Home › GPUs & Macs › Apple M1 Max (32-core GPU) vs Apple M2 Max (38-core GPU)

Apple M1 Max (32-core GPU) vs Apple M2 Max (38-core GPU) for local LLMs

Across 6 tests measured on both, the Apple M2 Max (38-core GPU) is faster in 6 and generates a median 1.09× the tokens/s of the Apple M1 Max (32-core GPU).

On Llama 2 7B (F16, llama-bench) the Apple M1 Max (32-core GPU) generates 23.0 tokens/s and the Apple M2 Max (38-core GPU) 24.6 tokens/s (measured).

Same model, same quant, same test

ModelQuant / testM1 Max (32-core GPU)
tok/s		M2 Max (38-core GPU)
tok/s		FasterSources and run details
Llama 2 7BF16
llama-bench		23.0
Metal		24.6
Metal		M2 Max (38-core GPU) +7%llama.cpp #4167 M1 Max (32-core GPU) · llama.cpp #4167 M2 Max (38-core GPU)
Llama 2 7BQ4_0
llama-bench		61.2
Metal		66.0
Metal		M2 Max (38-core GPU) +8%llama.cpp #4167 M1 Max (32-core GPU) · llama.cpp #4167 M2 Max (38-core GPU)
Llama 2 7BQ8_0
llama-bench		40.2
Metal		41.8
Metal		M2 Max (38-core GPU) +4%llama.cpp #4167 M1 Max (32-core GPU) · llama.cpp #4167 M2 Max (38-core GPU)
Llama 3.1 8B InstructQ4_K_M
LocalScore		36.2
llamafile		45.5
llamafile		M2 Max (38-core GPU) +26%LocalScore M1 Max (32-core GPU) · LocalScore M2 Max (38-core GPU)
Llama 3.2 1B InstructQ4_K_M
LocalScore		121
llamafile		145
llamafile		M2 Max (38-core GPU) +20%LocalScore M1 Max (32-core GPU) · LocalScore M2 Max (38-core GPU)
Qwen2.5 14B InstructQ4_K_M
LocalScore		19.8
llamafile		22.0
llamafile		M2 Max (38-core GPU) +11%LocalScore M1 Max (32-core GPU) · LocalScore M2 Max (38-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 M1 Max (32-core GPU)Apple M2 Max (38-core GPU)
MemoryUnified memoryUnified memory
Memory bandwidth400 GB/s400 GB/s

All results for the Apple M1 Max (32-core GPU) · All results for the Apple M2 Max (38-core GPU) · More comparisons