DGX Spark, Nvidia’s tiniest supercomputer, tackles large models at solid speeds HANDS ON Nvidia bills its long-anticipated DGX Spark as the "world's smallest AI supercomputer," and, at $3,000 to $4,000 (depending on config and OEM), you might be expecting the Arm-based mini-PC to outperform its less-expensive siblings. But the machine is far from the fastest GPU in Nvidia's lineup. It's not going to beat …
I reserved a DGX Spark and should be getting an invitation to buy soon, but... eh
I think I'll decline the invitation for a few reasons:
> M5 Max is right around the corner with double rate FP16 and "Neural Accelerators" (Tensor Cores). The advantage this thing had over an Apple Silicon machine is almost certainly null and void now. A Mac Studio with an M5 Max is far more useful to me than this very focused device.
> The Elon PR stunt is massively off-putting, yuck.
The machine looks really nice but the Ubuntu image is riced and gross, full of bizarre bloat but at the same time, lacking the docs/manpages. If its anything like the Jetson Nano, it just needs to be completely wiped and a clean (glibc based Linux) image prepared using the raw BSP (Board Support Packages).
Due to some weirdness with the bootloader, kernel modules and user-land tools, this will take time (took me about a week to get the Nano productive). And at the end, you have something clean, functional but completely non-standard from what Nvidia's support probably expects.
This post has been deleted by its author
This is a lot less exciting than it sounded on paper. It's a bit overpriced for that sort of performance...and for that reason, ah'm oot.
If it was half the price, even $2000 it would be an instant no-brainer. At $3,000, it's not even competitive...for $3,000 you could get 4x RTX 3090 ti cards and achieve better performance...you'd have less vram...you'd have 96GB instead of 128GB...but for certain types of model, I don't think that's much of a trade off...particularly for image diffusion. It's possible to get extremely high quality results out of a 30GB model...especially if you can achieve higher iterations per second...because there is a point where more iterations becomes more important than a bigger model for image diffusion for an individual user or a small team.
I might wait a year before I pick one up used, they will probably be dirt cheap in a years time.
