NVIDIA Ampere could be around 17.6 TFLOPs, according to a Twitter user

Robert Sole 3 February, 2020

0 1.099 1 minute of reading

This weekend we echoed the commissioning of the Big Red 200, the Indiana University supercomputer. This supercomputer has two special elements: it is based on AMD EPYC processors and NVIDIA Ampere graphics cards. Ampere graphics cards will be announced in May during the GTC. Before this, user @ dylan552p on Twitter has calculated the relative power of the NVIDIA Ampere.

The calculations are relative. We must bear in mind that the user has calculated that the Ampere have 25% more power in TFLOPs than the Turing. It is a very relative data that we must take for what it is, a hypothesis. According to the calculations made by this user, Casa Ampere would offer 18 TFLOPs in 64-bit floating point (FP64).

[amazon box="B0817726L8"]

Possible power of the NVIDIA Ampere

Big Red 200 mounts 1.344 AMD EPYC 7742 processors accompanied by 256 graphics cards. The 1.344 AMD processors offer us a performance of 3.15 PFLOPs based on FP64 performance. The Big Red 200 system has 8 power PFLOPs, which means that the remaining 5 PFLOPs belong to the graphics.

Approximately each of the NVIDIA Ampere cards individually offers 17.6 TFLOPs. We must bear in mind that all this is very relative and there is no data that, at the moment, confirms or denies the data. The combination of the AMD EPYC and NVIDIA Ampere processors would add about 18 TFLOPs of power.

Will this be the final power of the Ampere? Well, it is not known, because we do not know what silicon is being used in these systems. How many CUDA Cores does he have? At what frequencies is this graph working? How many Tensor Core does this system integrate?

It is fine as speculation and to move in a relative power framework, but there are many factors and variables that are not taken into account. We will leave doubts about its performance and power during the GTC 2020 in May.