Nvidia announces DGX desktop "personal AI supercomputers"

arstechnica.com

20 points·freeqaz··6 comments

It'd be nice to know how this compares to what we can run locally on today's "AI" optimized machines from Apple, Microsoft, etc.

These chips are a generation and a half ahead of what is shipping now from other vendors (including the barely shipping AMD Ryzen 9 AI Max+ H395). They will sell every one they can produce to money that otherwise would have been spent on legacy hardware like M3 Ultra desktops and M4Max laptops. NVDA reportedly gave a forecast of $1T in revenue from all of its products by 2027.

That said, these chips are not quite shipping and their stability has yet to be proven. nVidia is also unlikely to DISRUPT its AI Cloud customers too much.

At this point it's completely uninteresting for the price point.

It only has 276 GB/s memory bandwidth.

A framework desktop with a strix halo at 2/3rd the price and similar memory bandwidth looks pretty good in comparison.

Exactly, and it's a regular x86 machine without any potential compatibility / OS support issues.

The dgx spark definitely makes sense for some people, but $3k for an upgraded Jetson, with less tops than a last gen consumer 4090, does beg the question. Unless you're getting them to practice specifically on nvidia hardware before scaling up, why not get an m4 max Mac w/ 128GB+ ram instead?

Reply on news.ycombinator.com