How do the DGX Spark cores compare to that of Apple macbook M4 cores. Curious how the community is thinking about it from your experiences.
Last I checked the NVIDIA DGX Spark’s computational capacity was 26x higher than Apple’s M4 Max and M3 Ultra
DGX Spark is 1,000 TOPS
M4 Max is 38 TOPS
M3 Ultra is in the same ballpark
DGX Spark blows Apple’s M4/M3 Ultra out of the water from an AI compute capability
Apple edges the DGX Spark’s 128GB and the Clusters 256GB in unified 512GB memory
Depending on how you plan to use the DGX Spark for AI/ML the DGX Spark is the no brainer. Apple may be better in memory intensive creative tasks.
I looked at both and since I’m looking specifically for AI/ML focused metal having the desktop AI Supercomputer is my preferred choice, if it launches this year.
4 Likes
Also note that the 1,000 TOPS is for FP4 (I believe), whereas on the Mac we are probably talking about higher precision? I am looking to make good use of both systems and they are both relatively energy efficient they will serve as nice always-available inference units for my single user local network.
3 Likes
I’ve been thinking about investing in both as well but I’m also tempted to invest in scaling Spark Clusters if they perform and deliver as expected and am on the fence till Q1 next year.
The Apple Mac Studio customization can be pushed to include:
- M3 Ultra chip
- 32-core CPU
- 80‑core GPU
- 32-core Neural Engine
- 512GB unified memory
- 16TB SSD storage
… bearing in mind that’s also a $15,500+ investment just for that Mac Studio (without accessories).
1 Like
I hear there is work going on for an M5 Ultra for Studio (not sure when, the info was a leak I think?). If you can wait on the Mac, then maybe that will become available? I pulled the trigger on the maxed M3U and it works quite well and is ever so silent. The Ultra config (vs the Max config) of Studio uses a large copper heatsink instead of maybe aluminum for the Max? All I know for sure is that it works well.
2 Likes
Thanks for the reminder!
Looks like Apple Mac M5 Ultra Studio targets ML/AI Development.
Rumored specs:
- 36-core CPU ← (28P+8E)
- 84‑core GPU ← up 4
- 32-core Neural Engine
- 512GB unified memory
- 16TB SSD storage
- Speculated launch Mar 2026
2 Likes
Looks like they’re both around the same price. I think the spark would be much better for AI/ML compute intensive tasks. Plus if you’re on the list like me, the spark should be available quite soon.
1 Like
Depends on the usage. DGX Spark should be a good fit for large model fine-tuning as well as AI development on the DGX platform, but not the best for LLM inferences due to the bottleneck of memory bandwidth. The speed of LLM inferences is mainly limited by the memory bandwidth. M4 Max has 526 GB/s bandwidth while DGX Spark has only 273 GB/s.
1 Like
Thanks @marcwester it was exciting to see that email come through last week and hopefully we’ll have the DGX Spark soon.
Similar to @lionellee & @erikshop I was thinking about complementing the DGX Spark with the Mac as for inference but I also want to scale keeping infra physical connectivity in mind and by then I’m hoping Jensen will have created a mini-powerful switch without having to spend upwards of $35,000 on a relevant switch.
2 Likes
I have an Apple MAC M3 Ultra with 512GB of memory. I will do a benchmark between it and the DGX Spark
3 Likes
Does the inception discount apply to the Spark?
based on the performance numbers rolling in for the DGX Spark I am going to wait for the 20 Petaflop version of the DGX workstation.
1 Like
Which one did you see?
The values of the original llama.cpp looks much better than those from Ollama that popped up first. At least for the big MoEs. So it might be worth to wait while optimized kernels for FP4 are available to be used vLLM, SGLang.
If not seen yet, the best review on YouTube is the one by Level1Techs who do see the full potential of that box.
Do we know the cost benchmark for the GB-300 DGX Station? I saw Reddit had it around $150,000