Next-gen NVIDIA GPUs 70-75 percent faster for deep learning?

Late last week, news hit the web that the future Cray "Shashta" supercomputer of Indiana University (IU) will feature next-gen NVIDIA GPUs. The first-phase of the supercomputer, which is codenamed "Big Red 200", involves the installation of 672 dual-socket nodes with AMD EPYC 7742 "Rome" processors.

Where it gets interesting is that the second phase (slated for this Summer) will involve the installation of more EPYC nodes, which will be used in combination with next-gen NVIDIA Tesla GPUs. Big Red 200 promises up to 6 petaflops of computing power.

Brad Wheeler, IU vice president for IT and chief information officer, told The Next Platform that an opportunity presented itself to wait a little longer to implement NVIDIA's latest technology. The report from Next Platform mentions that the new NVIDIA chips are 70-75 percent faster than the current-generation, but it's not clear whether that is a direct quote from Wheeler or speculation on the website's part.

According to The Next Platform who are quoting Wheeler, the addition of new GPUs was “something of a fluke”, because “an opportunity presented itself to wait a bit longer and move up to Nvidia’s newer technology“.

The original plan for Big Red 200 was to use Tesla V100 (Volta) processors, but IU ended up buying a smaller number of next-Gen GPUs from NVIDIA, which are expected to deliver total computing power to 8 petaflops from original 5.9.

Below is a video of the installation of the first-phase of the Big Red 200 supercomputer.



Via: VideoCardz