There are only a few hours left for Jensen Huang to give his speech at this year’s GTC, and the first images of the NVIDIA GH100 have already been leaked. Based on Hopper architecture, it is the most powerful processor graphics card NVIDIA has released to date and the first under TSMC’s 5nm node. So let’s see what it looks like.
Hopper is a codename that has been buzzing around NVIDIA for some time, first in the form of a future chiplet-based RTX 40. However, as of today, the RTX 40 is called Lovelace and the graphics card we are talking about is a monolithic block. Information that we already knew in advance thanks to previous leaks, but that, as you will see later, has been confirmed through the first images of the GH100 that have been leaked.
However, it is through the NVIDIA leak where we know more data, for example we know that its maximum configuration is 144 SM organized in 8 CPGs where only the first one has the ability to display graphics. As for the latest rumors tell us about a total L2 cache of 48 MB. Although at the moment the details regarding the novelties of the architecture are scarce, how is the internal composition of its different nuclei and how these have evolved with respect to its predecessor. We also hope is the definitive presentation of the Big Accelerator Memory which we told you about a few days ago, as well as the commercial use for the first time of the optical version of the NVLink.
First images of the NVIDIA GH100
For some time now, the HPC market has become one of NVIDIA’s spearheads, since the same power and architectures used to generate extremely realistic graphics in games can be applied to other fields, e.g.n special to the scientific world, but this forces certain changes in a GPU that is focused on that market. The most important of all? The floating point support with 64-bit precision, something that has no value in video games, but that is crucial in certain areas, where a simple variation can lead to different results. Which has led to apart from the gaming architecture used in the GeForce RTX we also see designs like this GH100.
So, as you can see in the first images of the NVIDIA GH100 with its 6 HBM2E memory chips It is not designed for you to use it to play your favorite games, since it lacks a series of elements that are key when it comes to reproducing the graphics and, on the other hand, it has a series of features that are not found on a home PC unless that one is a professional in certain fields, because they do not have any added value. Namely, We are not facing the RTX 4090 chipbut if possibly in front NVIDIA’s most powerful graphics card as far as raw computing power it means.
However, we will have to wait for the NVIDIA conference to have all the full details of it, for which there are very few hours left. Let’s not forget that NVIDIA has stiff competition in the form of AMD’s Instinct MI200 and Intel’s future Ponte Vecchio.
