M3 Ultra (80-core GPU) — LLM Benchmarks

Measured LLM inference benchmarks for M3 Ultra (80-core GPU) across all RAM configurations (256 GB, 512 GB). 6 benchmark rows across 3 models. Compare how RAM affects throughput. Real runs, not estimates.

6Benchmark rows

3Models tested

2RAM configurations

178.8Fastest avg tok/s

RAM configurations

Each configuration differs only in unified memory. More RAM = larger models fit. Throughput is similar across RAM tiers at the same model size.

256 GBM3 Ultra (80-core GPU, 256 GB)3 rows · 3 models177.9 tok/s peak 512 GBM3 Ultra (80-core GPU, 512 GB)3 rows · 3 models178.8 tok/s peak

All benchmark rows — M3 Ultra (80-core GPU)

Sorted by avg tok/s descending. Click source badge to see original measurement.

Chip (RAM)	Model	Quant	RAM req.	Avg tok/s	Prompt tok/s	Runtime	Source
M3 Ultra (80-core GPU, 512 GB)	Llama 3.2 1B Instruct	Q4_K - Medium	—	178.8 tok/s	5601.0 tok/s	—	ref
M3 Ultra (80-core GPU, 256 GB)	Llama 3.2 1B Instruct	Q4_K - Medium	—	177.9 tok/s	4999.5 tok/s	—	ref
M3 Ultra (80-core GPU, 256 GB)	Llama 3.1 8B Instruct	Q4_K - Medium	—	63.3 tok/s	1062.2 tok/s	—	ref
M3 Ultra (80-core GPU, 512 GB)	Llama 3.1 8B Instruct	Q4_K - Medium	—	62.7 tok/s	1109.5 tok/s	—	ref
M3 Ultra (80-core GPU, 256 GB)	Qwen 2.5 14B Instruct	Q4_K - Medium	—	36.7 tok/s	568.3 tok/s	—	ref
M3 Ultra (80-core GPU, 512 GB)	Qwen 2.5 14B Instruct	Q4_K - Medium	—	35.8 tok/s	577.4 tok/s	—	ref

Models tested on M3 Ultra (80-core GPU)

Llama 3.1 8B Instruct Llama 3.2 1B Instruct Qwen 2.5 14B Instruct

Data

benchmarks.json — full dataset · chips.json — chip summaries · benchmarks.csv — CSV export

Data sourced from factory lab measurements and community reference runs. See all chips →