NVIDIA will introduce ExaScale Machine in 2017

Monday, November 2, 2009

NVIDIA Corporation has already tried to convince everyone that the union power graphics and central processors will achieve maximum computing power. And recently, head of research and experimental units NVIDIA Bill Dally (Bill Dally) has learned that the company engaged in the development of a powerful solution called ExaScale Machine, which is scheduled for release in 2017.

At the same time on the resource HPC Tech started the slide, from which some details are known about this promising development. According to information posted there, the graphics processor ExaScale will include 2,400 productive nuclei (7200 block of floating point, FPU) and 16 CPU on one chip.

This total TDP of such a decision would be 300 watts, and its performance will be up to 40 teraflops in operation single-precision floating-point and 13 teraflops - when similar calculations double precision. Each node of the GPU will have 128 GB of memory and bandwidth capacity of 2 TB / s, and also receive a 512 GB flash memory to implement the functions of the checkpoint and temporary memory.