llamaperf

AMD MI50 32GB vs M5 Ultra 96GB for local LLMs

12 reports on the AMD MI50 32GB and 3 on the M5 Ultra 96GB

Which is faster for local LLMs?

On Qwen3.8 27B at 8-bit, the AMD MI50 32GB runs at 22.8 tokens per second (one run) and the M5 Ultra 96GB at 15.0 (one run). Those are community reports whose engines and settings differ, so the gap is a rough sign, not a controlled measure of the two cards. Memory bandwidth, which caps how fast a card can write, is 20% higher on the M5 Ultra 96GB (1,024 GB/s on the AMD MI50 32GB, 1,229 on the M5 Ultra 96GB).

AMD MI50 32GB
Memory
32GB
Memory bandwidth
1,024 GB/s
FP16 compute
26.5 TFLOPS
Reports
12
M5 Ultra 96GB
Memory
96GB unified
Memory bandwidth
1,229 GB/s
FP16 compute
not on record
Reports
3

Measured on both

Median tokens per second from plain runs (one device, one request, no speculative decoding, the whole model in the device's memory) of the same model size at the same quant level. Engines and context lengths can still differ between the runs. How to read these

ModelAMD MI50 32GBM5 Ultra 96GB
Qwen3.8 27B8-bit22.8one run15.0one run

Estimated by the calculator

The quant the calculator recommends for each card and its estimated speed at a 32,768-token context with 32 GB of system RAM, from memory bandwidth and the model's shape. Where a model was also measured above, the measurement wins.

ModelAMD MI50 32GBM5 Ultra 96GB
Qwen3.8 27B25.7 t/sQ6_K14.9 t/sFP16
Qwen3.8 125B · 6B active19.6 t/sQ2_K, experts in RAM103.1 t/sQ3_K_M
DeepSeek V4 Flash 284B · 13B activedoesn't fit at these settingsdoesn't fit at these settings
Qwen3.6 35B · 3B active93.1 t/sQ5_K_M107.0 t/sQ8_0
Qwen3.6 27B25.7 t/sQ6_K14.9 t/sFP16
Gemma 4 26B · 4B active65.4 t/sQ6_K35.5 t/sFP16
DeepSeek V4.1 Flash 552B · 16B activedoesn't fit at these settingsdoesn't fit at these settings

Change the context, RAM or model in the calculator: AMD MI50 32GB · M5 Ultra 96GB

Every report on each card, with its source: AMD MI50 32GB · M5 Ultra 96GB

More comparisons