Nvidia “Lovelace” can get 18,432 CUDA cores

Just before Christmas, we reported that Nvidia looks to postpone the graphics architecture “Hopper” and its modular structure to the future and instead stop at one last monolithic architecture before that. This is said to be called “Lovelace”, where the GPU designation for the top model looks to be AD102 – and now the leaker means @ kopite7kimi that the circuit can have as many as twelve graphics clusters.

Videocardz is available with mathematics – the RTX 2000 series “Turing” circuits have a maximum of six graphics clusters (GPC) of six clusters with texture processors (TPC) and twelve stream processors (SM). As each SM unit contained 64 CUDA cores, the total number of CUDA cores in the Turing case was 4,608. With Ampere, another GPC was added, which together with a creative way of calculating meant that the number of CUDA cores increased up to 10,752.

With Nvidia’s latest number of CUDA cores per GPC, 1,536, the AD102 “Lovelace” with as many as twelve graphics clusters would thus be equipped with as many as 18,432 CUDA cores – 71 percent more than in the Geforce RTX 3090. To physically fit everything this on a circuit is believed the introduction of “Lovelace” hang on a shrinkage in the manufacturing process of the manufacturing partner Samsung.

Today’s “Ampere” is manufactured by Samsung at 8 nanometers, a process that has been anything but trouble-free. Samsung itself already has circuit production of 5 nanometers, a process the company after a problematic introduction now seems to have got the right side of. Exactly when an introduction of “Lovelace” can be expected is unclear, but the new graphics circuits will probably not be launched until late 2021.

Read more about Nvidia: