This is a multi-chip solution
Intel today hosted Architecture Day 2020, where it revealed a lot of new things. And, probably, one of the most impressive announcements was the announcement of the line of HP XE graphics accelerators.
It is worth mentioning right away that they are not aimed at the consumer segment, but at data centers, that is, you need to compare XE HP cards, for example, with the Nvidia A100 or the upcoming AMD Radeon Instinct Mi100 card.
What makes the Intel product so special? At least by its parameters and performance. Accelerators of the XE HP line will exist in three versions: with one chip, two and four. Intel itself calls the chips tiles. That is, the GPU XE HP, in addition to the base one, will be multi-chip, like modern desktop and server AMD processors.
Each chip will contain 512 execution units with 4096 “cores” each. Accordingly, the top-end solution of the line will boast 16 384 cores!
It’s interesting that this information surfaced on the Web at the beginning of the year, and now it turned out that all this was true. As a result, the performance of the top-end accelerator with four chips, clocked at 1.3 GHz, will be an incredible 42 TFLOPS (FP32)! For comparison, the same Nvidia A100 with its monstrous GPU containing 54 billion transistors has a performance of “only” 19.5 TFLOPS.
True, the advantage of Intel’s solution can be explained very simply. The fact is that this very performance in TFLOPS is calculated using a very simple formula: the clock frequency in MHz is multiplied by the number of cores in the chip and multiplied by the number of instructions processed per clock, and then all this is multiplied by 10 6 . That is, when calculating such performance, numerous architectural features and so on are not taken into account, but these dry numbers are not always needed. However, HP’s top-of-the-line XE accelerator is very impressive, and what’s even more interesting is what Intel has created.
Intel during the presentation showed what its product is capable of. The XE HP handles 10 streams of HEVC 4K video at 60 fps in real time with just one chip. At the same time, Intel talks about almost perfect scalability, that is, the top-end version of the XE HP will be able to handle 40 threads already!