March 22, 2022
The computer is built from 4,608 new graphics processors that include 80 billion transistors each, and 1,152 communication processors from the BlueField family developed in Israel
Anvidia today (Tuesday) announced the NVIDIA Eos supercomputer (pictured above) for artificial intelligence applications, which is based on a new GPU chip architecture called NVIDIA Hopper, and the BlueField-3 DPU high-speed communication processors developed in Israel. The computer is expected to start operating this year. It will be operated by the company Anvidia itself and will serve as a research laboratory in the fields of climate, biology and artificial intelligence. It is expected to provide processing power of up to 18.4 Exaflops in artificial intelligence calculations – four times that of the Japanese computer Fugaku, which is currently considered the most powerful in the field.
The computer will be built from a cluster of 576 new DGX H100 computers, each of which includes 8 graphics processors from the new NVIDIA H100 family. These processors are based on the new NVIDIA Hopper architecture (named after Admiral Grace Hopper of the American publishing house, who was the pioneer of computers in the world. It developed the world’s first compiler and cobalt language was defined based on its works). In total, there are 4,608 H100 graphics processors, manufactured in a 4-nm process by TSMC and including 80 billion transistors in each processor. Each of the H100 processors inside the DGX H100 computers is connected to the other processors via the Israeli ConnectX-7 chip.
The Hopper architecture includes BlueField-3 DPU communications and acceleration processors developed in Israel, which provide internal Ethernet connectivity at speeds of up to 400Gbps. In total, the new super-processor includes 1,152 BlueField chips. Anvidia said H100 processors are expected to be available starting in the third quarter of 2023. They have received support from major cloud providers such as Alibaba, Baidu AI Cloud, Oracle cloud and Tencent cloud, with major server makers already announcing that they will use them on future servers. These include: Atos, BOXX Technologies, Cisco, Dell, Fujitsu, Gigabyte, Lenovo and HPE.
At the same time, Anvidia has announced a new category of CPUs called Grace Superchips, and a fast Ethernet communication technology called NVIDIA Spectrum 4 which has been developed in Israel. Each of the new CPUs will contain 144 Arm cores and memory running at a rate of up to 1 terabyte / second. Each processor is built from two CPU chips connected to each other using NVLinke-C2C technology, which provides fast connectivity in the form of chip-to-chip communication (Chip-2-Chip).
Posted in categories: News, Computers and Embedded Systems, Semiconductors
Posted in tags: BlueField, GPU, Anvidia, PC
