Of course Nvidia was going to have not less than one supercomputer; I imply it has been making information centre GPUs for 1,000,000 years now. And I do know that DLSS requires a specific amount of pre-training to have the ability to make its funky upscaling algorithms as steady and efficient as they’re. But this dumb-ass Dave did not realise till final week that Nvidia has a devoted supercomputer, “with 1000s of our newest and best GPUs” that has been operating full-time for the previous six years simply to enhance the standard of DLSS.
At final week’s RTX Blackwell Editor’s Day in wonderful Las Vegas*, deep within the midst of CES 2025, Brian Catanzaro, Nvidia’s VP of utilized deep studying analysis, took to the stage to speak by DLSS 4 and the various adjustments and challenges it brings with it.
As nicely because the game-changing change from convolutional neural networks to the brand new transformer mannequin for DLSS 4, the opposite factor to catch my ear was the apart Catanzaro made about how Nvidia trains its fashions.
“How is it that we have been capable of make progress with DLSS through the years?” He asks. “You know, it has been a six 12 months, steady studying course of for us.
“Actually, we’ve got an enormous supercomputer at Nvidia, with many 1000s of our newest and best GPUs, that’s operating 24/7, three hundred and sixty five days a 12 months enhancing DLSS. And it has been doing that for six years.”
Maybe it is me being totally naïve, however I did not realise the quantity of useful resource that was being devoted to creating its upscaling resolution higher over time.
I figured Nvidia may give its DLSS gang some devoted time with a multi-million greenback machine for coaching functions from time to time, however I did not realise it was dwelling rent-free contained in the thoughts of its personal supercomputer.
A supercomputer purely devoted to grinding away on pictures that simply aren’t fairly adequate.
“What we’re doing throughout this course of,” Catanzaro continues, “is we’re analysing failures. When the DLSS mannequin fails it appears like ghosting or flickering or blurriness. And, you realize, we discover failures in lots of the video games we’re taking a look at and we strive to determine what is going on on, why does the mannequin make the fallacious selection about how to attract the picture there?
“We then discover methods to reinforce our coaching information set. Our coaching information units are all the time rising. We’re compiling examples of what good graphics appears like and what tough issues DLSS wants to resolve.
“We put these in our coaching set, after which we retrain the mannequin, after which we take a look at throughout tons of of video games with a view to work out how you can make DLSS higher. So, that is the method.”
And it is a hell of a course of. And from my expertise of the brand new transformer mannequin inside DLSS 4 it actually is barely getting higher. Not only for new RTX 50-series people, both.
* Las Vegas, in all honesty, is decidedly unglorious particularly for an inveterate introvert reminiscent of myself, however you have to say that or they will not allow you to out of the airport. Tight monitoring they have over there.
