Nvidia was the target of a major data leak. We have a lot of information about the future of its catalog. The company is working on different graphics architectures, including “Ada Lovelace”, “Hopper” or even “Blackwell”.
These different codenames are interesting because they highlight a change. Nvidia seems willing to use two names for graphics architectures to better differentiate its “consumer” offerings from its grossing offerings. Currently, the Ampere architecture is at the heart of both offerings, despite some mechanical tweaks. The compute offering has FP64 and Tensor components, while the Ampere graphics line removes them in favor of RT cores and other more relevant graphics components.
Graphics Architectures Ada Lovelace, Hopper, and Blackwell
Much like AMD and its “RDNA” and “CDNA” offerings, the next generation of cards will be organized around Ada Lovelace and Hopper. Ada Lovelace will spawn several GPUs, including the “AD102”, “AD103”, “AD104”, “AD106”, “AD107” and “AD10B”. The “computations” division, code-named “Hopper,” will produce the “GH100” and “GH202” graphics processors. Another architecture for computations or data centers is also mentioned “Blackwell” with solutions named “GB100” and “GB102”.
Nvidia plans to launch the GeForce 40 series “Ada Lovelace” graphics cards in the second half of 2022, with the possibility of updating its professional line (Hopper) by the end of 2022 or next year. Eventually, “Blackwell” could follow “Hopper.”
GeForce RTX 4090, what do we know?
The AD102 GPU is said to benefit from a 5nm etch performed by TSMC. This GPU is said to use a monolithic design with an estimated matrix size of 600 mm². Its equipment is said to include 18,432 CUDA cores, which is almost double the equipment of the Ampere architecture.
A frequency of 2.2 GHz would allow 81 TFLOPs in FP32, more than double the 36 TFLOPs of the GeForce RTX 3090 series.