Rumor | NVIDIA 5 nm Lovelace AD102 (RTX 4080/4090?) specs leak, looks to be a monster GPU with 18,432 CUDA cores and nearly 66 TFLOPs of FP32 performance

NVIDIA's Ampere launch is still pretty fresh with many prospective buyers not being able to buy them even nearly four months after launch. However, NVIDIA is on a fast track to develop Ampere's successor. Though initially thought to be Hopper, we are now learning that Lovelace (named after British mathematician Ada Lovelace) is the architecture that is next in line after Ampere, and the initial specs already seem pretty droolworthy. 

Reports are coming in of a possible AD102 GPU in the works. First revealed by known leaker @kopite7kimi, the good folks over at 3DCenter have deduced some of the specs of the upcoming GPU. According to 3DCenter, NVIDIA AD102 will sport 2x the graphics processing clusters (GPCs) of Turing i.e. 12. This indicates that AD102 may house as many as 72 texture processing clusters (TPCs) and 144 streaming multiprocessors (SMs).

This eventually means that we might see a total of 18,432 CUDA cores — a 71% increase compared to Ampere's GA102 die. Assuming a clock speed of about 1.8 GHz, we can see that NVIDIA Lovelace could theoretically offer nearly 66 TFLOPs of FP32 single-precision performance. 

Of course, such massive core counts means a shift to a 5 nm process. The question remains whether NVIDIA will continue with Samsung or go with TSMC. There were reports of NVIDIA pre-booking TSMC's 5 nm capacity for 2021 while also leveraging Samsung for low-volume orders.

NVIDIA Lovelace could be a stop-gap solution branded as the RTX 40 series until a total architecture revamp sees the light of the day in the form of Hopper, which is expected to be a multi-chip module (MCM). It is also possible that Hopper would be restricted to datacenters as well, but it is too early to speculate.