32 comments

[ 4.0 ms ] story [ 59.2 ms ] thread
Article doesn't seem to mention price which is $4,000 which makes it comparable to a 5090 but with 128GB of unified LPDDR5x vs the 5090's 32GB DDR7.
I wonder why they didn't test against the broadly available Strix Halo with 128GB of 256 GB/s memory bandwidth, 16 core full-fat Zen5 with AVX512 at $2k... it is a mystery...
The strix halo can also be used as a capable gaming/dev Pac with your OS of choice.
You guys that continue to compare DGX Spark to the Mac Studios, please remember two things:

1. Virtually every model that you'd run was developed on Nvidia gear and will run on Spark. 2. Spark has fast-as-hell interconnects. The sort of interconnects that one would want to use in an actual AI DC, so you can use more than one Spark at the same time, and RDMA, and actually start to figure out how things work the way they do and why. You can do a lot with 200 Gb of interconnect.

At best this is a cheap setup to test distributed training/inference code.
Also remember that the Mx Ultras have 2-3x the memory bandwidth. Looking at the benchmarks even Strix Halo seems to beat the Spark. Buying a 200 Gbps switch is $10k-$100k+ so don't imagine anyone actually will use the interconnect. The logical thing for Nvidia would be to sell a kit with three machines and cabling, and make it a ring with the dual ports per machine. Helps for some scenarios but not others with the 10 times slower network than memory bandwidth.
> ollama gpt-oss 120b mxfp4 1 94.67 11.66

This is insanely slow given its 200+GB/s memory bandwidth. As a comparison, I've tested GPT OSS 120B on Strix Halo and it obtains 420tps prefill and >40tps decode.

Probably the quants have higher perplexity, but the Sparks performance seems to be lack lustre. The reviewer videos I've seen so far tries their best not to offend Nvidia or, rather, not break their contracts.
Looks like MLX is not a supported backend in Ollama so the numbers for the Mac could be significantly higher in some cases.

It would be interesting to swap out Ollama for LM Studio and use their built-in MLX support and see the difference.

How representative is this platform of the bigger GB200 and GB300 chips?

Could I write code that runs on Spark and effortlessly run it on a big GB300 system with no code changes?

All three (GB10, GB200 and GB300) are part of the Blackwell family, which means they have Compute Capability >= 10.X. You could potentially develop kernels to optimize MoE inference (given the large available unified memory, 128Gb, it makes the most sense to me) with CUDA >= 12.9 then ship the fatbins to the "big boys". As many people have pointed out across the thread, the spark doesn't really has the best perf/$, it's rather a small portable platform for experimentation and development
M5 Macs may be launching as early as today. Inference should see a significant boost w/ matmul acceleration.
It isn't that good for local LLM inferencing. It's not designed to be as such.

It's designed to be a local dev machine for Nvidia server products. It has the same software and hardware stack as enterprise Nvidia hardware. That's what it is designed for.

Wait for M5 series Macs for good value local inferencing. I think the M5 Pro/Max are going to be very good values.

I wish I could run Linux on them (the m5)
What is the value proposition for buying one of these vs renting time on similar hardware from a cloud provider?
I don't think there is one. Honestly this version 1 is dead on arrival.
Given that most of Nvidia's enterprise software products are all single server designed to run on DGX boxes, like NIMs, this makes sense.

I am still amazed at how many companies buy a ton of DGX boxes and then are surprised that Nvidia does not have any Kubernetes native platform for training and inferencing across all the DGX machines. The Run.ai acquisition did not change anything, as you leave all the work to the user to integrate with distributed training frameworks like Ray or scalable inference platforms, like KServe/vLLM.

Fascinating that we didn't have to wait too long. Apple announced M5 this morning. Does it compare though?
Nvidia always short changes its own products and stunts them in some way.

No doubt that’s present here too somehow.

Gotta cut off something important so you’ll spend more on the next more expensive product.

I think my 2001 MBP M1 Pro is ~200GB/s memory bandwidth, but it handles qwen3:32b quite nicely, albeit maxed out at ~70W.

I somehow expected the Spark to be the 'God in a Box' moment for local AI, but it feels like they went for trying to sell multiple units instead.

I'd be more tempted by a 2nd hand 128GB M2 ultra at ~800GB/s but the prices here are still high, and I'm not sure the Spark is going to convince people to part with those, unless we see some M5 glutenous RAM boxes soon. An easy way for Apple to catch up again.

That memory bandwidth choked out their performance. How can you claim 1000 tflops if it's not capable of delivering it. Seems they chose to sandbag the spark in favour of the rtx pro 6000.

I guess my next one I'm looking out for is the Orange Pi AI studio pro. Should have 192gb of ram, so able to run qwen3 235b, even though it's ddr4, it's nearly double the bandwidth of the spark.

Good luck with any kind of coherent ecosystem and support. Also, if you're in the U.S., there is a good chance you'll get hit with tariffs which would wipe out any potential value. I'd much rather stick with nVidia that has an ecosystem (even Apple for that matter), than touch a system like this off of Alibaba.
>Good luck with any kind of coherent ecosystem and support.

Admittedly I'm not a huge fan of debian; likely would end up going Arch on this one.

>Also, if you're in the U.S.,

Im not.

> I'd much rather stick with nVidia that has an ecosystem (even Apple for that matter), than touch a system like this off of Alibaba.

I get that. Realistically I'm waiting for medusa halo, some affordable datacenter card, something.

Two questions:

a) what is the noise level? In that small box, it should be immense?

b) how many frames do we get in Q3A at max. resolution and will it be able to run Crysis? ;-) LOL (SCNR)

GPT 120B is your goto model:

DGX Spark

pp - 1723.07/s

tg - 38.55/s

Ryzen AI Max+ 395

pp - 711.67/s

tg - 40.25/s

Is it worth the money?

"Metal foam" sounds cool but it just looks like a steel wool pad you would use for cleaning dishes.
Any views on how Isaac Lab / Isaac Sim would perform on DGX Spark?
Bruh, if it were priced at like $2,499 it would make sense, but this is just too much.