Home › GPUs & Macs › Apple M2 (10-core GPU) vs Apple M4 (10-core GPU)
Apple M2 (10-core GPU) vs Apple M4 (10-core GPU) for local LLMs
Across 6 tests measured on both, the Apple M4 (10-core GPU) is faster in 6 and generates a median 1.12× the tokens/s of the Apple M2 (10-core GPU).
On Llama 2 7B (F16, llama-bench) the Apple M2 (10-core GPU) generates 6.7 tokens/s and the Apple M4 (10-core GPU) 7.4 tokens/s (measured).
Same model, same quant, same test
| Model | Quant / test | M2 (10-core GPU) tok/s | M4 (10-core GPU) tok/s | Faster | Sources and run details |
|---|---|---|---|---|---|
| Llama 2 7B | F16 llama-bench | 6.7 Metal | 7.4 Metal | M4 (10-core GPU) +11% | llama.cpp #4167 M2 (10-core GPU) · llama.cpp #4167 M4 (10-core GPU) |
| Llama 2 7B | Q4_0 llama-bench | 21.9 Metal | 24.1 Metal | M4 (10-core GPU) +10% | llama.cpp #4167 M2 (10-core GPU) · llama.cpp #4167 M4 (10-core GPU) |
| Llama 2 7B | Q8_0 llama-bench | 12.2 Metal | 13.5 Metal | M4 (10-core GPU) +11% | llama.cpp #4167 M2 (10-core GPU) · llama.cpp #4167 M4 (10-core GPU) |
| Llama 3.1 8B Instruct | Q4_K_M LocalScore | 14.7 llamafile | 16.8 llamafile | M4 (10-core GPU) +14% | LocalScore M2 (10-core GPU) · LocalScore M4 (10-core GPU) |
| Llama 3.2 1B Instruct | Q4_K_M LocalScore | 55.6 llamafile | 74.3 llamafile | M4 (10-core GPU) +34% | LocalScore M2 (10-core GPU) · LocalScore M4 (10-core GPU) |
| Qwen2.5 14B Instruct | Q4_K_M LocalScore | 7.5 llamafile | 9.0 llamafile | M4 (10-core GPU) +20% | LocalScore M2 (10-core GPU) · LocalScore M4 (10-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 (10-core GPU) | Apple M4 (10-core GPU) | |
|---|---|---|
| Memory | Unified memory | Unified memory |
| Memory bandwidth | 100 GB/s | 120 GB/s |
All results for the Apple M2 (10-core GPU) · All results for the Apple M4 (10-core GPU) · More comparisons