The future of high-performance GPUs for the professional sector lies in multi-chip designs. AMD already confirmed it recently with its new Instinct MI200 series, and NVIDIA will soon do the same with Hopper, an architecture that will coexist with Ada Lovelace, although in different markets, since the former will be used in graphics accelerators for servers and data centers, and the latter in graphics cards for general consumption and for design, rendering and publishing professionals. .
NVIDIA Hopper is therefore the name given to the next architecture of the green giant for the professional sector. Like Ada Lovelace, it will be manufactured in the 5nm process, a significant leap from the 8nm node Ampere uses, and it will use a multi-chip design, which means that NVIDIA will be able to combine two GPUs to create a super GPU. Ada Lovelace will instead maintain a monolithic GPU design.
To better understand the differences between the two designs, it is enough to look at the specifications that, in theory, the top of the range of both architectures could have. In the case of Ada Lovelace, the latest information points to the AD102 chip, a real monster that would add 18,432 shaders or CUDA cores distributed in 144 SM units (72 TPCs, 12 GPCs), and that it would have an approximate power of 90 TFLOPs in FP32. Its memory bus would be 384 bits, and it could be mounted on a graphics card with 24 GB of GDDR6X memory. The TGP of this graphics card would be around 600 watts.
In contrast, the top-of-the-range Hopper-based chip could combine two AD102 GPUs to create a super GPU of 36,864 shaders or CUDA cores, which is equivalent to a total of 288 SM units (144 TPs, 24 GPCs). This GPU bestiality would be divided into two interconnected silicon chips that would use different packages and that, in theory, will use the TSMC CoWoS encapsulation technology. Its launch would take place at the end of next year.
We don’t have much information about the rest of its possible specs yet, but we can almost certainly assume that NVIDIA Hopper will use HBM2E or HBM3 memory (stacked in 3D), and that it will offer enormous bandwidth. (between 3 TB / s and 4 TB / s). In terms of raw power we do not have a defined value, but it is rumored that the jump could be so great that it would allow NVIDIA to reach, with this GPU, the 150 FTLOPs in FP32. In the image that we leave at the end of the article we can see the distribution of the two GPUs in their respective packages.
