~/gpu/m3-ultra-512 vs m5-max-128

M3 Ultra 512 manufacturerM3 Ultra 512vsM5 Max 128 manufacturerM5 Max 128

Head-to-head for local LLM inference. The honest comparison: VRAM, bandwidth, compute, and which of the 30 catalog models actually fit on each.

The specs.

$ diff specs m3-ultra-512 m5-max-128
Stat
m3-ultra-512
m5-max-128
Δ
VRAM
512 GB
128 GB
-75%
Memory bandwidth
819 GB/s
614 GB/s
-25%
FP16 compute
80 TFLOPS
55 TFLOPS
-31%
Weights budget at 8K ctx
413 GB
103 GB
-75%

Model fit difference.

$ models that change with the card
Fits on both
27of 30
Only on m3-ultra-512
3
Only on m5-max-128
0

// showing 12 of 30 models; differing fits first

Model
m3-ultra-512
m5-max-128
Llama 3.1 405B405B
fitsFP8/INT8
overQ4_K_M
DeepSeek V3671B
fitsQ4_K_M
overQ4_K_M
DeepSeek R1671B
fitsQ4_K_M
overQ4_K_M
Llama 3.2 1B1.23B
fitsFP16/BF16
fitsFP16/BF16
Llama 3.2 3B3.21B
fitsFP16/BF16
fitsFP16/BF16
Llama 3.1 8B8B
fitsFP16/BF16
fitsFP16/BF16
Llama 3.3 70B70.6B
fitsFP16/BF16
fitsQ8_0
Qwen 2.5 7B7B
fitsFP16/BF16
fitsFP16/BF16
Qwen 2.5 32B32.5B
fitsFP16/BF16
fitsFP16/BF16
Qwen 2.5 Coder 32B32.5B
fitsFP16/BF16
fitsFP16/BF16
Qwen 2.5 72B72.7B
fitsFP16/BF16
fitsQ8_0
Qwen3 30B A3B30.5B
fitsFP16/BF16
fitsFP16/BF16

Which one wins for…

$ ./recommend --by-workload
More VRAM headroom

M3 Ultra 512 has 384 GB more.

Faster decode (bandwidth)

M3 Ultra 512 by +33%.

Faster prefill (compute)

M3 Ultra 512 by +45% TFLOPS.

Catalog models that fit

M3 Ultra 512: 30 fit · M5 Max 128: 27.

Drill into either card.

$ ./vrambudget --gpu
$ openM3 Ultra 512512GBapple · monster$ openM5 Max 128128GBapple · flagship

Discussion.

$ gh discussion list

// sign in with github to leave a comment. threads live in the repo's discussions tab.