Home › GPUs & Macs › Apple M1 Ultra (48-core GPU) vs Apple M2 Ultra (60-core GPU)

Apple M1 Ultra (48-core GPU) vs Apple M2 Ultra (60-core GPU) for local LLMs

Across 6 tests measured on both, the Apple M2 Ultra (60-core GPU) is faster in 6 and generates a median 1.20× the tokens/s of the Apple M1 Ultra (48-core GPU).

On Llama 2 7B (F16, llama-bench) the Apple M1 Ultra (48-core GPU) generates 33.9 tokens/s and the Apple M2 Ultra (60-core GPU) 39.9 tokens/s (measured).

Same model, same quant, same test

ModelQuant / testM1 Ultra (48-core GPU)
tok/s
M2 Ultra (60-core GPU)
tok/s
FasterSources and run details
Llama 2 7BF16
llama-bench
33.9
Metal
39.9
Metal
M2 Ultra (60-core GPU) +18%llama.cpp #4167 M1 Ultra (48-core GPU) · llama.cpp #4167 M2 Ultra (60-core GPU)
Llama 2 7BQ4_0
llama-bench
74.9
Metal
88.6
Metal
M2 Ultra (60-core GPU) +18%llama.cpp #4167 M1 Ultra (48-core GPU) · llama.cpp #4167 M2 Ultra (60-core GPU)
Llama 2 7BQ8_0
llama-bench
55.7
Metal
62.1
Metal
M2 Ultra (60-core GPU) +12%llama.cpp #4167 M1 Ultra (48-core GPU) · llama.cpp #4167 M2 Ultra (60-core GPU)
Llama 3.1 8B InstructQ4_K_M
LocalScore
48.9
llamafile
59.4
llamafile
M2 Ultra (60-core GPU) +21%LocalScore M1 Ultra (48-core GPU) · LocalScore M2 Ultra (60-core GPU)
Llama 3.2 1B InstructQ4_K_M
LocalScore
138
llamafile
174
llamafile
M2 Ultra (60-core GPU) +26%LocalScore M1 Ultra (48-core GPU) · LocalScore M2 Ultra (60-core GPU)
Qwen2.5 14B InstructQ4_K_M
LocalScore
27.8
llamafile
34.1
llamafile
M2 Ultra (60-core GPU) +23%LocalScore M1 Ultra (48-core GPU) · LocalScore M2 Ultra (60-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 Ultra (48-core GPU)Apple M2 Ultra (60-core GPU)
MemoryUnified memoryUnified memory
Memory bandwidth800 GB/s800 GB/s

All results for the Apple M1 Ultra (48-core GPU) · All results for the Apple M2 Ultra (60-core GPU) · More comparisons