Microsoft has announced the next iteration of its DirectX API, dubbed DirectX Raytracing or DXR. Coinciding with the announcement, Nvidia is talking about its support for DXR, plus its own GameWorks Ray Tracing / RTX that offers raytracing features built on DXR's functionality. (RTX is apparently the acronym for Ray Tracing for Gameworks, as in 'gameworX.') But perhaps most interesting in all of this are the hints of how Volta hardware will help accelerate DXR and RTX.
I gave the overview of raytracing and DXR already, and Nvidia has their own overview of rasterization and raytracing, but then Nvidia adds a few additional nuggets. First, RTX is the culmination of 10 years of algorithms and GPU architecture enhancements. Then we get this bit: "Highly optimized ray tracing technology on Volta GPUs." This is interesting, because if you've been following the GPU rumor mill, there has been plenty of talk of late about the future of Nvidia GeForce cards. And of course, there's the above video from Remedy Entertainment showing RTX in action.
There are two codenames floating around, Ampere and Turing, which has led to speculation that Volta will end up relegated to the supercomputing and machine learning markets. Right now, the only Volta products that exist are the Tesla V100 and the Titan V, neither of which are practical for gaming use—Tesla V100 doesn't even have display outputs. The only Volta GPU so far is the GV100, and beyond the normal FP32 CUDA cores (which is what we use in most games), GV100 includes FP64 cores and Tensor cores, plus the ability to do FP16 on the CUDA cores at twice the rate of FP32.
It's those Tensor cores that have been a big question mark for gaming, as Tensor operations are primarily used in machine learning. I speculated earlier that Turing could be a product focused solely on machine learning, by putting a bunch of Tensor cores and nothing else (relatively speaking) into the chip. Ampere meanwhile has been bandied about as the next GeForce target, without Tensor cores. Now that we've heard about DirectX Raytracing and Nvidia's talk of Volta acceleration for DXR and RTX, however, I'm beginning to think Ampere is simply the next part after Volta, and it could be a year or more away from launch.
Nvidia goes on to talk about RTX specifically offering libraries for area shadows, glossy reflections, and ambient occlusion. While DXR will run on Nvidia's existing hardware, RTX will require a Volta (or later) GPU. Again, that makes the likelihood of Volta being limited to Titan V and Tesla extremely unlikely. But what's the actual hardware in Volta that enables RTX?
Nvidia didn't get into the details of the hardware yet—that's 'reserved for a later discussion'—but there was a brief mention at one point on how machine learning tools are being used to optimize the creation of accurate real-time ray tracing. It might be a bit of a tenuous link, but if I'm reading between the lines correctly, those Tensor cores in Volta hardware can be used to optimize performance for ray tracing. With nearly an order of magnitude more TFLOPS available in Tensor cores vs the traditional FP32 CUDA cores (using the GV100 ratio), Volta may be the first high performance ray tracing GPU to hit the consumer world.
