speed-bench: add Mac Studio M1 Ultra 128 GB numbers - #706
Open
EnMerovingi wants to merge 1 commit into
Open
Conversation
Run with the canonical README command (ctx 2048..65536, --step-incr 2048, --gen-tokens 128) on the ds4f-q2 imatrix GGUF, Metal backend, macOS 26.1. M1 Ultra sits in an unusual spot: generation is close to M2 Ultra (23.5 -> 18.5 t/s from 2k to 64k) while prefill is roughly half of it, around 225 t/s at 2k. Note that the Metal 4 tensor API path is disabled on pre-M5 devices, which is visible in the prefill column.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Hi! Thank you so much for making this work of art public.
I will be glad to contribute more going forward; for now I am only adding this benchmark result. Thanks to everyone involved, and especially to Salvatore Sanfilippo (@antirez) for supporting a community that fights for a better world.
Adds
speed-bench/m1_ultra.csvand the generatedspeed-bench/m1_ultra_ts.svg. There were no M1 Ultra numbers in the directory yet.Setup
ds4f-q2imatrix GGUF (DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-0731)iogpu.wired_limit_mb=112640speed-bench/README.md:Results
Generation is very close to the existing M2 Ultra CSV (23.5 -> 18.5 t/s across the sweep, versus 23.2 -> 20.4), which matches the two chips sharing the same 800 GB/s memory bandwidth. Prefill is roughly half of M2 Ultra (225 vs 410 t/s at 2k), which is where the older GPU shows. The engine prints
Metal 4 tensor API disabled for pre-M5/pre-A19 deviceson this machine.Memory footprint was 82.12 GiB planned at 64k context (80.76 GiB model + 0.86 GiB KV), so there is comfortable headroom for much longer contexts on a 128 GB machine.