AMD MI50 32GB vs AMD Strix Halo 128GB for local LLMs
13 reports on the AMD MI50 32GB and 111 on the AMD Strix Halo 128GB
Which is faster for local LLMs?
On Ornith1.5 35B · 3B active at 4-bit, the AMD MI50 32GB runs at 108.9 tokens per second (one run) and the AMD Strix Halo 128GB at 72.6 (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 4.0x higher on the AMD MI50 32GB (1,024 GB/s on the AMD MI50 32GB, 256 on the AMD Strix Halo 128GB). Across the 3 models both have been measured on, the AMD MI50 32GB's median is higher on 2 and the AMD Strix Halo 128GB's on 1.
- Memory
- 32GB
- Memory bandwidth
- 1,024 GB/s
- FP16 compute
- 26.5 TFLOPS
- Reports
- 13
- Memory
- 128GB unified
- Memory bandwidth
- 256 GB/s
- FP16 compute
- 59 TFLOPS
- Reports
- 111
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 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 | AMD MI50 32GB | AMD Strix Halo 128GB |
|---|---|---|
| Ornith1.5 35B · 3B active4-bit | 108.9one runengine not statedcontext not stated | 72.6one runllama.cppcontext not stated |
| Qwen3-Coder 30B · 3B active4-bit | 66.1one runllama.cppcontext not stated | 70.3one runllama.cpp8K context |
| Qwen3.8 27B8-bit | 22.8one runengine not stated65K context | 7.2one 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 | AMD MI50 32GB | AMD Strix Halo 128GB |
|---|---|---|
| Ornith1.5 35B · 3B active | 93.1 t/sQ5_K_M | 12.6 t/sFP16 |
| Qwen3-Coder 30B · 3B active | 104.7 t/sQ5_K_M | 12.6 t/sFP16 |
| Qwen3.8 27B | 25.7 t/sQ6_K | 3.5 t/sFP16 |
| Qwen3.8 125B · 6B active | 19.6 t/sQ2_K, experts in RAM | 12.6 t/sQ6_K |
| DeepSeek V4 Flash 284B · 13B active | doesn't fit at these settings | 33.5 t/sQ2_K |
| Qwen3.6 35B · 3B active | 93.1 t/sQ5_K_M | 12.6 t/sFP16 |
| Qwen3.6 27B | 25.7 t/sQ6_K | 3.5 t/sFP16 |
| GLM-5.3 320B · 18B active | doesn't fit at these settings | 11.1 t/sQ2_K |
| Gemma 4 26B · 4B active | 65.4 t/sQ6_K | 9.5 t/sFP16 |
Change the context, RAM or model in the calculator: AMD MI50 32GB · AMD Strix Halo 128GB
Every report on each card, with its source: AMD MI50 32GB · AMD Strix Halo 128GB