Nvidia's Project Digits is a 'personal AI supercomputer'

If they end up actually shipping this, lots of people will buy these machines to get an AArch64 Linux workstation—even if they are not interested in AI or Nvidia GPUs.

At $3,000, it will be considerably cheaper than alternatives available today (except for SoC boards with extremely poor performance, obviously). I also expect that Nvidia will use its existing distribution channels for this, giving consumers a shot at buying the hardware (without first creating a company and losing consumer protections along the way).

> At $3,000, it will be considerably cheaper than alternatives available today

$3000 gets me a 64-core Altra Q64-22 from a major-enough SI today: https://system76.com/desktops/thelio-astra-a1-n1/configure

And of course if you don't care about the SI part, then you can just buy that motherboard & CPU directly for $1400 https://www.newegg.com/asrock-rack-altrad8ud-1l2t-q64-22-amp... with the 128-core variant being $2400 https://www.newegg.com/asrock-rack-altrad8ud-1l2t-q64-22-amp...

That Altra may be a good choice for certain server applications, like a Web server, but when used as a workstation it will be sluggish, because it uses weak cores, with much lower single-threaded performance than the Arm cores used in NVIDIA Digits.

For certain applications, e.g. for those with many array operations, the 20 cores of Digits might match 40 cores of Altra at equal clock frequency, but the cores of Digits are likely to also have a higher clock frequency, so for some applications the 20 Arm cores of Digits may provide a higher throughput than 64 Altra cores, while also having a much higher single-thread performance, perhaps about double.

So at equal price, NVIDIA Digits is certainly preferable as a workstation instead of a 64-core Altra. As a server, the latter should be better.

I mean I can get a Snapdragon X Elite laptop for $1200 that'll have a faster CPU than the one in the Digits, too...

Sure, but less than half the memory bandwidth and inference is largely bandwidth bound.

The context was people that don't care about AI or Nvidia GPUs and just want an AArch64 system. So inference performance is irrelevant here

> lots of people will buy these machines to get an AArch64 Linux workstation—even if they are not interested in AI or Nvidia GPUs.