NVIDIA has not yet completed the rollout of the RTX 30 series, but they have already started to emerge first rumors about the GeForce RTX 4000, a new generation of graphics cards that will arrive between the end of 2021 and mid-2022, and that according to the latest information could be based on Ada Lovelace architecture, and not on Hopper.
I am sure that our more advanced readers will have been able to see, without problems, why is this distinction so important at the architectural level. If you haven’t been able to discern it, don’t worry, let’s explain it right now.
Numerous information has been pointed out that Hopper to Debut MCM-Type Design in GPU Industry. This means that with such an architecture, NVIDIA will adopt an approach similar to what we have seen in Ryzen processors: join several chips to create a “super chip.”
Currently, an NVIDIA GPU integrates, in a single silicon chip, a certain number of SM units, which leaves us with a specific count of shaders, texturing units, raster units, RT cores and tensor cores. Well, with Hopper, that monolithic core design will theoretically be a thing of the past. The GPU will be made up of several silicon chips intercommunicating, which will create a much more powerful GPU with a huge number of shaders.
GeForce RTX 4000 will use monolithic GPUs
As long as this information is fulfilled, and they are based on the Ada Lovelace architecture. The truth is that it makes a lot of sense, Ampere architecture has shown that doubling the number of shaders per SM unit has not had the expected impact at the performance level, that is, linear scaling has not been achieved compared to the previous generation, but it has increased the complexity when transferring the design to the wafer.
With that in mind, Ada Lovelace architecture is likely to focus in improvements that could go beyond gross power, that is, the large-scale increase in SM units. However, this is not to say that Hopper has been canceled, just that it will not be used to bring the GeForce RTX 4000 to life.
The Hopper architecture would become, according to this information, NVIDIA’s big bet for your professional graphics cards from 2023-2024. GPUs with MCM designs that allow the integration of huge amounts of shaders, also known as CUDA cores, make much more sense at a professional level, especially due to the boom that certain sectors are experiencing, such as artificial intelligence and Big Data.
We do not yet have specific details on the GeForce RTX 4000, but I am convinced that NVIDIA will continue to bet on specialization at the core level, and that said generation will again differentiate between shading engines, geometry engines, texturing engines, tensor cores for AI and RT cores for ray tracing.