Is the DGX Spark about as fast as a Mac Studio for image generation? Niklas's test and the newly announced 64GB option

Give a DGX Spark and a Mac Studio the same prompts, and which one hands it back first? Niklas (@niklaslenz_ai) answers with a stopwatch.

According to the post, Qwen-Image-2.1 takes 28.7 seconds per 1024² image on a DGX Spark, 29.9 seconds on a Mac Studio M5 Ultra, and 7.6 seconds on an RTX 5090 measured for reference ¹.

The setup and the ratings

Niklas writes that the test used the same four prompts with three seeds each, and that the enhancer was off for the speed measurement ².

There are quality ratings too. In Niklas's own ratings (a single rater, enhancer off and on averaged), the Mac scores 7.1 and the Spark 6.8 out of 10 ³. He also notes that enhancer time was not measured on the Mac, and that the Mac's enhancer-on images use texts written on the 5090 ⁴. Because the rater is the author alone, this is not a basis for saying the Mac's images are better in general.

Why the gap is not just the hardware

The machines did not run the same software. Niklas says the Mac ran mflux with 8-bit weights, while the Spark and the 5090 ran ComfyUI INT8 ⁵. With different software (mflux versus ComfyUI), the difference in time, or its smallness, cannot be read as a hardware difference alone ⁵.

Another point: the DGX Spark that was measured sits under the 128 GB heading on Niklas's own site ⁶. Read that way, it is a result from a 128GB machine, and it tells you nothing about image-generation speed on any other configuration ⁶.

NVIDIA's public playbook covers installing ComfyUI on a DGX Spark so the unified memory can be used for large models ⁷.

What a 64GB configuration changes

In its October 21 blog post, NVIDIA wrote that DGX Spark 64GB systems would start rolling out from partners on Friday, October 23 ⁸. That was a plan at the time of writing, not a statement that the systems are on sale or in stock.

According to the product page, memory comes in 64 GB or 128 GB configurations ⁹, and the 64 GB configuration is sold exclusively through participating OEM partners ¹⁰.

As a guide to model size, NVIDIA lists up to 100B parameters for the 64GB system ¹¹ and up to 200B parameters for the 128GB system ¹². The blog names Nemotron 3.5 Bolt 74B 24GB, Qwen3.8 27B and Gemma 4 26B-A4B as examples of models that 64GB systems can run ¹³.

NVIDIA also writes that a vLLM update makes agentic tasks with Qwen3.8 27B up to 1.3x faster on DGX Spark ¹⁴.

Adding more units

If a single machine is not enough, you can link them. NVIDIA's blog says clusters of up to eight units are supported in manual configuration ¹⁵, and that such a cluster delivers 1TB of unified memory, enough for models like GLM 5.3 ¹⁶. It does not say which memory configuration each unit has.

The product page body, on the other hand, says up to four DGX Spark systems can work together on larger models ¹⁷. The sources disagree on the count, so check NVIDIA's latest wording before buying around a particular number.

Points to weigh

• Mostly image generation: in Niklas's test, per-image times on the DGX Spark and the Mac Studio M5 Ultra came out close [[C261]]. The software differed [[C264]], and the Spark measured appears under the 128GB heading on Niklas's site [[F312]]. The ratings come from a single rater, with the Mac at 7.1 and the Spark at 6.8 [[C263]]
• Running a large LLM: NVIDIA's guide is up to 100B parameters for 64GB [[F305]] and up to 200B parameters for 128GB [[F306]]
• Aiming for even larger models: clustering is an option, though the sources differ on the maximum count [[F315]] [[F307]]
• Choosing 64GB means choosing a machine from a participating OEM partner [[F304]]

This article is about the DGX Spark and says nothing about the performance of RTX Spark machines. If you are weighing both, check the official support list for the tools you want to run first.

Image and LLM models are large, and they need somewhere to live.

DGX Spark-class machines can also be found on Amazon.

Read next

If memory size is your sticking point, see our read of the memory-size benchmarks for RTX Spark laptops. For more examples of local AI on these machines, see our roundup of local AI showcases.