Nvidia is back in the news, this time not for details on its graphics, but on the CPUs. According to Jensen Huang, the company will use the next generation of Intel xeonthe Sapphire Rapids in their DGX-H100. This supposes a substitution of all AMDs EPYC which the company has been using for years, seeing how they cut ties with AMD and decided to go with Intel.
Jensen Huang, the current CEO of NVIDIA, has announced that the company will make a complete transition to future Intel processors. xeon. With this decision, all AMD EPYCs with Zen 3 architecture will be replaced by Intel Xeon sapphire rapids. According to Huang, the main reason for this decision is that Intel has a higher performance per core than the competition.
The NVIDIA DGX H100 will have Intel Sapphire Rapids inside
However, this replacement of processors would only affect its DGX H100 units, which are workstations with a lot of computing power. They can give shelter 8x GPUss NVIDIA H100thus obtaining a total of 32 PFLOPSperformance for AI. Additionally, it contains a total of 640 GB by heart HBM3 thus giving a bandwidth of 24 TB/s per system, almost nothing.
Apart from that, each DGX H100 can be linked with others to give rise to the POD DGX H100, achieving a performance of 1 exaflop in AI, with 20 TB by heart HBM3. But it doesn’t stop there, since we can continue with the same formula, joining 18x DGX POD, for a total of 256 Nvidia H100.
And now, if that wasn’t enough, we have its definitive version: the NVIDIA EOSa supercomputer created by joining 576 NVIDIA Systems DGX H100, with a total of 4,608 NVIDIA H100. This should reach an overwhelming number of 18.4 exaflops AI performance, 4 times faster than the fastest computer in the world.
Intel and the serial production problems of these Xeons
Now this change from AMD EPYC to Intel Xeon Sapphire Rapids makes us somewhat controversial. While AMD’s Zen 3 architecture processor has been with us for some time, the new xeon He didn’t even come out. To make matters worse, a few hours ago it was stated that the mass production Sapphire Rapids processors does not meet expectations.
The latter clashes with previous Intel product updates, which claimed it had already reached end customers. This again leaves the company in a situation with many uncertainty, since we do not know its final specifications or price. So, it could be that NVIDIA decided to make this decision in order to cut ties with AMD or because they saw the real performance of these Intel Xeon Sapphire Rapids.
Regarding the specifications of the Intel Sapphire Rapids, we know that they will reach the 56 hearts Yes 112 threadswith an architecture golden cove under knot Intel 7. To all this, a total of 105 MB cache memory L3, HBM2E80x lines PCI-E 5.0 and a TDP of 350W. The whole would imply an improvement of a 19% more of CPI compared to its predecessor, easily surpassing AMD’s current performance per core.
However, let’s not forget that AMD will also release the new architecture Zen 4 for its EPYC Genoa and Bergamo with up to 128 hearts a 5nm. Maybe they won’t compete to have the best IPC, but it looks like they will be the core leaders in the HEDT CPU market.
via: Toms Hardware