Nvidia’s Director of Technical Marketing, Tom Peterson, joined HotHardware on its 2.5 Geeks podcast yesterday and revealed details about the upcoming RTX 20-series graphics cards that Nvidia had hitherto kept secret.
He talked about the Turing architecture’s Tensor Cores and RT Cores, he explained more about the Deep Learning Super Sampling technology, and perhaps most importantly, he admitted that Nvidia “could have done a better job” highlighting the performance of the new GPUs. He also talked somewhat candidly about what kind of performance gains that we could expect with the GTX 2080 cards.
Nvidia spent most of its press event discussing ray tracing technology and the performance gains that the Turing-based cards bring to the table for that task. The company spent no time highlighting how well the new cards would handle existing games, though it followed up a few days later with a performance chart that suggests high generation-to-generation performance gains.
During the Q&A with HotHardware, Peterson said that Turing is a “beast” and that “if you are on high-resolution and not CPU limited” you could expect to see between 35 percent and 45 percent performance gains in existing games when you step up from a GTX 1080 to an RTX 2080. And when queried about the RTX 2080 outperforming the GTX 1080 Ti, he said that he thinks there would be cases that would happen but couldn’t say for sure.
On Deep Learning Super Sampling
Deep Learning Super Sampling (DLSS) is a new technology that Nvidia introduced with the Turing architecture. The company didn’t explain the inner workings of the process, but the basic idea is that DLSS uses an AI algorithm to increase the resolution of your video output to improve the image quality. Nvidia said this technology improves performance compared to other super sampling and antialiasing technologies.
HotHardware asked Petersen what DLSS does and how its delivered to gamers. He didn’t go deep into the details, but Petersen explained that there are two methods for developers to implement DLSS. They could create a custom algorithm which runs natively on the Tensor Cores, or Nvidia can pump the game code through a Saturn V supercomputer to allow a neural net to create the algorithm for the developer. For that reason, DLSS will be largely used by AAA studios that have relationships with Nvidia, though Petersen said the company is working on “a lot of games and not everything is AAA titles.”
Nvidia Turing is the first GPU architecture to embrace the company’s NVLink bridge technology, which offers higher bandwidth data transfer than the traditional SLI bridge. At Siggraph, where Nvidia launched Turing and the Quadro RTX lineup, the company revealed that NVLink enables frame buffer scaling. In traditional SLI, each card needs the same data in its memory pool. With Quadro RTX cards, NVLink combines the memory of each card to create a single, larger memory pool.
Petersen explained that this would not be the case for GeForce RTX cards. The NVLink interface would allow such a use case, but developers would need to build their software around that function. “While it's true this is a memory to memory link; I don't think of it as magically doubling the frame buffer. It's more nuanced than that today,” said Petersen. “It's going to take time for people to understand how people think of mGPU setup and maybe they will look at new techniques. NVLink is laying a foundation for future mGPU setup.”
NVLink’s current implementation offers higher bandwidth, which should improve performance, but it’s not a fix-all solution. Anyone who’s used an SLI setup would be familiar with the dreaded micro-stutter. Unfortunately, it appears the new bridge doesn’t completely solve that problem. Not yet, at least.
Petersen said that the stuttering issue has more to do with alternate frame rendering (AFR), which is the current rendering technique for SLI. Nvidia is working on new mGPU technologies that would improve scaling in the future, but the company isn’t at liberty to discuss those yet.
HotHardware also asked Peterson about overclocking performance, to which he said, “it’s really good!” Nvidia built the power and cooling solutions of the Founder’s Edition cards specifically for overclocking, and Petersen said he'd seen 2.1Ghz on multiple cards, which suggests that at least the FE cards would be strong overclockers.
Petersen also mentioned Nvidia Scanner and GPU Boost 4.0, of which we know nothing about yet.