"Works like a brain": a new IBM processor will help the US defeat China in the Microchip war

NorthPole chips have already interested the military as a new step towards improving computer vision, real -time data processing and creating artificial intelligence. IBM has introduced a unique processor prototype for NorthPole artificial intelligence, which can accelerate the development of the field of view, machine vision segments and large data in real -time data. This was reported on the company's official website.

Although the processor is created on a 12-nm technological process and contains 22 billion transistors in an area of ​​800 square millimeters, it has 256 nuclei and can perform 2048 operations on the nucleus per cycle with 8-bit accuracy, with the ability to double and twice the number of operations with 4- 4- bit and 2-bit accuracy, respectively. "This is a whole network on a chip," says Dharmenndra Modchi, author of a project who has worked on a similar concept of 14 years.

NorthPole is much more efficient than usual 12-Hm processors, and 25 times more energy efficient when it comes to the number of interpreted frames on joules of the necessary energy. According to Modkhi, in a special test of the RESNET-50, the NorthPole processor exceeds all the main common architecture-even those that use more advanced technological processes, such as a graphic processor implemented by 4-Hm technical process.

According to the Editor from the Science of the Defense One Patricka Taker, the IBM NorthPole processor can create more intelligent, more efficient and independent devices that can even help the US win the microchip war against China. The processor design prototype will be useful for soldiers with additional reality headsets that control drones or land.

The processor will help to perceive and interpret the outside world, take a wider range of data, including audio, optical, infrared, infrared, hydro -location and other sensors. NorthPole will help you perform segmentation much faster (distinguish people in the image) and classify sounds for you directly on the battlefield and without connecting to the Internet. It can also make a revolution in unmanned vehicles not only for the army but also in the commercial sector.

In addition, the new processor, created on the human brain, is paving the way for another type of AI, which will not rely on large cloud services and data processing companies, such as Amazon or Google. Unlike traditional chips that separate memory from processing chains, the NorthPole processor combines both - like synapses in the brain, which store and process information on the basis of their connection with other neurons.

In the article in Science, IBM researchers call it "an architecture of a neural output that erases this limit, eliminating outsotal memory, intertwining the calculations with a memory on a crystal that has an outward appearance of active memory. " Today's computers have at least two characteristics that limit the development of AI. First, they need a lot of energy.

Our brain, which consumes only 12 watts of energy, can store and extract the information necessary for a detailed conversation, simultaneously absorbing, properly interpreting and deciding on a huge amount of sensory data needed to drive a car. But a desktop computer requires 175 watts only to process units and zeros in an orderly spreadsheet. This is one of the reasons why computer vision in cars and drones absorbs a lot of energy because of the autonomy of devices.

This energy inefficiency is one of the reasons why many of today's artificial intelligence tools depend on the huge corporate cloud farms that consume enough energy to ensure a small city's power supply, Patrick Taer said. The second problem is that we are approaching the atomic limit of the number of transistors, which can be placed in the processor without harming its size, according to the law of Mura.

Increasing the number of transistors will inevitably increase the area of ​​the processor, and technologies need small chips. NorthPole processor prototype can help solve both problems. The Defense One portal editor is convinced that IBM will probably be interested in the Pentagon. Moreover, the company has been working on such neuromorphic chips for over ten years with the financial support of the military under the Synapse Darpa program.

Over the past four years, the Ministry of Defense has invested $ 90 million in the Meinard Holdey, Assistant Defense Minister for Critical Technologies, various military laboratories, including the Air Force Research Laboratory and the National Laboratory of Sandia, already studying the possibility of using the prototype. Holdey said NorthPole was already competing with the most advanced chips from Asia, and it is expected that future versions will be even more effective.