After weeks of leaks, Nvidia is finally formally launching its next-generation GPU family. As expected, the new cards are strongly focused on ray tracing, which the company claims the new GPUs are stunningly positioned to deliver.
Much of the company’s demonstrations and product capability discussion revolve around the use of real-time ray tracing in consumer graphics for the very first time. The GTX 1080 Ti, for example, is capable of calculating 1.21 gigarays per second, while Turing is capable of 10 gigarays at the same time. At one point, Nvidia claimed that a single Turing-class GPU will deliver higher ray tracing performance than 4 DGX Voltas. When tasked with the same workload, Pascal GPUs take 308ms to run, a quad of Volta finishes in 55ms, and one Turing-class GPU can finish in just 45ms. That’s still not a playable frame rate, but it’s being used to render a photo-realistic Star Wars scene as opposed to a shipping title.
The full-fat version of Turing (it’s not clear which GPU this specifically refers to) is capable of 14 TFLOPS of FP32, 110 FP16 tensor FLOPS (that’s the half-precision mode) and 78T RTX-OPS. That last metric isn’t really a metric at all since we don’t really know what an RTX-OP is, exactly, but presumably, that kind of information will be fleshed out at a later date. The current Titan X, in contrast, is capable of 12 RTX-OPS.
Nvidia has worked on developing Turing and its ray tracing capabilities for a decade, which would put the date on this project as kicking off not long before Intel started publicly talking about Larrabee, its own aborted push to move to a real-time ray tracing rendering system. And that brings us to the GPUs themselves:
The gap between the 2080 and the 2080 Ti is enormous. The 2080 Ti is 4352 cores, up 1.47x from the 2080 (the gap between the 1080 and 1080 Ti was 1.4x). Clock speeds are again slightly lower, (1350 base clock on 2080 Ti, down from 1480 on the 1080 Ti) with 11GB of RAM and 616GB/s of memory bandwidth.