Nvidia sells tiny new computer that puts big AI on your desktop (opens in new tab)

> AMD AI Max PCs with 128gb of unified memory? The spec sheet say "128 GB LPDDR5x", so how is this better?

Framework's AMD AI Max PCs also come with LPDDR5x-8000 memory: https://frame.work/desktop?tab=specs

The GPU is significantly faster and it has cuda, though I'm not sure where it'd fit in the market.

At the lower price points you have the AMD machines which are significantly cheaper, even though they're slower and with worse support. Then there's apple's with higher memory bandwidth and even the nvidia agx Thor is faster in GPU compute at the cost of worse CPU and networking, and at the 3-4K price point even a threadripper system becomes viable that can get significantly more memory

CUDA.

Given that it runs on ARM chips and is specifically designed for AI tasks, I would be more surprised to see it running Windows by default

I would use it for locally hosted RAG or whatever tech has supplanted it instead of paying API fees. We have ~20TB of documents that occasionally need to be scanned and chatted with and $4,000 one time (+ electricity) is chump change compared to the annual costs we would otherwise be looking at.

Though the adage “this is the worst it’ll ever be” is parroted daily by AI cultists, the fact is it’s still yet to be proven that currently available LLMs can be made cost effective. For now every ai company is lighting tens of billions of dollars on fire every year and hoping better algorithms, hardware, and user lock in will ensure profits eventually. If this doesn’t happen, they will design more and more “features” in the LLM to monetize it - shopping, ads, sponsored replies, who knows? It may get really awful. And these companies will have so much of our data and eventually the need to make profits will lead them to sell that data and just generally try to extract as much out of us as they can.

This is why in the long run I believe we all should aspire to do LLM inference locally. But unfortunately we just are not anywhere close to par with the SoTA cloud models available. Something like DGX spark would be a decent step in this direction, but this platform appears to mostly be for prototyping / training models meant to eventually be run on data center nvidia hardware.

Personally I think I will probably spec out an M5 max/ultra Mac Studio once that’s a thing, and start trying to do this more seriously. The tools are getting better every day and “this is the worst it’ll ever be” is much more applicable to locally run models.