Of course, Nvidia was going to have at least one supercomputer. That means we’ve been making data center GPUs for a million years. And we know that DLSS requires some pre-training to make its funky upscaling algorithm stable and effective. But this stupid Dave has a dedicated supercomputer “equipped with 1000 of our latest and greatest GPUs” that Nvidia has been running full-time for the past 6 years just to improve the quality of DLSS. I didn’t know I had it until last week.
At RTX Blackwell Editor’s Day, held last week in glorious Las Vegas* in the midst of CES 2025, Nvidia’s VP of Applied Deep Learning Research, Brian Catanzaro, took to the stage to talk about DLSS 4 and the many changes it brings. We talked about the challenges. that.
As well as the innovative switch from convolutional neural networks to DLSS 4’s new transformer models, another thing that caught my attention was Catanzaro’s aside about how Nvidia trains its models.
“Why have we been able to make progress in DLSS over the years?” he asks. “As you know, this has been a six-year continuous learning process for us.
“In fact, Nvidia has a big supercomputer with thousands of the latest and greatest GPUs that work 24/7 to improve DLSS. And we’ve been doing it for six years.”
Maybe I’m completely naive, but I didn’t realize how much resources are being spent on improving upscaling solutions over time.
I thought Nvidia might give its DLSS group time to use its multi-million dollar machines for training purposes, but now it’s clear that Nvidia is living rent-free inside its supercomputers. I didn’t know that.
A supercomputer that specializes in thoroughly removing images that aren’t good enough.
“What we’re doing in this process is we’re analyzing failures. When a DLSS model fails, it looks like ghosting, flickering, blurring. And, you know, we’ve developed We found failures in many of the games,” continued Catanzaro. I try to understand what’s going on, but why does the model make wrong choices about how to draw the image there?
“Then we find ways to enhance the training data set, which is growing all the time. We are collecting examples about.
“We put them into a training set, retrain the model, and test it on hundreds of games to find ways to make DLSS better. Here’s the process.”
And it’s a tough process. And my experience with the new transformer model in DLSS 4 is that it really only gets better. It’s not just the new RTX 50 series guys.
* To be honest, Las Vegas obviously has a bad reputation, especially for a die-hard introvert like me, but they won’t let you leave the airport unless you say so. They keep a close eye on it.