Home › GPUs & Macs › Apple M3 Max (30-core GPU) vs Apple M5 Pro (20-core GPU)

Apple M3 Max (30-core GPU) vs Apple M5 Pro (20-core GPU) for local LLMs

Across 6 tests measured on both, the Apple M5 Pro (20-core GPU) is faster in 6 and generates a median 1.18× the tokens/s of the Apple M3 Max (30-core GPU).

On Llama 2 7B (F16, llama-bench) the Apple M3 Max (30-core GPU) generates 19.5 tokens/s and the Apple M5 Pro (20-core GPU) 21.6 tokens/s (measured).

Same model, same quant, same test

ModelQuant / testM3 Max (30-core GPU)
tok/s
M5 Pro (20-core GPU)
tok/s
FasterSources and run details
Llama 2 7BF16
llama-bench
19.5
Metal
21.6
Metal
M5 Pro (20-core GPU) +10%llama.cpp #4167 M3 Max (30-core GPU) · llama.cpp #4167 M5 Pro (20-core GPU)
Llama 2 7BQ4_0
llama-bench
56.6
Metal
66.3
Metal
M5 Pro (20-core GPU) +17%llama.cpp #4167 M3 Max (30-core GPU) · llama.cpp #4167 M5 Pro (20-core GPU)
Llama 2 7BQ8_0
llama-bench
34.3
Metal
38.9
Metal
M5 Pro (20-core GPU) +13%llama.cpp #4167 M3 Max (30-core GPU) · llama.cpp #4167 M5 Pro (20-core GPU)
Llama 3.1 8B InstructQ4_K_M
LocalScore
37.6
llamafile
44.7
llamafile
M5 Pro (20-core GPU) +19%LocalScore M3 Max (30-core GPU) · LocalScore M5 Pro (20-core GPU)
Llama 3.2 1B InstructQ4_K_M
LocalScore
133
llamafile
175
llamafile
M5 Pro (20-core GPU) +32%LocalScore M3 Max (30-core GPU) · LocalScore M5 Pro (20-core GPU)
Qwen2.5 14B InstructQ4_K_M
LocalScore
20.1
llamafile
24.0
llamafile
M5 Pro (20-core GPU) +20%LocalScore M3 Max (30-core GPU) · LocalScore M5 Pro (20-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 M3 Max (30-core GPU)Apple M5 Pro (20-core GPU)
MemoryUnified memoryUnified memory
Memory bandwidth300 GB/s307 GB/s

All results for the Apple M3 Max (30-core GPU) · All results for the Apple M5 Pro (20-core GPU) · More comparisons