As planned, NVIDIA has taken advantage of the online celebration of the GTC 2020 to present Ampere, its new graphic architecture, and has focused solely on its professional side. The Tesla A100 has been the main protagonist of the event, a graphic solution that comes with 6,192 shaders (108 SM active), 432 tensor cores, a 5,120-bit bus and 40 GB of HBM2 memory with a bandwidth of 1,134 GB / s.
There is no doubt that the Tesla A100 is a true monster, but it is a “cropped” version, that is, does not use GA100 graphics core with all its active SM drives, which means that NVIDIA could surprise us with a new top of the range graphics card equipped with up to 8,192 shaders.
It was clear that NVIDIA comfortably leads the professional sector, and with the announcement of Ampere the company has only reinforced that leadership, but what about Ampere in its version for general consumption? As I said we have no news in the strict sense, but the CEO of NVIDIA, Jen-Hsun Huang, has confirmed that the RTX 30 series will use the same architecture, that is, that there will be no distinctions between Tesla, Quadro and RTX series graphics cards. 30 based on Ampere, at least as far as base configuration is concerned.
NVIDIA will maintain a common base and bet on specialization
That is the idea that the CEO of the company has transmitted, and we are not surprised since in the end it is the same that we have seen in the RTX 20 series. NVIDIA completely abandons the distinction between graphics cards in the professional and consumer sectors which he established with Pascal and Volta, and adopts the common approach that we have seen in Turing.
What has been said means that a 30 series RTX will use the same architecture as a Tesla or Quadro, And it may even have the same core-level base (GPU) configuration, but NVIDIA will tailor each design to the specific needs of the industry that each graphics card is targeting.
So, for example, a 30 series RTX will have high power at FP16 and FP32, but it will come with its severely undercut FP64 capabilities, and the same will apply to Quadro. In contrast, the Tesla will offer more power in FP64 and FP8-16. I remind you that FP64 loads are mainly used in scientific simulations, and that FP8 loads are mainly used in inference, artificial intelligence and deep learning.
I don’t want to end without making an important note, and I think we could see a RTX TITAN 30 series with 6,192 shaders and 24 GB + of GDDR6 memory. In the video that I leave at the end you can see an interesting analysis of the evolution of ray tracing and DLSS 2.0 technology, which as we know has turned out to be one of the most important advances that Turing has left us.