NVIDIA intends to circumvent Moore's Law, developing a graphics chip based on several cores or modules, thus allowing it to continue increasing power even though the manufacturing process cannot be lowered.

Moore's Law is a problem that is about to be fulfilled and that is that it is increasingly difficult to lower the architecture of the processors. We have seen the problems of Intel to reduce from 14nm to 10nm and things look complex to reach 7nm. NVIDIA has had an interesting idea to skirt Moore's Law and is to develop the 'Mutli-Chip-Module (MCM), which is really nothing more than a set of processors integrated into the same system, something like integrating a processor and a graph in a DIE.

NVIDIA would be developing, together with experts from Arizona State University, the University of Texas and the Barcelona Supercomputing Center, the development of these MCMs for graphics cards, achieving significant performance improvements. The problem of reducing the dimensions of the transistors is a reality and manufacturers are forced to stretch the current architectures, until the next stage is ready, so they are trying to find other solutions that allow to continue increasing power and improve efficiency. , but from other points of view or in more ingenious ways.

What these experts are looking for is the development of graphic card improvements, integrating different GPU modules in the same package. The idea is more or less similar to the integration of a processor and a graphics in the same package, creating in this case a system with several GPUs, which improve performance and power. NVIDIA is the promoter of this idea and is releasing its products to see if systems can be developed that allow improved performance with optimal functionality. This MCM system for graphics cards would be achieving a significant increase in SM and the use of GPU applications.

They have analyzed the possibilities of 256SMs MCM-GPUs in a single system and it seems that they are very happy with the potential of this technology. By using simple GPM-based building blocks with advanced interconnects for this 256SM chip, it "achieves a 45.5% increase in speed within the GPU over 128SM systems," the researchers have said. Other tests show a 26.8% improvement better than multi-GPU systems and a 10% performance improvement on a hypothetical monolithic GPU that cannot be built on the basis of a roadmap based on today's technology.

This MCM GPU system will not arrive shortly, it is a technology that will take a long time to arrive, at least two generations or perhaps two architectures, since at the moment NVIDIA and its partner in the manufacture of chips, TSMC, are managing to lower the manufacturing process quite satisfactorily.

Source: Hexus