M3 Ultra 96GB vs AMD Strix Halo 128GB for local LLMs
10 reports on the M3 Ultra 96GB and 318 on the AMD Strix Halo 128GB
Which is faster for local LLMs?
On gpt-oss 117B · 5.1B active at 4-bit, the M3 Ultra 96GB has 73.0 generation tokens/s (one run); prompt processing 347.0 tokens/s (one run); the AMD Strix Halo 128GB has 52.6 generation tokens/s (4 runs); prompt processing 572.5 tokens/s (4 runs). 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 3.2x higher on the M3 Ultra 96GB (819 GB/s on the M3 Ultra 96GB, 256 on the AMD Strix Halo 128GB). Across 3 shared models with generation measurements on both cards, the M3 Ultra 96GB's median is higher on 3 and the AMD Strix Halo 128GB's on 0.
- Memory
- 96GB unified
- Memory bandwidth
- 819 GB/s
- FP16 compute
- 21.3 TFLOPS
- Reports
- 10
- Memory
- 128GB unified
- Memory bandwidth
- 256 GB/s
- FP16 compute
- 59 TFLOPS
- Reports
- 318
Measured on both
Generation and prompt processing speeds 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 bit class. Each side shows the engines and reported context behind its median. Context can mean a configured window or prompt depth, so these are community comparisons with differing settings. How to read these
| Model | M3 Ultra 96GB | AMD Strix Halo 128GB |
|---|---|---|
| gpt-oss 117B · 5.1B active4-bit | 73.0one runPP 347.0 (one run)MLXcontext not stated | 52.64 runsPP 572.5 (4 runs)llama.cpp, 2 not statedcontext not stated |
| Qwen3.5 35B · 3B active4-bit | 109.7one runPP 1770.1 (one run)MLXcontext not stated | 52.42 runsPP 1017.0 (one run)llama.cpp, 1 not stated128K context, 1 not stated |
| Qwen3.5 35B · 3B active8-bit | 73.4one runPP 1808.0 (one run)llama.cppcontext not stated | 49.7one runllama.cppcontext not stated |
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.
| Model | M3 Ultra 96GB | AMD Strix Halo 128GB |
|---|---|---|
| gpt-oss 117B · 5.1B active | 68.5 t/sQ3_K_M | 15.8 t/sQ6_K |
| Qwen3.5 35B · 3B active | 57.1 t/sQ8_0 | 12.6 t/sFP16 |
| Qwen3.8 27B | 7.9 t/sFP16 | 3.5 t/sFP16 |
| Qwen3.8 125B · 6B active | 55.1 t/sQ3_K_M | 12.6 t/sQ6_K |
| Qwen3.6 35B · 3B active | 57.1 t/sQ8_0 | 12.6 t/sFP16 |
| DeepSeek V4 Flash 284B · 13B active | doesn't fit at these settings | 33.5 t/sQ2_K |
| Qwen3.6 27B | 7.9 t/sFP16 | 3.5 t/sFP16 |
| Gemma 4 26B · 4B active | 19.0 t/sFP16 | 9.5 t/sFP16 |
Change the context, RAM or model in the calculator: M3 Ultra 96GB · AMD Strix Halo 128GB
Every report on each card, with its source: M3 Ultra 96GB · AMD Strix Halo 128GB