NVIDIA RTX 5090 vs AMD Strix Halo 128GB for local LLMs
171 reports on the NVIDIA RTX 5090 and 321 on the AMD Strix Halo 128GB
Which is faster for local LLMs?
On Qwen3.8 27B at 4-bit, the NVIDIA RTX 5090 has 80.0 generation tokens/s (7 runs); the AMD Strix Halo 128GB has 13.1 generation tokens/s (4 runs); prompt processing 378.2 tokens/s (3 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 7.0x higher on the NVIDIA RTX 5090 (1,792 GB/s on the NVIDIA RTX 5090, 256 on the AMD Strix Halo 128GB). Across 7 shared models with generation measurements on both cards, the NVIDIA RTX 5090's median is higher on 7 and the AMD Strix Halo 128GB's on 0.
- Memory
- 32GB
- Memory bandwidth
- 1,792 GB/s
- FP16 compute
- 419 TFLOPS
- Reports
- 171
- Memory
- 128GB unified
- Memory bandwidth
- 256 GB/s
- FP16 compute
- 59 TFLOPS
- Reports
- 321
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 | NVIDIA RTX 5090 | AMD Strix Halo 128GB |
|---|---|---|
| Qwen3.8 27B4-bit | 80.07 runsllama.cpp, TensorSharp +1, 2 not stated40K to 256K context, 2 not stated | 13.14 runsPP 378.2 (3 runs)llama.cpp, 1 not stated32K context, 3 not stated |
| Qwen3.6 35B · 3B active4-bit | 215.1one runllama.cppcontext not stated | 60.07 runsPP 1005.3 (7 runs)llama.cpp, 3 not stated30K context, 6 not stated |
| Qwen3 30B · 3B active4-bit | 352.1one runllama.cppcontext not stated | 73.73 runsPP 920.0 (3 runs)llama.cpp48K context, 2 not stated |
| Gemma 4 31B4-bit | 64.0one runOllama32K context | 11.52 runsPP 559.3 (2 runs)Gufocontext not stated |
| Gemma 4 26B · 4B active4-bit | 243.9one runPP 13809.2 (one run)llama.cppcontext not stated | 67.32 runsPP 4061.0 (2 runs)Gufocontext not stated |
| Qwen3.5 35B · 3B active4-bit | 194.0one runllama.cppcontext not stated | 52.42 runsPP 1017.0 (one run)llama.cpp, 1 not stated128K context, 1 not stated |
| Qwen3.5 27B4-bit | 30.1one runPP 1515.8 (one run)llama.cpp1K context | 12.1one runPP 310.0 (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 | NVIDIA RTX 5090 | AMD Strix Halo 128GB |
|---|---|---|
| Qwen3.8 27B | 44.9 t/sQ6_K | 3.5 t/sFP16 |
| Qwen3.6 35B · 3B active | 163.0 t/sQ5_K_M | 12.6 t/sFP16 |
| Qwen3 30B · 3B active | 262.8 t/sQ4_K_S | 12.6 t/sFP16 |
| Gemma 4 31B | 46.2 t/sQ5_K_M | 3.1 t/sFP16 |
| Gemma 4 26B · 4B active | 114.5 t/sQ6_K | 9.5 t/sFP16 |
| Qwen3.5 35B · 3B active | 163.0 t/sQ5_K_M | 12.6 t/sFP16 |
| Qwen3.5 27B | 44.9 t/sQ6_K | 3.5 t/sFP16 |
| Qwen3.8 125B · 6B active | 20.3 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 27B | 44.9 t/sQ6_K | 3.5 t/sFP16 |
| DeepSeek V4.1 Flash 552B · 16B active | doesn't fit at these settings | doesn't fit at these settings |
| GLM-5.3 320B · 18B active | doesn't fit at these settings | 11.1 t/sQ2_K |
| Qwen3.5 9B | 54.8 t/sFP16 | 10.5 t/sFP16 |
Change the context, RAM or model in the calculator: NVIDIA RTX 5090 · AMD Strix Halo 128GB
Every report on each card, with its source: NVIDIA RTX 5090 · AMD Strix Halo 128GB