Click for Full Size
HPCTech.com has leaked a slide purportedly coming from NVidia’s Chief Scientist Bill Dally that details an ExaScale machine. How big is an “Exascale” ? Well:
- GigaScale = 1G
- TeraScale = 1000G = 1T
- PetaScale = 1000T = 1P (The biggest HPC’s in the world right now are considered PetaScale)
- ExaScale = 1000P = 1E
The problems in building a traditionally architected ExaScale machine are daunting, typically requiring Power and Cooling beyond anyone’s imagination. However, using NVidia’s new GPU’s (Next generation of Fermi?) it’s a possibility.
According to the slide leaked by hpctech.com, the ExaScale GPU will pack 2400 throughput cores (7200 FPUs) and 16 CPUs on a single chip with a TDP of 300W, delivering up to 40TFLOPs of single-precision floating-point processing power or 13TDLOPs of double-precision floating-point processing capabilities. Besides, each node of the chip will feature 128GB of memory, 2TB/s bandwidth, 512GB Phase-change/Flash for checkpoint and scratch.
via NVIDIA Scientist Promises ExaScale Machine in 2017 – Expreview.com.
