Even during the announcement of the first GeForce RTX 20-series video cards, many considered that Turing GPUs owe their not at all small dimensions to the presence of additional blocks: RT cores and tensor cores. Now, a Reddit user has analyzed infrared images of Turing TU106 and TU116 GPUs and concluded that the new compute units do not take up as much space as originally thought.
First, let's recall that the Turing TU106 GPU is the youngest and most compact NVIDIA chip with special RT cores for ray tracing and tensor cores for accelerating artificial intelligence functions. In turn, the Turing TU116 GPU, which is related to it, is deprived of these special computing units, and that is why it was decided to compare them.
NVIDIA Turing GPUs are divided into TPC blocks, which include a pair of streaming multiprocessors (Streaming Multiprocessors), which already include all the computing cores. And as it turns out, the Turing TU106 GPU has only 1,95mm² more TPC block area than the Turing TU116 GPU, which is 22%. Of this area, 1,25 mm² are tensor cores, and only 0,7 mm² are RT cores.
It turns out that without the new tensor and RT cores, the flagship Turing TU102 GPU, which underlies the GeForce RTX 2080 Ti, would not occupy 754 mm², but 684 mm² (36 TPC). In turn, the Turing TU104, which is the basis of the GeForce RTX 2080, could occupy 498 mm² instead of 545 mm² (24 TPC). As you can see, even without tensor and RT cores, older Turing GPUs would be quite large chips. Significantly more Pascal GPUs.
So what is the reason for such a considerable size? For starters, Turing GPUs have had their cache size increased. The size of shaders has also been increased, and Turing chips have larger instruction sets and larger registers. All this made it possible to significantly increase not only the area, but also the performance of Turing GPUs. For example, the same TU2060-based GeForce RTX 106 provides almost the same level of performance as the GP1080-based GeForce GTX 104. The latter, by the way, has a 25% larger number of CUDA cores, although it occupies an area of 314 mm2 versus 410 mm2 for the new TU106.
Source: 3dnews.ru