The parameters of the first supercomputer on the Ampere GPU hint at this.
Nvidia is due to launch a new generation of GPUs called Ampere this year. They will be produced according to the seven-nanometer process technology and, of course, will be more productive than the current generation of GPUs.
But how much more productive? We do not yet have any intelligible technical information about the new architecture and new GPUs, but there are already certain hints of speed.
Reportedly, Indiana University was the first to receive the Cray Shasta supercomputer, based, among other things, on the new generation of Nvidia graphics accelerators. The installation, also known as Big Red 200, cost the institution $ 9.6 million.
To clarify, Big Red 200 is the name of a particular installation installed at the university, and Shasta is the name of a new class of Cray supercomputers. In particular, soon such a supercomputer of the pre-exascale level, which will be called Perlmutter, will receive the National Laboratory named after Lawrence at Berkeley.
Returning to Big Red 200, it is reported that initially this supercomputer was supposed to be based on Nvidia Tesla V100 accelerators and have a performance of 5.9 PFLOPS. However, in the end, the project managers decided to wait a bit to get the system already on the new Nvidia 3D-cards. There are no names, but it is said that, previously, the new GPUs are approximately 70-75% more productive than their predecessors.
Of course, we are talking about specialized accelerators, but this is not particularly important, since we are talking about growth, and it will be approximately the same for all classes of video cards. That is, it is quite possible that Ampere 3D-cards will be 70-75% more productive than similar Turing solutions, and this is more than impressive.
It is also worth clarifying that now Indiana University has received only part of the supercomputer based solely on processors, moreover, it is AMD Epyc 7742 CPU. In summer, Cray will install the second part of the system, which will contain not only processors, but also the same new accelerators.
And this gives us an understanding of when the first Nvidia accelerators on the new architecture will hit the market. Most likely, in the summer we will receive only specialized Tesla adapters, and GeForce gaming cards will be released at best in the fall.