llama.cpp discussion #4167 comment: Mac mini M6 llama-bench (jhen0409, 2026-10-01)

Fields

id

sources/llama-cpp-discussion-4167-comment-18694932-m6-mac-mini

URL

https://github.com/ggml-org/llama.cpp/discussions/4167#discussioncomment-18694932

Publisher

llama.cpp community (jhen0409, collaborator)

Kind

forum

Published

2026-10-01

Accessed

2026-10-16

notes

Single comment on discussion #4167, not the maintainer scoreboard table. Poster (collaborator jhen0409) reports a Mac mini M6, 12-core GPU, 32 GB, 170 GB/s, macOS 27.0.1, llama-bench of ggml-org Llama-2-7B at build 0c1e570: F16 pp512 952.53 / tg128 11.59, Q8_0 pp512 970.46 / tg128 21.01, Q4_0 pp512 988.93 / tg128 35.74, backend BLAS,MTL, threads 2. NOT promoted into the canonical scoreboard row: the summary table on the same page lists an M6 8-core / 170 GB/s row attributed to commit c1d0e7a with the same six numbers, while the comment itself cites build 0c1e570 and 12 GPU cores. That disagreement is why this source is kind=forum and confidence 0.45, and why the numbers were not filed as benchmarks.

Notes

Single comment on discussion #4167, not the maintainer scoreboard table. Poster (collaborator jhen0409) reports a Mac mini M6, 12-core GPU, 32 GB, 170 GB/s, macOS 27.0.1, llama-bench of ggml-org Llama-2-7B at build 0c1e570: F16 pp512 952.53 / tg128 11.59, Q8_0 pp512 970.46 / tg128 21.01, Q4_0 pp512 988.93 / tg128 35.74, backend BLAS,MTL, threads 2. NOT promoted into the canonical scoreboard row: the summary table on the same page lists an M6 8-core / 170 GB/s row attributed to commit c1d0e7a with the same six numbers, while the comment itself cites build 0c1e570 and 12 GPU cores. That disagreement is why this source is kind=forum and confidence 0.45, and why the numbers were not filed as benchmarks.

Cross-references

This record references no other records.