Recently, the company Nvidia has presented the long-awaited graphics card the new generation. Their parameters I described in the initial news, but at that time Nvidia itself said not everything.
NDA article about the architecture of Turing will be lifted on Friday, but for now, those interested in the segment of graphics cards a little deeper, I have some interesting data. Almost all of them are present here in this table.
Source — Videocardz
To begin with, I note that the three presented maps are based on three different GPU, which is unusual. For example, the GTX 1080 GTX 1070 use the same graphics processor — GP104. Immediately we have RTX 2070 based on GPU “lower class” — TU106. However, it is believed that Nvidia just moved the hierarchy, so now TU102 as GP100, TU104 — how GP102, and TU106 — how GP104. However, I have a question which GPU will get RTX 2060 (or is it GTX 2060), but not yet about this.
The most interesting thing for me personally is the number of cores RT. Those nuclei which are responsible for ray tracing in real time, but it is unclear exactly how they work. As you can see, their number is different for different GPUs.
If we are really for geeks in this direction, will paint more. Each GPU contains multiple clusters GPC (Graphics Processing Cluster), which can be seen in the table. Each cluster contains several blocks SM (Streaming Multiprocessor). Interestingly, before the number for all GPUs of the same generation had the same (if I remember correctly), but now — no. Have TU104 each GPC cluster contains 8 units of SM, and the TU102 and TU106 — 12 blocks. So on. Each block SM already contains CUDA kernel directly, and a new generation of tensor kernel and the RT kernel. More specifically, each block includes eight tensor SM cores and only one core RT. What does it mean?
See, RTX 2070 total of 36 cores RT, which is twice less than the RTX 2080 Ti. Model RTX/GTX 2060, apparently, will receive about 1536 CUDA cores. This means that nuclei RT she is only 24. Accordingly, the RTX/GTX 2050 them and will do about 16. And then the question arises, and whether use of such a large number of cores RT? In General, my point is that, apparently, the recent speculation that Nvidia card for hardware accelerated ray tracing will not receive are true.