In recent days, with this news, we spoke to the first time of Lovelace : yes tra tta of the code name which identifies what should be the future generation of NVIDIA architecture for GPU . There are very few certainties with respect to this new architecture: the only data on which we can converge with relative certainty is linked to the use of the production technology at 5 nanometers for the construction of these GPUs.

Precisely the passage to a more sophisticated production process than that adopted at the moment for the solutions of the Ampere family (GeForce RTX 3000) will allow you to considerably increase the amount of processing units inside these GPUs and thus obtain higher overall performance.

The indiscretions that are emerging these days are to be taken with all the necessary cautions : we are very far from debut of these GPUs and the internal plans of NVIDIA, as well as not having been officially communicated, are subject to possible upheavals in a very short time. That said, for the higher-end Lovelace proposal indicated by the code name of AD 102 we are talking about a number of CUDA cores equal to well 18. 432 .

GA 102 has a “7 6” structure.

Maybe AD 102 will get a “12 6 “structure.

– kopite7kimi (@ kopite7kimi) December 28, 2020

This figure would be obtained starting from a structure with a number of GPC, Graphics Processor Clusters, which increases from the current 7 of the GA GPU 102 used in the GeFroce RTX cards 3090 and GeForce RTX 3080 up to 12 . This would therefore imply the presence of well 144 Streaming Multiprocessors, each of which integrates 128 CUDA cores for a total of 18. 432 which represents the theoretical maximum achievable with the AD GPU 102.

On paper, an increase of this type in the number of CUDA cores, and in parallel of the other processing units inside the GPU, would allow to record a significant boost in overall performance . All this, of course, provided that these are the actual specifications of the top-of-the-range Lovelace GPU and that other components do not become a bottleneck for the GPU: let’s think for example of the innovations required by the video memory sector to keep up with a chip so complex and powerful.

We remember in any case that these are only preliminary rumors , not official data from NVIDIA or partner manufacturers. The expected debut of the Lovelace solutions is still a long way off, given that the first Ampere boards made their debut on the market only in September 2020.