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

Would consider at a lower price of $500 USD, way too expensive for what it brings.

I think it's cheap at $3000. 128gb RAM, top of the line GPU capabilities, 4tb storage... it's much better than what a top shelf mbp can do and much cheaper

correct, I just spent $4k on an "AI" machine to do stuff. 96GB ram, ryzen 9 9950x 16 core, 4TB nvme, 24tb hdd, rtx 4090.

If this thing was available six months ago I would have bought it instead!

upto 4tb storage

What do you do on it?

This definitely loses on CPU performance.

I originally thought so, since the previous Grace CPUs used the neoverse N2, which loses to Apple's M4 cores.

However apparently 10 of the cores are the Cortex-X925 CPUs, which are a serious upgrade. Basically 10 performance cores and 10 efficiency cores that should be pretty competitive with any current apple CPU.

M4 Max is 12 performance cores and 4 efficiency cores, the former of which are basically the fastest single-core performance you can buy right now, and the latter of which have never been touched by any architecture for their energy class. It seems highly unlikely that what you say is the case.

$500? that won't even buy you a decent new graphics card, anybody claiming this is overpriced doesn't have a clue.

The cortex X925 is pretty competitive. Seems similarly aggressive to the Apple M4. Both have 10 wide decode and disbatch, pretty exceptional single thread, and good overall performance.

Seems close enough that it might well come down to if your application uses SVE (which the X925 has) or SME (which apple has). I believe generally SVE is much easier to use without using Apple proprietary libraries.

Of if you need significant memory bandwidth, apple M4 peaks at around 200GB/sec or so, the other 300GB/sec or so is available for the GPUs.

Seems quite plausible that 10 * x925 and 10 * A725 might well be more collective performance than apples 12 p-cores + 4 e-cores. But sure it's a bit early to tell and things like OS, kernel, compiler, thermal management, libraries, etc will impact actual real world performance.

Generally I'd expect the Nvidia Project Digit 10 p cores + 10 e cores + healthy memory system to be in the same ball park as the apple M4 max.

I guess you are talking about something different; I'm not really treating the cores as ML accelerators.