IBM and the US Air Force are developing a new generation neuromorphous supercomputer

Back in 2014, IBM introduced a new type of chip, which is designed to implement neural networks. Then the developers said that the chip contains 1 million neurons and 256 million synapses, which, with some reservations, is similar to the architecture of the neocortex. Such a processor can be used to solve problems where high and ultra-high performance is required. For example, it can be used to classify objects in a video stream (this is already being done) in real time.

Somewhat later, the Livermore National Laboratory (LLNL) announced the creation of a productive computer, the principle of which is similar to the principle of the human brain. The system includes 16 million neurons and 4 billion synapses. In the computer, which was developed by the laboratory, only 16 TrueNorth chips, and the system itself was named the IBM Neuromorphic System. What can such systems be used for?

“Neuromorphic computations open up quite impressive new possibilities, and, in general, this area is in line with our vision of the future of high-performance computer systems,” said Jim Brase, one of the project leaders. “The potential of neuromorphic computing and machine intelligence will help change the way we conduct research,” he continues. According to experts, on the basis of such chips, you can create systems an order of magnitude more productive than the most powerful computers of today.
')
IBM is trying to improve the new technology. Now she is involved in a joint project with the US Air Force. Partners are going to create a new supercomputer, which will include 64 million neurons and 16 billion synapses. In this case, it will consume only 10 W of energy, that is, it can be connected to a conventional power grid.


According to the developers, the new computer (its name is TrueNorth Neurosynaptic System) will allow working with several data sources simultaneously (video, images, audio and text). Now one neuromorphic processor consists of 5.4 million transistors that make up the core, there are 4096 of them in total. The cores, in turn, create an array of 1 million digital neurons that interact with each other through 256 million electrical synapses.

Why is all this military? The fact is that the US Air Force, namely the Air Force Research Lab (AFRL), uses the capabilities of the processor to identify military and civilian vehicles during airborne radar. The military claim that the chip works no worse than a powerful military computer. But the energy consumed is twenty times less. In the future, with an increase in chip performance, the efficiency of the systems will be much higher.



TrueNorth chip is made on 28 nm process technology. It contains 5.4 billion transistors and is a neuromorphic system with the following characteristics:

• one million emulated "neurons"
• 256 million emulated connections between neurons - “synapses”
• about 400 megabytes of SRAM memory (approximately 50 megabytes)



But not only the military are interested in such processors. Samsung has previously created a TrueNorth based vision system. Its principle of operation differs from that of ordinary cameras. And this system provides video stream processing with a speed of 2 thousand frames per second and higher. The most modern cameras, this figure does not exceed 120 fps.

But all this is only the initial stage. After all, such processors allow you to create self-learning systems that can come as close as possible to the architecture of the human brain. TrueNorth features are being explored (and expanded by) more than 40 research organizations, including the most famous.