Adherence to Moore’s law means that monolithic graphics processors will eventually hit a performance ceiling and if you’re a company like NVIDIA that pushes the envelope on graphics performance, you have to avoid hitting that ceiling. No wonder the company is researching multi-ship-module GPUs which may help it keep Moore’s law relevant to graphics processors.
A paper has been published recently by researchers at NVIDIA, the University of Texas, Arizona State Univeristy, and the Barcelona Supercomputing Centre. The paper is a study on the methods for bypassing the deceleration in the advancement of transistor density.
In order to avoid hitting the performance ceiling that is inevitable for monolithin graphics processors, the researchers have proposed the manufacturing of basic GPU modules that will be integrated on a single package coupled with high bandwidth and power-efficient signaling technology. This should allow for multi-chip-module graphics processor designs.
NVIDIA’s in-house GPU simulator was used by the researchers to evaluate their designs. Their findings reveal that multi-chip-module graphics processors can help to increase the number of Streaming Multiprocessors.
The researchers simulated a 256 Streaming Microprocessors chip that hit a 45.5 percent speedup compared to the largest possible monolithic GPU with 128 SMs. The design performed 26.8 percent better than a discrete multi-GPU with a similar number of Streaming Microprocessors.
Whether or not one can be built on the existing technology roadmap is another matter entirely.