Hacker News new | ask | show | jobs
by GodelNumbering 5 days ago
I do not see how it is a 'beast of a' anything. It has 300GB/s memory bandwidth, barely above AMD Strix halo (256GB/s) with the same 128GB RAM and less than half memory bandwidth of M5 Max 128GB (614GB/s). Emphasizing memory bandwidth because most people interested in it I suppose are AI enthusiasts. Also, Windows.
2 comments
bigyabai 5 days ago
Unlike the M5 Max, it should have usable context prefill. It's feasible to run 256k token workflows that would take the better half of an hour for TTFT on the M5.
link
moffkalast 5 days ago
They have a lot of software groundwork ahead of them to make an ARM CPU viable for any kind of desktop use outside direct inference or training usage too.

AMD has the advantage that their x86 machines run everything, Apple maintains the whole MacOS stack, while Nvidia can barely scrape together one Ubuntu release per Jetson generation, it's beyond embarrassing. Maybe they ought to put those agents they keep droning about to some actual work on their OS support.

link
bigyabai 5 days ago
> Nvidia can barely scrape together one Ubuntu release per Jetson generation

Why would they do more? It's an LTS distro, the Nvidia drivers are updated for as long as the hardware's compute capability is supported.

Nvidia's ARM drivers are updated constantly, and battle-tested as the backbone in hundreds of thousands of Grace ARM servers.

link
moffkalast 4 days ago
Yes it's an LTS, which means it only makes sense to build something on it early in its release cycle, so the platform itself needs to keep up with new releases. Orins will go EoL next year, so good luck with that.

That's not even considering the lazy out tree patchwork support Nvidia does for their products on top of that. Maybe it's different in this case for Windows since it forces a rolling release, but I seriously doubt they'll do it properly instead of forking some version and keeping it around for 10 years like absolute idiots.

link
bigyabai 4 days ago
> That's not even considering the lazy out tree patchwork support Nvidia does for their products

For their ARM SOCs? Almost every single ARM OEM on the consumer market is begging you to use out-of-tree blobs for basic firmware support. Nvidia's stance isn't ideal but it's also not unique (or damning) to the rest of their ARM competitors.

link