If you are curious about a specific product, it is sure to be a next-generation Intel GPU within the Xe family. To whet your appetite and seeing that the months go by, Raja Koduri wanted to raise a bit of enthusiasm among the community and has published the first real and official image of one of its GPUs, specifically the one that, in theory, is going to be the most powerful on the market for network training and AI: Xe HPC, combining 7 packaging technologies in a single chip.
What Koduri showed is without a doubt an example of what is to come and Intel's muscle. The chip and GPU are huge, and as expected it is shaded in key aspects of the architecture so as not to show too much information to rivals, but everything is perfectly intuited. Is it able to compete in its segment with AMD and NVIDIA?
Intel Xe HPC, the largest GPU built to date with 7 technologies
Xe HPC ready for power on!7 advanced silicon technologies in a single packageSilicon engineers dreamThing of beauty @intel pic.twitter.com/RF8Prsy05f— Raja Koduri (@Rajaontheedge) January 26, 2021
Intel has undoubtedly surprised more than one user and engineer. That the chip was gigantic is not a surprise as such because we already saw it live and compared to its brothers, but today is another level.
It does not have IHS and the full dies are shown in their arrangement on the interposer and PCB, as well as a few more features. Most of these dies are hidden with a shadow, but information about it is glimpsed since Intel has left the divisions visible.
Therefore, we can see the clusters, the cache, and the memory of the UEs without any problem, without having the memory controllers or the PHYs. But from here come the questions, the speculations and the rumors, because there is a lot of cloth to cut in the hidden by the company, so take a little salt and let's get to it.
What we know for now about Xe HPC
This chip will have 16 clusters with 128 EU each, which gives us a total of 1024 EU per die and 2048 in total and 8192 shaders. The architecture used internally will be called Intel Graphics 12.5 and according to Intel, it would run at a frequency of 1290 MHz to offer 21.1 TFLOPS in FP32.
This GPU cataloged within Ponte Vecchio presents according to Koduri 7 advanced silicon technologies, where for example 4 different lithographic processes are found in the same package, different types of chips such as HBM, Foveros, CO-EMIB, ODI and new interconnection technologies by TSV of the company.
Although at first Intel said that it would manufacture these GPUs in its 10nm SuperFin process, there is news in this regard, as they will eventually reach 7nm through external manufacturing (TSMC?).
What we don't know yet
There is speculation, for example, with the fact of including HBM2E or even HBM3 in their VRAMs, but the use of memories such as 3D XPoint or another type of NAND Flash is not ruled out. This is thought by the asymmetric arrangement by Intel in the design, where there are dies that are not correctly explained and all our assumptions about it.
As we see, Intel takes it very seriously and is going for the GPU market ahead of AMD and NVIDIA, can it dominate the market in both deep learning and gaming?
0 Comments