The miniaturization of processors is almost constant for decades, billions of transistors now coexist within a few millimeters of space and the solutions made with the production process to 10nm are already on the market (mobile). But the room for maneuver is running out, well aware of the chipmaker, which already talk of the next generation of 7nm but at the same time confirmed the need for a breakthrough in the industry. It will not be easy to go beyond this limit, for this study, there are new approaches, like the one just unveiled by DARPA.
For the first time you want to make a completely different processor, which no longer follows the rules of architecture von Neumann. DARPA has put on flat funding for $ 80 million, money that will be used to advance the development of the HIVE ( Hierarchical Identify Verify Exploit ) with thick partners like Intel, Qualcomm , Northrop Grumman, Georgia Tech and Pacific Northwest National Laboratory. These last two will concentrate on the development of software tools for the processor, defined as the first GAP (graph analytic processor).
HIVE is a non-von Neumann because the concentration of its data and its ability to simultaneously work with different processes in different areas of memory – said Trung Tran, program manager at DARPA
The new graph analytic processors do not yet exist, the whole thing was only theorized and announced partnership with this program is aimed right at the physical realization of this component. Change the architecture and also the manner in which it operates, no longer exist CPU and GPU: the processed elements are then located on a universal memory, and a second memory – a new concept – can thus access in these very high-speed points ( terabytes per second) for the execution of processes.
Keystone of the new processors will also be energy efficiency, today the new arithmetic-processing-unit (APU) consumes 1,000 times less than what is happening today on the supercomputer. Intel, Qualcomm and other partners will be entitled to market the processors and the memories made to create HIVE , thus not subject to any copyright.
This new approach is completely experimental, we must emphasize this, we will not see for sure HIVE processors on the next Christmas notebook. DARPA will be awarded $ 50 million to those making the best hardware design, with the sole condition that the same winning company and so forth as many for the realization of the processors. If the project is successful we will enter into a new technological era, with the possibility of analysis of large flows of unpublished data at rates, an idea that goes hand in hand with the explosion of the universe Internet of Things.