Nvidia’s attempt to keep Moore’s Law alive
A team of Nvidia researchers at Arizona State University, the University of Texas, and the Barcelona Supercomputing Center have published a paper outlining ways to circumvent the problem of increasing density transistor has been proposed.
According to today’s technologies, a ceiling has been defined for the maximum performance that can be achieved in the processor sector, which is very difficult to achieve and has a high cost. To avoid such a situation, Nvidia researchers have proposed to build GPU modules (GPM). These modules are integrated in one package using signaling technologies with high bandwidth and low energy consumption, so that it can be used for designing and making graphics processors with multi-chip module (MCM).
- Nvidia unveiled its next-generation graphics processor platform called Volta
- Nvidia announced the next two generations of Tegra named Logan and Parker
Based on the researchers’ findings from Nvidia’s proprietary simulator, MCM GPUs can greatly increase the number of stream processing units (SMs). Increasing the number of these units will increase the execution speed. By improving the basic GPM blocks and using advanced interconnects, the researchers have also simulated a chip with 256 stream processing units, which has increased the speed ratio by 45.5 percent compared to the largest GPU with 128 stream processing units. In addition, compared to the same multi-GPU design, their performance has increased by 26.8% and is only 10% less than a GPU with 256 stream processing units, which is impossible to make with today’s technologies.