If the information is true, it won’t be long before NVIDIA presents Hopper for the first time, a GPU architecture for HPC that brings with it many unknowns and novelties. The first of these is whether we are facing the first graphics card built by chiplets and therefore by several chips instead of just one. The second concerns rumors about its internal organization, which will undergo a major change, so much so that we had not seen one like this since the days of the GTX 8800.
Hopper specs according to NVIDIA leak
The latest rumors regarding the Hopper architecture are that it will be completely monolithic and therefore composed of a single chip. In any case, we should remove the doubts within a few weeks, since NVIDIA will soon celebrate its GTC where its new and gigantic graphics for HPC should be presented. And what does the NVIDIA leak tell us about the GH100? Well, for starters, we can know its configuration.
- 8 GPC and total.
- 9 TPCs for GPCs.
- 2 SM per TPC.
who gives us a total of 144 SMs, the same figure as the GA100, but it must be borne in mind that these GPUs are not designed for the gaming market and within their SM they replace certain elements with others compared to those used in the RTX . For example, the intersection unit for Ray Tracing, the RT Core, is of little use in scientific calculations, so it is usually removed. At the other extreme, we have 64-bit floating point units, which are superfluous in games, but not in areas where data precision is very important.
A 144 SM configuration might seem small compared to the MI200’s 220 compute units and its dual GPU configuration. However, it is more than possible that NVIDIA will do the same than with the RTX 30 and doubles the number of floating point units, by commuting them with those of integers. So, in the end, the amount of compute units in the GPU would end up being greater than AMD’s own design. Expected performance? 3 times larger than the current GA100, something impossible without doubling these units.
New unit and organization at Hopper: the CPC
However, there could be changes to Hopper’s organization according to the NVIDIA leak, since it would be introduced a unit between the TPC and the GPC called CPC, which would be composed of 3 CPCs. Which makes us suppose that for the new architecture they could implement an additional level of cache between the L1 of the SM and the global L2 of the GPU. Something that wouldn’t be impossible as long as that’s what AMD has done with their RDNA and it’s a way to relieve stress on the L2, especially in terms of read requests. A change that would make the architecture totally different from what is expected of the RTX 40 for gaming.
NVIDIA Hopper: very limited gaming capabilities
Although we call them graphics cards, these hardware components use GPU cores for large-scale parallel computations because they are inherently more efficient than a CPU in these parts of the algorithms. Thus, instead of increasing the number of processor cores, the card is used as a coprocessor, not to generate images on the screen, but to obtain mathematical results at a much higher speed.
We’ve already seen how, for example, AMD removed all graphics capabilities from CDNA, well, according to the NVIDIA leak in Hopper, the same will happen in all GPCs except for the first, which will be the only one that can generate graphics. This makes us think that it will have asymmetric GPCs where the fixed function units will only be present in the former. So it will have basic graphics capabilities, but when it comes to playing games it will be quite limited and at a much worse level than an RTX 4050, therefore NVIDIA Hopper is not intended or intended for games.