DGX Spark vs Mac Studio vs Strix Halo for local LLMs: measured tok/s and tok/s per watt
Updated 21 Oct 2026 · every number links to its source build
These three small boxes are the usual shortlist for running big models at home: NVIDIA's DGX Spark (GB10, 128 GB), Apple's Mac Studio (M3/M5 Ultra, up to 512 GB) and AMD Strix Halo (Ryzen AI Max+ 395, 128 GB). Here is what they measured, and how many tokens each gets per watt.
Same benchmark, same method: speed and tok/s per watt
Only one public series tested all three with the same tool and logged power: Jeff Geerling's ai-benchmarks. Power is the peak whole-system draw during the run as logged there (Strix Halo: the "around 133 W" the tester reported for those runs). tok/s per watt = measured tok/s ÷ that power, so it is a conservative figure.
| Model | DGX Spark Dell Pro Max with GB10 · build | Mac Studio M3 Ultra, 512 GB · build | Strix Halo Framework Desktop, Ryzen AI Max+ 395 · build |
|---|---|---|---|
| Llama 3.2 3B | 97.93 tok/s 128.5 W → 0.762 tok/s per W | 154.6 tok/s 223 W → 0.693 tok/s per W | 88.14 tok/s 133 W → 0.663 tok/s per W |
| gpt-oss 20B | 87.97 tok/s 122.8 W → 0.716 tok/s per W | 115.29 tok/s 227 W → 0.508 tok/s per W | not measured |
| Llama 3.1 70B | 4.71 tok/s 156 W → 0.030 tok/s per W | 14.08 tok/s 243 W → 0.058 tok/s per W | 4.97 tok/s 133 W → 0.037 tok/s per W |
Generation tokens/sec (llama.cpp / Ollama eval rate). Strix Halo rows ran on the Radeon 8060S iGPU with llama.cpp Vulkan, Q4_K_M; quantization for the other two is not stated in the summary table. gpt-oss 20B was not run on the Framework board. Higher tok/s per W = less energy per token.
2026 models: measured tok/s (power not measured)
For newer models no source measured all three systems with the same method or logged power, so this table shows tok/s only, from different testers, engines and quants. Use it for orders of magnitude, not single-digit differences.
| Model | DGX Spark | Mac Studio | Strix Halo |
|---|---|---|---|
| Qwen3.8 27B | 11.0 vLLM, 128K ctx · source build 19.0 vLLM + MTP · source build | 19.69 M5 Ultra, BF16, MLX · source build 12.46 M3 Ultra, BF16, MLX · source build | 9.59 Q6_K, Vulkan · source build 18.01 Q6_K + MTP · source build 7.76 Q8_0, llama.cpp ROCm · source build |
| Qwen3.6 35B-A3B | 78.1 NVFP4, vLLM · source build 102.3 NVFP4 + MTP · source build | no measured result yet | 54.22 Q8_0, llama.cpp (Vulkan) · source build |
| gpt-oss 120B | 44.4 vLLM, 128K ctx · source build | 79.68 M2 Ultra 76-core, MXFP4, llama.cpp · source build | 55.8 Ryzen AI Max+ PRO 495 (192 GB), MXFP4, Vulkan · source build |
| gpt-oss 20B | 96.2 vLLM · source build 87.97 llama.cpp (Geerling) · source build | 115.29 M3 Ultra, llama.cpp (Geerling) · source build | no measured result yet |
Power for these runs: not measured at the wall. Apple reports peak power of about 300 W (M5 Ultra) and 200 W (M3 Ultra) at 65K-token context (source), and one DGX Spark test logged a GPU-only peak of 82 W in nvidia-smi (source); neither is comparable to the wall-power figures above, so we don't compute tok/s per watt from them. The gpt-oss 120B Mac result is an M2 Ultra (the chip used in the 2023 Mac Studio).
FAQ
Which is fastest for local LLMs: DGX Spark, Mac Studio or Strix Halo?
In the one test series that ran all three the same way (Jeff Geerling's ai-benchmarks), the Mac Studio M3 Ultra generated fastest: 154.6 tok/s on Llama 3.2 3B vs 97.93 (DGX Spark) and 88.14 (Strix Halo), and 14.08 tok/s on Llama 3.1 70B vs 4.71 and 4.97. Generation speed tracks memory bandwidth, and the M3 Ultra has the most.
Which uses the least power per token (tok/s per watt)?
It depends on model size. On small models the DGX Spark was most efficient: 0.762 tok/s per watt on Llama 3.2 3B and 0.716 on gpt-oss 20B (peak system power ~122.8-128.5 W). On the dense Llama 3.1 70B the Mac Studio was most efficient at 0.058 tok/s per watt, vs 0.037 (Strix Halo) and 0.030 (DGX Spark), because it was almost 3x faster while drawing well under twice the power.
Is the DGX Spark faster than Strix Halo?
Only slightly, and not always. On Llama 3.2 3B the DGX Spark was faster (97.93 vs 88.14 tok/s); on Llama 3.1 70B Strix Halo was marginally faster (4.97 vs 4.71). On gpt-oss 120B a DGX Spark measured 44.4 tok/s in vLLM at 128K context and a Ryzen AI Max+ PRO 495 box measured 55.8 in llama.cpp; those are different engines and sources, so treat them as the same ballpark.
What the numbers say
- Mac Studio: fastest generation and best efficiency on big dense models, thanks to the most memory bandwidth. Best when the model is large and you chat with one model at a time.
- DGX Spark: lowest power per token on small and MoE models, CUDA and vLLM, and speculative decoding (MTP) that lifts Qwen3.6 35B-A3B from 78.1 to 102.3 tok/s.
- Strix Halo: close to the DGX Spark on generation speed in the same-method test above, but slow on dense 27B+ models without MTP (9.59 tok/s on Qwen3.8 27B Q6_K). See the AMD guide for Vulkan vs ROCm.
More: all comparisons · Mac vs PC for local AI · will the model fit? · Qwen3.8 27B on every rig · gpt-oss 120B on every rig
Referral link: if you sign up with Google and add $10, Runpod gives you and RigBase bonus credit (terms). It never affects our data.