_perm_b32 for Q1_0 vec_dot on AMD (+110% decode on gfx1201)

Fields

id

sources/bench-llamacpp-pr-28398-radeon-r9700-q1

URL

https://github.com/ggml-org/llama.cpp/pull/28398

Publisher

llama.cpp contributors (community)

Kind

forum

Accessed

2026-10-03

notes

GitHub PR #28398 in ggml-org/llama.cpp about using hardware v_perm_b32 for Q1_0 vec_dot on AMD. Published 2026-09-04 by The-Monk. +110% decode on gfx1201 (Radeon R9700). Verified 2026-10-10: Read PR 'CUDA/HIP: use hardware v_perm_b32 for Q1_0 vec_dot on AMD (+110% decode on gfx1201)' (https://github.com/ggml-org/llama.cpp/pull/28398); confirms implementation and benchmarks by The-Monk showing mean 29.402 -> 61.657 t/s (+109.7%) decode on gfx1201.

Notes

GitHub PR #28398 in ggml-org/llama.cpp about using hardware v_perm_b32 for Q1_0 vec_dot on AMD. Published 2026-09-04 by The-Monk. +110% decode on gfx1201 (Radeon R9700). Verified 2026-10-10: Read PR 'CUDA/HIP: use hardware v_perm_b32 for Q1_0 vec_dot on AMD (+110% decode on gfx1201)' (https://github.com/ggml-org/llama.cpp/pull/28398); confirms implementation and benchmarks by The-Monk showing mean 29.402 -> 61.657 t/s (+109.7%) decode on gfx1201.

Cross-references

This record references no other records.