In recent months there has been a lot of talk about the plans for NVIDIA as for the next generation video cards, that is the solutions GeForce RTX 4000 architecture-based Ada Lovelace.
The data theft carried out by the Lapsus $ hackers should have already revealed the main technical specifications of the upcoming GPUs, but one element that remains the subject of much debate is that of consumption. The use of a connector capable of conveying up to 600W obviously suggests very high values, and in fact there has been rumors of a power required up to 850W for an unspecified top of the range.
An indiscretion that, although initially considered “too excessive to be true”, seems to find confirmation in the latest rumors spread by the leaker kopite7kimi. According to a tweet, NVIDIA would be testing a next generation card with two 16-pin connectors of new generation necessary to support a request for well 900W.
In fact, there is another full-fat AD102 SKU with 900W TGP, 48G 24Gbps GDDR6X, 2*16pin and higher frequency.
But no one knows whether it will become an actual product. Because the test board of AD102 has more than two 16pin connectors, so everything is possible.– kopite7kimi (@ kopite7kimi) April 27, 2022
This solution may not be a GeForce RTX 4090 but a possible RTX 4090 Ti or a TITAN RTX, it is not yet known. The leaker claims that the model in question would be accompanied by 48GB of 24Gbps GDDR6X memory, which would make it a much more powerful solution than the RTX 3090 Ti which today represents the top of the offer. However, it is necessary to specify how the leaker points out that it is not certain that this card will actually become a commercial product.
Apart from this unspecified model, which should be based on an AD102 GPU with all active units (therefore 18432 CUDA core, as per leaked specifications), NVIDIA would be working on a GeForce RTX 4090 based on the same GPU but with less CUDA coreflanked by 24GB of 21Gbps GDDR6X memory for a TDP of 600W.
The solutions RTX 4080 e RTX 4070 they should instead rely on two different GPUs, respectively AD103 e AD104. The first should have a maximum of 10752 CUDA cores, but it is not known at the moment how many will be active. The card in question should have 16 GB of memory, an increase of 4-6 GB compared to the current RTX 3080 (12-10 GB) for a TDP of around 350W. The GeForce RTX 4070, on the other hand, should have 7680 or fewer CUDA cores and 12 GB of memory, an increase from the 8 GB of the current RTX 3070 / Ti. As for the TDP we talk about approx 300W.
| GPU specs leaked in the Lapsus $ leak | |||||||||
|---|---|---|---|---|---|---|---|---|---|
| GeForce RTX 3000 (Ampere) | GeForce RTX 4000 (Available) | ||||||||
| GPU | SM | Cache L2 | Bus | CUDA Core | GPU | SM | Cache L2 | Bus | CUDA Core |
| GA102 | 84 | 6 MB | 384 bit | 10752 | AD102 | 144 | 96 MB | 384 bit | 18432 |
| GA103 | 60 | 4 MB | 256 bit | 7680 | AD103 | 84 | 64 MB | 256 bit | 10752 |
| GA104 | 48 | 4 MB | 256 bit | 6144 | AD104 | 60 | 48 MB | 192 bit | 7680 |
| GA106 | 30 | 3 MB | 192 bit | 3840 | AD106 | 36 | 32 MB | 128 bit | 4608 |
| GA107 | 20 | 2 MB | 128 bit | 2560 | AD107 | 24 | 32 MB | 128 bit | 3072 |
Recall that the GeForce RTX 4000 series GPUs will debut by the end of the year, probably between September and October. These models should then be followed by others in the medium and low range, from the GeForce RTX 4060 to scaling.