Hacker News new | past | comments | ask | show | jobs | submit login
Nvidia's next DGX supercomputer is all about generative AI (tomshardware.com)
49 points by ripvanwinkle 6 months ago | hide | past | favorite | 15 comments



After Nvidia is done selling crypto mining and AI generating hardware, maybe they can also focus a bit on their PC gaming side-hustle and throw us a bone in the form of a mid-range GPU that can run AAA games and doesn't cost both kidneys, that would be great.


Right now AI/datacenter is outbidding the gamers for constrained fab capacity. Until that changes, recent generation nodes won't come down in price and neither will gamer GPU's.


>Right now AI/datacenter is outbidding the gamers for constrained fab capacity.

Allow me to doubt this. Intel Arc A770 16GB has a TSMC manufactured die but it's much larger in size that an RTX 3060 12GB sells for way less, also at a (small) profit.


This is a weird comparison.

I don't think either of these are intended for the datacenter? So this would be a price comparison between two gaming / general purpose GPUs?

The A770 is going for $350ish, is manufactured on a higher tech node (TSMC N6), and has a die size of 406mm square.

The 3060 is going for $250ish, was manufactured on two lower tech nodes (originally Samsung 8nm then moved to TSMC 7nm), and has a die size of 392.5.

I'm not sure the description "much larger in size" makes sense here. If anything it seems like pricing follows node-area costs.


What about the Intel a770? : D


It's more like that AAA devs always target high-end GPUs.


That wasn't the main point though. The point was that a few generations ago, Nvidia was selling mid range silicon, 60 and 70 series cards, for around 300 to 400 USD price point. Now with the 4000 series, the 60 and 70 series silicone have the specs of the older 50 and 60 silicon, at much higher price points.

Nvidia is now selling us less silicone die for way more money, and their sales numbers reflect that.


From what I can tell, AAA devs typically target compatibility with consoles first, usually whatever is latest generation at the anticipated release date for their game. And these are equivalent to mid-range gaming rigs at best. Then they work on PC compatibility afterwards and poorly optimize performance of their assets against a common GPU compatibility layer that adjusts the features it enables (if they exist in this version of what's likely a shared underlying game engine) based on hardware compatibility.

If we're very lucky, the game will get tested against a wide range of hardware. Sane defaults for GPU feature flags based on detection of hardware capability and the results of that testing will result in a default user experience that doesn't need manual adjustment. If we're lucky.

So tl;dr, they're not targeting the top 1% of hardware. Typically they just screw things up and don't catch it. But higher end hardware will handle poor optimization better and provide a better experience. So it can feel like they're 'targeting' that hardware when really its a poorly optimized port.


That’s sort of the point of AAA games, though. There’s an entire universe of non-AAA games that will run just fine on extremely dated hardware.


AAA devs target the consoles level of performance, and its been like that for a while. A lot of "extreme" settings in games are usually just poorly optimized and sometimes even looks ugly or almost does nothing artistically or experience wise.


> Nvidia uses a new NVLink Switch System with 36 NVLink switches to tie together 256 GH200 Grace Hopper chips and 144 TB of shared memory into one cohesive unit that looks and acts like one massive GPU

Is there more information about this?


The direct announcement (posted at https://news.ycombinator.com/item?id=36113129) is probably more useful.


Archer: Do you want a singularity? Because this is how you get a singularity.


here is another interesting link explaining the reference architecture across 256 nodes and 24 racks:

https://www.anandtech.com/show/18877/nvidia-grace-hopper-has...


Just in time for elections.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: