Nvidia investigating the design of graphics processors, which consists of several crystals
Modern GPUs are among the most complex semiconductor solutions in the world. Suffice it to say that the GPU V100 contains 21.1 billion transistors spread over an area of 815 mm2!
Needless to say, infinitely increase the complexity of the GPU will not work. Specialists Nvidia already thinking about how to solve this problem in the foreseeable future. And they believe that one option is the use of several individual crystals to create a single GPU (Multi-Chip-Module GPU; MCM-GPU). That is talking about the same principles that are used in some modern processors with lots of cores.
On the project the specialists of Nvidia worked together with scientists from the University of Arizona, the University of Texas and the centre of Barcelona Supercomputing Center. The researchers suggest mnogoshipovyh layout as a way to circumvent Moore’s law.
Experts suggest that to «break» the GPU on a fairly simple basic module GPU Module (GPM) and using a high speed interconnect connection, to form the final graphics processor.
Of course, the experts used a software emulator to understand how appropriate this approach. In particular, the virtual Multi-Chip-Module GPU 256 multiprocessors (Streaming Multiprocessor; SM). He was 45.5% faster than the largest to date, monolithic GPU with 128 multiprocessors. In addition, it is 26.8% faster configuration of two graphics cards with the same 256 multiprocessors. If to speak about a hypothetical monolithic GPU with 256 blocks of SM, create that today it is impossible, the difference in performance would be no more than 10%.
Of course, all this is research, and not even the most Nvidia and its specialists together with scientists of various universities. However, in the not very distant future we may expect such graphics processors.