NVIDIA DGX Spark vs M3 Ultra 512GB for local LLMs
91 reports on the NVIDIA DGX Spark and 41 on the M3 Ultra 512GB
Which is faster for local LLMs?
On Qwen3.8 125B · 6B active at 4-bit, the NVIDIA DGX Spark has 29.8 generation tokens/s (2 runs); the M3 Ultra 512GB has 28.0 generation tokens/s (2 runs); prompt processing 558.8 tokens/s (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 3.0x higher on the M3 Ultra 512GB (273 GB/s on the NVIDIA DGX Spark, 819 on the M3 Ultra 512GB). Across 2 shared models with generation measurements on both cards, the NVIDIA DGX Spark's median is higher on 1 and the M3 Ultra 512GB's on 1.
- Memory
- 128GB unified
- Memory bandwidth
- 273 GB/s
- FP16 compute
- 125 TFLOPS
- Reports
- 91
- Memory
- 512GB unified
- Memory bandwidth
- 819 GB/s
- FP16 compute
- 28.4 TFLOPS
- Reports
- 41
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 DGX Spark | M3 Ultra 512GB |
|---|---|---|
| Qwen3.8 125B · 6B active4-bit | 29.82 runsllama.cpp, SGLang128K to 256K context | 28.02 runsPP 558.8 (one run)llama.cpp, oMLX250K context, 1 not stated |
| Gemma 4 26B · 4B active4-bit | 59.12 runsPP 2657.0 (one run)llama.cpp, vLLM256K context, 1 not stated | 87.8one runPP 2595.3 (one run)llama.cppcontext not stated |
| Qwen3.8 27B4-bit | -0 runsPP 899.9 (one run)llama.cpp16K context | 31.8one runPP 391.2 (one run)MLXcontext 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 DGX Spark | M3 Ultra 512GB |
|---|---|---|
| Qwen3.8 125B · 6B active | 14.4 t/sQ6_K | 13.6 t/sFP16 |
| Gemma 4 26B · 4B active | 10.8 t/sFP16 | 21.5 t/sFP16 |
| Qwen3.8 27B | 4.0 t/sFP16 | 8.0 t/sFP16 |
| Qwen3.6 35B · 3B active | 14.5 t/sFP16 | 28.7 t/sFP16 |
| DeepSeek V4 Flash 284B · 13B active | 38.3 t/sQ2_K | 22.6 t/sQ8_0 |
| Qwen3.6 27B | 4.0 t/sFP16 | 8.0 t/sFP16 |
| DeepSeek V4.1 Flash 552B · 16B active | doesn't fit at these settings | 30.0 t/sQ4_K_M |
| GLM-5.3 320B · 18B active | 12.7 t/sQ2_K | 7.4 t/sQ8_0 |
| Qwen3.5 9B | 12.0 t/sFP16 | 23.9 t/sFP16 |
Change the context, RAM or model in the calculator: NVIDIA DGX Spark · M3 Ultra 512GB
Every report on each card, with its source: NVIDIA DGX Spark · M3 Ultra 512GB