Presented 3D-accelerator Nvidia Quadro RTX, based on the architecture of Turing

Last night, Nvidia introduced a new graphics card. Unfortunately, it’s not game models of GeForce and professional Quadro accelerators. However, their characteristics allow us to learn about future flagships for gamers.

So, was presented three adapters: Quadro 5000 RTX, RTX, RTX 6000 and 8000. Together with them, Nvidia introduced and the architecture of Turing, which formed the basis of new products. Personally, I adhere to the opinion that it is actually the same architecture Volta, only slightly modified. But in order to confirm or refute the need to wait for the game cards and the first full-fledged tests.

Let’s start with the common data. Nvidia is positioning their new products as the first solution for ray tracing. Here I tried to explain what it is and whether it is to us. But in any case, now we are talking about professional adapters, and they just need.

Interestingly, but Nvidia has not indicated the name of the graphic processor, which is based on the adapters. Given the logic of naming Nvidia, I think it’s called GT100 (for older cards).

The GPU contains 18.6 billion transistors and has an area of 754 mm2. In both cases this is less than GV100-generation GPU, Volta. Less new solutions and Executive units. CUDA cores, the new GPU has 4608 and tensor blocks (used for accelerating calculations that are associated with AI) — 576. More precisely, it is the parameters of the eldest of new products, while the graphics processor can have more. For comparison, the GPU GV100 contains 5376 CUDA cores, 672 tensor kernel, 21.1 billion transistors and has an area of 815 mm2. That is GV100 remains the most monstrous graphics processor on the market. I think it will remain so for at least another year or two.

Now about parameters of new products.

The prices are: $ 2,300 for RTX 5000, 6300 dollars for RTX 6000 and $ 10,000 for top — end RTX 8000. As you can see, two high cards differ only in memory size and price. And for an additional 24 GB of memory will have to pay almost 4000 bucks, which is impressive.

In addition, in the table, and this table with the official press release, you will find performance data in TFLOPS, which is actually for these cards is the main criterion that moving company. But there is a performance ghalicheh, which once again would hint at the positioning. In General, I offer you this sign, where the data more.

It has published the resource AnandTech, which is often somewhere finds something that does not advertise the manufacturers. But keep in mind that these data are unofficial.

I was also interested in the performance of the RTX 8000 compared to Quadro GV100. Despite the smaller number of execution units, the first card is higher. Although this is most likely due to the significantly higher core frequency. Speaking of frequencies. If the source of rights, and the peak frequency GT100 professional cards reaches 1730 MHz, in the game you can expect at least something like that, and more and more high values. Also interesting that Nvidia did not mention the level of energy consumption. In General, some kind of weird announcement happened, and it is unclear why it came out that way.

In addition, I note that the cards are equipped with four ports and one DisplayPort 1.4 VirtualLink. Bus NVLink allows you to combine innovations in bundles of two accelerator.

Well, about the future of the gaming adapters. About the frequency I have already said, now about the rest. Assume that the model GeForce RTX 2080 (and it will be called exactly sure about GTX) will get just 3072 cores CUDA, and Quadro 5000 RTX. By the way, it is likely that the latter is based on the GPU GT104, not GT100. Usually between the “hundredth/stultorum” and “locatortype” new GPU generations difference in the number of CUDA cores, just about time and a half. Tensor of the nuclei of the game card should not receive because they do not need them. In terms of future RTX 2080 Ti, which will be released sometime next year, so it will just be GPU GT102 c 4608 CUDA cores. Perhaps they would be a little less or even a bit more if the GPU itself GT100 (GT102 and as his game incarnation) actually contains more nuclei.

It remains to wait for the announcement of the game cards and to check if my guesses. Presumably, they will already show on August 20.

Leave a Reply

Your email address will not be published. Required fields are marked *