Faster, power-conserving chip seen as future for AI

Universities and electronics manufacturers are working feverishly to develop a new type of high-performance computer chip that also conserves power for use in robots, cars and drones.

The field programmable gate array, or FPGA as the new chip is known, is being used to fuel, for example, AI robots and big data that require lightning-quick processing power. The FPGA is a core part similar to the central processing unit in PCs, but used in combination with CPUs in devices.

There are various types of computer chips, from general-purpose types to those specialized in image processing. However, FPGA is characterized by its ability to reconnect its wiring to create the optimum circuit for the computations it is required to perform, and its processing speed is extremely fast.

These features make it ideal for applications that call for real-time processing, such as robots.

With the onset of the internet of things, when almost anything will be virtually connected, devices will have to be controlled swiftly. Scientists are developing FPGAs that are highly efficient and consume less power, for use in cars, home appliances, robots and portable information devices.

At Tohoku University, Professor Takahiro Hanyu and his team have developed an FPGA that can operate with about 80% less power compared with previous models.

Electronics-maker NEC has developed an FPGA on which copper becomes solid and forms circuitry when electricity is applied.

Scientists expect that next-generation robots will operate on combinations of a motor and AI that drive joints so that each of these combinations will locally “learn” and control movements. NEC hopes the new FPGA will facilitate equipping robots with such functions.

Yuichiro Tsuchiya, a graduate student at the University of Tokyo, has developed software that makes use of deep learning, a core technology for AI, onto, for example, drones and home appliances.

Share on Pinterest
More share buttons
Share with your friends










Submit