M3 Ultra 512GB vs AMD Strix Halo 128GB for local LLMs
41 reports on the M3 Ultra 512GB and 320 on the AMD Strix Halo 128GB
Which is faster for local LLMs?
On Qwen3.8 125B · 6B active at 4-bit, the M3 Ultra 512GB has 28.0 generation tokens/s (2 runs); prompt processing 558.8 tokens/s (one run); the AMD Strix Halo 128GB has 24.5 generation tokens/s (6 runs); prompt processing 1700.5 tokens/s (5 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 512GB (819 GB/s on the M3 Ultra 512GB, 256 on the AMD Strix Halo 128GB). Across 7 shared models with generation measurements on both cards, the M3 Ultra 512GB's median is higher on 7 and the AMD Strix Halo 128GB's on 0.
- Memory
- 512GB unified
- Memory bandwidth
- 819 GB/s
- FP16 compute
- 28.4 TFLOPS
- Reports
- 41
- Memory
- 128GB unified
- Memory bandwidth
- 256 GB/s
- FP16 compute
- 59 TFLOPS
- Reports
- 320
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 512GB | AMD Strix Halo 128GB |
|---|---|---|
| Qwen3.8 125B · 6B active4-bit | 28.02 runsPP 558.8 (one run)llama.cpp, oMLX250K context, 1 not stated | 24.56 runsPP 1700.5 (5 runs)Gufo, llama.cpp, 1 not stated128K to 160K context, 3 not stated |
| Qwen3.6 35B · 3B active8-bit | 81.2one runPP 2324.6 (one run)engine not statedcontext not stated | 50.44 runsPP 1353.3 (4 runs)llama.cpp, Gufo, 2 not stated512K context, 4 not stated |
| Qwen3.8 27B4-bit | 31.8one runPP 391.2 (one run)MLXcontext not stated | 13.14 runsPP 378.2 (3 runs)llama.cpp, 1 not stated32K context, 3 not stated |
| Gemma 4 26B · 4B active4-bit | 87.8one runPP 2595.3 (one run)llama.cppcontext not stated | 67.32 runsPP 4061.0 (2 runs)Gufocontext not stated |
| Qwen3.5 35B · 3B active4-bit | 80.9one runPP 2779.0 (one run)llama.cppcontext not stated | 52.42 runsPP 1017.0 (one run)llama.cpp, 1 not stated128K context, 1 not stated |
| Gemma 4 26B · 4B active8-bit | 84.9one runPP 2634.8 (one run)llama.cppcontext not stated | 41.6one runPP 1698.9 (one run)llama.cppcontext not stated |
| Qwen3.5 122B · 10B active4-bit | 49.5one runPP 784.5 (one run)MLXcontext not stated | 22.4one runPP 287.0 (one run)engine not statedcontext 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 512GB | AMD Strix Halo 128GB |
|---|---|---|
| Qwen3.8 125B · 6B active | 13.6 t/sFP16 | 12.6 t/sQ6_K |
| Qwen3.6 35B · 3B active | 28.7 t/sFP16 | 12.6 t/sFP16 |
| Qwen3.8 27B | 8.0 t/sFP16 | 3.5 t/sFP16 |
| Gemma 4 26B · 4B active | 21.5 t/sFP16 | 9.5 t/sFP16 |
| Qwen3.5 35B · 3B active | 28.7 t/sFP16 | 12.6 t/sFP16 |
| Qwen3.5 122B · 10B active | 8.3 t/sFP16 | 7.7 t/sQ6_K |
| DeepSeek V4 Flash 284B · 13B active | 22.6 t/sQ8_0 | 33.5 t/sQ2_K |
| Qwen3.6 27B | 8.0 t/sFP16 | 3.5 t/sFP16 |
| DeepSeek V4.1 Flash 552B · 16B active | 30.0 t/sQ4_K_M | doesn't fit at these settings |
| GLM-5.3 320B · 18B active | 7.4 t/sQ8_0 | 11.1 t/sQ2_K |
| Qwen3.5 9B | 23.9 t/sFP16 | 10.5 t/sFP16 |
| Gemma 4 31B | 6.9 t/sFP16 | 3.1 t/sFP16 |
Change the context, RAM or model in the calculator: M3 Ultra 512GB · AMD Strix Halo 128GB
Every report on each card, with its source: M3 Ultra 512GB · AMD Strix Halo 128GB