Hopper, from NVIDIA, has generated enormous expectation because, according to the information we had been seeing, it was going to become the first architecture of the green giant to make the leap to an MCM core design, that is, multichip module. As we told you at the time in this article, this new design would allow you to create a super GPU by combining two or more “modest” GPUs.
It is a simple concept at its base, but in case you have any questions I want to use a concrete example, and pit it against its equivalent on a monolithic core GPU. A monolithic core GPU with 18,432 shaders would have to pack that huge number of shaders onto a single chip of silicon. This has advantages, since all the elements of the GPU are on a single silicon chip and there is no need to use interconnect systems, but it also has disadvantages. The most important is that translating that design to the wafer is tricky, since you are dealing with a huge number of shaders on a single silicon wafer, and this means there is a higher risk of getting units that don’t have all the working shaders.
By cons, that same GPU under an MCM design could be divided into two silicon wafers with 9,216 shaders each. This design has an important advantage, and that is that it would be easier to transfer to the wafer, but it also has disadvantages, and that is that it would be more complicated to put into practice, since it would be necessary to use an efficient interconnection system capable of joining both. GPUs. To give you an idea, the problems that could arise, in the first foray into the field of MCM GPUs, would be similar to those we saw with the arrival of the first generation Ryzen, which also used a multichip module design.
The fact is that, according to new information, NVIDIA’s Hopper architecture will not make the leap to an MCM design, but rather will keep the traditional monolithic core design. If this is confirmed, I think that NVIDIA could surprise us with a really amazing GPU because, in the end, we would be talking about many thousands of shaders built into a single chip of silicon. I already found the jump to 10,752 shaders that occurred with the Ampere-based RTX A6000 impressive, so imagine what Hopper could do.
However, maintaining that monolithic core design could have a negative consequence, and that is that the consumption of the most powerful GPUs based on Hopper could be quite high. We have to wait and see what NVIDIA finally does, but everything we have said in this article makes sense, and it is perfectly feasible.