Products You May Like
At this week’s GTC, Nvidia unveiled its much-anticipated “Ada Lovelace” RTX 40-series GPUs. The keynote included several headline-grabbing claims. One of those claims was an “up to 4X” uplift in RTX performance in some games compared with the RTX 3090 Ti. This is made possible by the company’s latest iteration of Deep Learning Super Sampling, aka DLSS. Now on its third iteration, DLSS 3 is only coming to the RTX 40-series GPUs. This wasn’t specifically mentioned in the keynote, to the best of our recollection, but has since been revealed via the company’s website. However, since previous Nvidia GPUs were able to run earlier versions of DLSS, why can’t they run DLSS 3? As it turns out, they probably can—just not very well.
The big innovation with DLSS 3 is it’s capable of predicting entire frames instead of just pixels. It’s able to do this thanks to advances in the GPU’s optical flow accelerator. This technology tracks the flow of pixels and has apparently been in Nvidia GPUs since Turing. However, in Ada Lovelace, it’s able to predict and construct entire frames. That’s possible because it can access pixel-level information not included in game-engine motion vector calculations, such as shadows, lighting, reflections, and particles. By using the game’s motion vectors and its optical flow analysis, it’s able to predict geometry and effects. This allows it to reconstruct 7/8ths of displayed pixels, according to Nvidia, which has obvious benefits to performance since they’re not being rendered with “brute force.”
Nvidia’s “benchmarks” showing next-generation DLSS performance. (Image: Nvidia)
However, none of this answers the question as to why that’s not possible with Ampere or Turing’s Tensor Cores. The answer comes from Bryan Catanzaro, who is a VP of Applied Deep Learning Research at Nvidia. He was asked on Twitter why it’s only possible on Ada, but not Ampere. His answer was pretty straightforward. He wrote, “DLSS3 relies on the optical flow accelerator, which has been significantly improved in Ada over Ampere—it’s both faster and higher quality.” This sounds like the Tensor Cores built into Ada are more powerful, and the flow accelerator is as well.
All that said, couldn’t it still boost frame rates on older GPUs? Catanzaro’s answer is pretty clear in that it would work, but not well. When asked why not just let customers try it anyway, he wrote, “Because then customers would feel that DLSS3 is laggy, has bad image quality, and doesn’t boost FPS.”
Though it might sound like a marketing move, it seems like there’s some truth to what he’s saying. For example, the optical flow accelerator (OFA) is apparently up to 2.5 times faster than that in Ampere, according to Tom’s Hardware. This likely means that even if you could run DLSS 3 on Ampere or Turing, it wouldn’t run well, or look good. This is reminiscent of when Nvidia launched RTX for owners of older GeForce GTX cards. It ran, sure, but it wasn’t fast.
Obviously, Nvidia is using the exclusivity of DLSS 3 as a reason to buy a 40-series GPU. However, that doesn’t mean DLSS 3 will forever be unobtainium for those with older GPUs. It is theoretically possible the company could continue to iterate on DLSS 3 to make it run faster on older hardware. It has zero incentive to that today, but a year from now things could change. For now, people who run one of the 35 games that will support DLSS 3 will have to run it on the older version—which is still quite impressive.
Now Read: