A lot of attention on the speech of NVIDIA head , dedicated to the opening of the exhibition Computex 2017, was given to calculations acceleration . In fact, it was in this context that the event mentioned the Volta generation accelerators. It was announced that the partners of the company for the production of server systems HGX-1 are concentrated in Taiwan: they are Foxconn, Inventec, Quanta and Wistron.
In one such system it is possible to combine up to eight Tesla P100 or Tesla V100 accelerators, which will be reported via the NVLink interface, and will provide performance at the level of one petaflops. These accelerators will be connected to the motherboard via the SMX connector, which we will discuss below.
Once again, the characteristics of NVIDIA GV100 - the largest processor in the world, were mentioned, according to the head of NVIDIA. The development of this processor was conducted for three years, and cost the company a decent amount. Actually, it is the crystal size and the amount of cheap memory such as HBM2 that allow us to assume that this kind of graphics processor will not exactly appear in gaming graphics cards.
The Tesla V100 module in SXM version was demonstrated not only on sketches, but also alive - the head of the company held it in his hands, as at the previous event.
It was said that due to the use of new cores for tensor computations, Volta processors provide a 12-fold advantage in arithmetic speed in comparison with Pascal.
Finally, the "dark side of Volta" appeared in front of camera lenses:
The Tesla P100 computation accelerators of the Pascal generation also have a similar design, which allows them to provide compatibility with HGX-1 systems. Related Products :
|