Where I think the DGX Spark shines over the Apple ecosystem for local AI is scalability.
You can start experimenting with one Spark (128 GB) and progressively add a 2nd, then a 4th if you want to run larger models. At 4 nodes you’d have 512 GB, enough for Nemotron Ultra or GLM 5.2 NVFP4.
That means your previous hardware investment scales. I really hope they continue with that philosophy and ensure the Spark 2 is backward-compatible with the Spark 1.
Reminds me a lot of the photography market, where companies like Nikon and Leica have maintained consistent standards and protected retro-compatibility across their hardware.
It’s linux with a GPU, so it’s unclear what you mean by compatible outside of what is already afforded by that.
If the spark 2 is a compelling upgrade in a few years (it may not be, I might be content with the 256 GB models I can run on my dual Spark 1’s by then), then I would probably imagine keeping the Spark 1s to run some specialized smaller model(s) and using the Spark 2s to run the best/biggest model I could.
Nothing would prevent you from running a heterogeneous cluster of spark 1s and spark 2s, but it’ll never be the optimal configuration for performance.
Yeah, I agree, I meant keeping the same design philosophy, and the care for making sure that they build on top of the previous version instead of against, creating synergies between generation 1 and 2.
Just giving one example, no idea if possible, but a prefill/decode split where you would split the two phases across hardware. Even if Spark 1 has lower memory bandwidth, prefill is compute-bound, so older nodes can perform prefill while the newer node handles decoding.
I just think the Spark is an amazing machine, and the fact that they put the ConnectX-7 at the heart of the product was pretty smart.
There is much to say, too, on optimizing for Nemotron on Sparks clusters to make sure the user has the best experience.
Compatible in what way? I don’t see why you wouldn’t be able to do PP with spark 1 and spark 2. You could do TP, but the performance is going to be bottlenecked by the slower spark 1 unless you do proportional sharding, which is a pain on top of all the other pains of this half-assed software stack.
If Spark 2 is not just a rebranded Spark 1 and is actually based on the next-gen Rubin chips, it feels like we can rule out any backward compatibility.
Even within the Blackwell lineup, we have SM100, SM120, and SM121, which aren’t exactly cross-compatible as it is. What’s going to happen when we transition to Rubin then? That’s a rhetorical question.
Personally, I don’t see any reason to expect any kind of compatibility at all… (I’d be happy to be proven wrong.)