Home › GPUs & Macs › Apple M2 Ultra (76-core GPU) vs Apple M3 Ultra (80-core GPU)
Apple M2 Ultra (76-core GPU) vs Apple M3 Ultra (80-core GPU) for local LLMs
Across 5 tests measured on both, the two are about even (median ratio 1.02×).
On Llama 2 7B (F16, llama-bench) the Apple M2 Ultra (76-core GPU) generates 41.0 tokens/s and the Apple M3 Ultra (80-core GPU) 39.8 tokens/s (measured).
Same model, same quant, same test
| Model | Quant / test | M2 Ultra (76-core GPU) tok/s | M3 Ultra (80-core GPU) tok/s | Faster | Sources and run details |
|---|---|---|---|---|---|
| Llama 2 7B | F16 llama-bench | 41.0 Metal | 39.8 Metal | M2 Ultra (76-core GPU) +3% | llama.cpp #4167 M2 Ultra (76-core GPU) · llama.cpp #4167 M3 Ultra (80-core GPU) |
| Llama 2 7B | Q4_0 llama-bench | 94.3 Metal | 92.1 Metal | even | llama.cpp #4167 M2 Ultra (76-core GPU) · llama.cpp #4167 M3 Ultra (80-core GPU) |
| Llama 2 7B | Q8_0 llama-bench | 66.6 Metal | 63.9 Metal | M2 Ultra (76-core GPU) +4% | llama.cpp #4167 M2 Ultra (76-core GPU) · llama.cpp #4167 M3 Ultra (80-core GPU) |
| gpt-oss 20B | MXFP4 llama-bench | 116 Metal | 116 Metal | even | llama.cpp #15396 M2 Ultra (76-core GPU) · llama.cpp #15396 M3 Ultra (80-core GPU) |
| Qwen2.5 14B Instruct | Q4_K_M LocalScore | 36.6 llamafile | 36.3 llamafile | even | LocalScore M2 Ultra (76-core GPU) · LocalScore M3 Ultra (80-core GPU) |
Generation speed in tokens/s, best measured backend for each chip (shown under the number). Backends can differ between the two chips; check them before drawing conclusions. Measured results only, nothing estimated.
Specs
| Apple M2 Ultra (76-core GPU) | Apple M3 Ultra (80-core GPU) | |
|---|---|---|
| Memory | Unified memory | Unified memory |
| Memory bandwidth | 800 GB/s | 819 GB/s |
All results for the Apple M2 Ultra (76-core GPU) · All results for the Apple M3 Ultra (80-core GPU) · More comparisons