Hacker News new | ask | show | jobs
by smoldesu 1350 days ago
Actual use cases? Ludicrously quick AI inferencing, media transcoding, 3D render offloading, SIMD-bound simulations and anything that uses CUDA. Dedicated graphics will dominate in these scenarios, and it's such a large market that the 4090 is hardly the tip of the iceberg. Nvidia's server-grade AI chips are genuine monsters. People definitely use these cards, and there's a demand to scale them for datacenters too.

Gaming still drives the sales, but you'd be surprised by what you can do with a $500 GPU plugged into a Linux box.
1 comments
kkielhofner 1350 days ago
Is there really datacenter demand for these anymore?

In addition to the Nvidia EULA against "datacenter use" they (allegedly) shut down the last dual slot 3090 (with server compatible cooling configuration) because it was eating at higher margin datacenter SKUs[0].

I don't see a lot of datacenters making use of a 3.5U card with challenging power and cooling requirements but I certainly could be wrong.

0 - https://www.crn.com/news/components-peripherals/gigabyte-axe...

link
my123 1350 days ago
The NVIDIA datacenter use clause is present in the license agreement for the Windows driver, but not on the Linux one.
link
kkielhofner 1350 days ago
Do you have a reference?

What’s the point of the infamous “No Datacenter Deployment. The SOFTWARE is not licensed for datacenter deployment, except that blockchain processing in a datacenter is permitted.” clause if it only applies to Windows?

link
my123 1349 days ago
Yes.

Windows: https://www.nvidia.com/en-us/drivers/geforce-license/

Linux: https://download.nvidia.com/XFree86/Linux-x86_64/520.56.06/l...

link
kkielhofner 1349 days ago
Wow, thanks!

This is fascinating and now I'm really confused. Who is using CUDA on Windows in any datacenter application?

link
smoldesu 1349 days ago
I wonder if it's the other way around... maybe this rule was grandfathered in from a time when Microsoft feared companies offering "DirectX as a Service" or whatever. It doesn't really seem like the sort of thing Nvidia would care about, especially considering the licensing of their Unix drivers and considerable enterprise support efforts.
link