NVIDIA's next generation of GPUs, likely to be called Ampere, will offer significant performance gains over current solutions, according to The Next Platform. True, we are talking about graphics processors used in computing accelerators.
Accelerators based on next-generation NVIDIA GPUs will be used in Indiana University's (USA) Big Red 200 supercomputer built on the Cray Shasta platform. They will be added to the system this summer during the second phase of supercomputer construction.
At the moment, it is not specified which GPUs these will be, because NVIDIA has not yet introduced them, but apparently we are talking about a new generation of Tesla accelerators based on Ampere. It is highly likely that NVIDIA will announce a new generation of its GPUs as early as March at its own event. , and then, just in time for the summer, new accelerators based on them should be ready.
It is reported that initially the Big Red 200 system was planned to be equipped with current Tesla V100 accelerators on NVIDIA Volta GPUs. This would make it possible to achieve a peak supercomputer performance of 5,9 PFlops. However, after that it was decided to wait a bit, dividing the construction of the Big Red 200 into two stages, and use newer boosters.
During the first phase of construction, a system of 672 dual-socket clusters was created based on 64-core AMD Epyc 7742 generation Rome processors. The second phase will include the addition of new Epyc Rome-based nodes that will be powered by one or more next generation NVIDIA GPUs. As a result, the performance of the Big Red 200 will reach 8 Pflops, and at the same time, fewer GPU accelerators will be used than planned.
It turns out that the performance of next-generation GPUs will be 70-75% higher compared to Volta. Of course, this refers to the "naked" performance in single precision operations (FP32). Therefore, it is now difficult to say how much the statements about such a significant increase in performance are relevant for the new generation of GeForce consumer video cards. Let's hope that ordinary consumers will also get significantly more powerful GPUs.
Source: 3dnews.ru