The US government handed Intel and Cray a contract to build an exascale computer called Aurora. This supercomputer is capable of performing quintillion operations per second and will go live in 2021. The machine is being built to run AI at extraordinary speeds.
Speeds of powerful supercomputers today are measured in petaflop. The next level is an exaflop, which the US, China,
The department of Energy formally announced the plan and also mentioned that the machine will cost $500 million. It will be housed at the DOE’s Argonne National Laboratory near Chicago. Not much was revealed about the architecture of the computer, but Intel did mention that it will be powered by a future generation of its Optane memory chips and to-be-released Xe GPUs.
As for Cray, they will be housing Intel’s chips inside its Shasta supercomputer. The data will travel around the system with fast optical cables, with the company’s proprietary ‘Slingshot Interconnect’. Intel has learned from recent supercomputers, Summit and Sierra, how important the hybrid CPU-GPU approach is. But instead of relying on a GPU from external specialists, the company is planning to build one themselves.
The growing field of AI and its algorithms run far more efficiently on GPUs rather than CPUs. Aurora is being developed specifically to run deep learning models, with many AI applications already under development.
We don’t know yet whether the US will be the first to cross the exascale barrier, as China had to push back their initial goal of revealing their machine in 2020. Japan is also targeting to launch their computer in 2021, which uses ARM CPUs and uses rapid data movement rather than raw FLOPs.
More information will be revealed with time.