Today, Intel’s Chief Architect Raja Koduri showed off the company’s supercomputer GPU. As we saw from his SC19 keynote, this was set to be a big milestone in terms of GPU architecture. See Intel Xe HPC GPU is Something to Get Excited About. Now, that GPU is closer to reality.
Intel Xe HPC Ponte Vecchio Packaging
At Raja’s SC19 presentation, he outlined how the Ponte Vecchio GPU would be different and designed specifically for exascale supercomputing.
At the time, he said the new chip would use various technologies for connecting HBM as well as what Intel is calling the “RAMBO” cache.
As part of that discussion, we expected a number of tiles to be integrated with different components from different fabs onto a single package.
Today, we have a glimpse of what that packaging looks like and it roughly aligns with what we saw in that artist rendering around 5 quarters ago.
— Raja Koduri (@Rajaontheedge) January 26, 2021
Rotating and brightening the image a bit gives us a look at the package:
From what we can see, there are 25 tiles on each side of the massive GPU. That means we are potentially looking at a GPU with what is a fairly extreme level of silicon packaging and integration effort. There also may be silicon being stacked here that we cannot see.
This aligns very well with the Future Intel Methodology outlined at Intel Architecture Day 2020. There, the company focused on a next-generation of packaging where it would take different IP blocks and rely on packaging to integrated into a larger system.
If we look at the Ponte Vecchio package shot today, it looks like that is exactly where Intel is heading with its new CPU.
This is truly impressive for Intel, assuming it powers on. At the higher end of the market, we get to see some of the more exotic technologies that will eventually filter into more products. What we see in the Intel Xe HPC GPU aligns with what we are seeing Intel message as its methodology well beyond its high-end GPUs.
Intel has been touting packaging technologies as a key differentiator since it can use them to meld IP from different fabs, not just its own. This is certainly a few steps above what current-generation multi-die NVIDIA and AMD solutions offer in the GPU/ CPU space so it will be interesting to see how those players evolve to address Intel taking the next step in products.