Nvidia Nfinity Fabric

NVIDIA Unveils Beastly 2 Petaflop DGX-2 AI Supercomputer With 32GB Tesla V100 And NVSwitch Tech (Updated) | HotHardware

Nvidia has declined to release their newest Tesla architecture on the consumer market, but this is the first indication that they have their own "infinity fabric" like AMD, that can scale across gpu cores at incredibly fast rates of 2.7 terabites per second. They are obviously supporting Intel's new NVlink as a new standard for this sort of thing. The current known version of the consumer Tesla graphics card is capable of 38tf, and this one should be capable of easily double and quadruple that, by merely cramming a ton of HBM3 memory onto the same infinity fabric they use for the gpu cores.

160tf is the estimated requirements for running a Star Trek holodeck. The current 38tf Tesla should come on the market at around $3,000.oo next year, and consumer electronics prices are cut in half every 3 years. 14tf are the estimated requirements for running any existing video game at high frame rates and running VR and real time ray tracing. 26-38tf are estimated for photo-realistic virtual reality. AMD's new Navi graphics cards are scalable, use more open source architecture and programming, and are better suited for consumer applications in the long run, but are usually less powerful. However, with these kinds of graphics cards the question becomes what the hell do you want to do with all the power you can afford to buy? And, how noisy is the damned thing! Adding all those gpu cores and memory means power and heat are being pushed to the limit.