New rumors about NVIDIA GeForce RTX 4090: the flagship graphics processor AD102 enters into the test phase, GDDR6X 24 Gb / s memory is ready:

It was the assumption that NVIDIA can choose the name of the GeForce RTX 50 series nameplate instead of the expected branding of the GeForce RTX 40 series, but it seems that the Kopite7kimi insider declared the NVIDIA decision to adhere to the Naming Scheme of the 40 series. In addition, one important milestone is that NVIDIA is possible , I have already started testing and evaluating your flagship graphics processor ADA Lovelace, AD102, which will work on a series of video cards, such as RTX 4090 and RTX 4080.

Now it is definitely not known whether NVIDIA has launched Ada Lovelace graphics processors into mass production, but it can occur in the middle of 2022, as previously hinting. So the chips tested here can be early samples. But if these chips really are in the wild, then we can see some leaks in the coming months. Of course, initial performance indicators will not be as high as these graphics processors have not yet received final drivers, but it can give us an idea of ​​what clock frequencies we can expect in the final versions.

Another interesting thing that Kopite7kimi says is that NVIDIA is going to use 24 Gb / s memory crystals on its Ad102 graphic processor. Insider mentions it only for RTX 4090, and we know that RTX 3090 is the only series that currently has the fastest GDDR6X memory solutions. It is not yet known whether NVIDIA output speed will select 24 or less for the rest of the line. One thing you can say for sure: we will get the speed speed of more than 1 TB / C on consumer gamecraft, which will give us a 15 percent increase in bandwidth indicators, and this is excluding completely new compression technologies that will be on board the ADA Lovelace.

NVIDIA GeForce RTX 4090 "AD102 GPU" - what do we know now (?):

The video cards of the NVIDIA GeForce RTX 40 series, equipped with AD102 graphics processors, will have a TDP to 600 W. At least, this is what the current BIOS is calculated, supplied by partners on the board, so rumors about the TDP 450-600 W can be true, but the final figures are not yet known. Power indicators are usually overestimated at the test stage, so they can be optimized with the actual launch of cards. Maps will be equipped with PCIe Gen 5 connectors and will be supplied with an adapter 4 x 8-pin per 1 x 16-pin to support huge power consumption. The upcoming GeForce RTX 3090 Ti will be supplied with an adapter with 3 8-pin per 1 16-pin.

In the leaked design of the printed circuit board it was said about memory, which will have 12 points of soldering on the printed circuit board, and all of them are compatible with Micron GDDR6X memory. Cards of a higher level can be delivered with unilateral and double memory, since they provide the best balance of power and temperature and have a container of up to 24 GB, but at higher speeds (up to 24 Gbit/s). As for the main segment, we will most likely see structures at a speed of 20 Gb/s and higher, but in versions of 8 GB and up to 16 GB, which can help reduce energy consumption, since power regulation will be reduced to 3 VRM for memory.

As for the cooling of these huge printed circuit boards, NVIDIA, as reported, is going to reuse its BFGPU design with three slots, while the partners are harvested to use cooling solutions with 3.5 and even four slots weighing more than 2 kg. Most AIB can simply use the AIO and hybrid cooling designs that you will see in RTX 3090 Ti. It is expected that cards will have up to 28 VRM phases on the flagship video card NVIDIA GeForce RTX 4090, so all this additional cooling will be used with benefit.

The family of graphic processors NVIDIA ADA Lovelace is expected to make a jump between generations, like Maxwell to Pascal. It is expected that it will be launched in the second half of 2022, but it is expected that shipments and prices will be similar to the current cards, despite the fact that NVIDIA spends billions of dollars to purchase these good good 5-nm TSMC plates.