IBM introduces power-efficient NorthPole machine learning processor

IBM today be familiar with NorthPole is an artificial intelligence chip designed in-house that is significantly more power efficient than competing processors.

NorthPole was developed over an eight-year period by staff at the company’s research lab in Almaden, California. The chip builds on his previous IBM research project that produced an AI-optimized processor called TrueNorth. According to IBM, NorthPole is approximately 4,000 times faster.

The newly debuted chip is made up of 22 billion transistors manufactured using a 12-nanometer process. According to IBM, these transistors are organized into 256 cores, each of which can perform 2,048 calculations per clock cycle with 8-bit precision. This means that calculations are performed on relatively small numbers that occupy up to 8 bits of space.

Each of NorthPole’s 256 cores contains an integrated memory pool that stores information used in computations. Because the chip’s memory circuits and core are close to each other, data can be moved between them in a short amount of time. This reduces processing delays associated with data movement and improves performance.

“Architecturally, NorthPole blurs the line between compute and memory,” said Dharmendra Modha, principal scientist for Brain Inspiration Computing at IBM. “At the level of the individual core, NorthPole appears as memory close to the compute; outside the chip, at the input/output level, it appears as active memory.”

Another benefit of the memory circuitry built into NorthPole’s core is increased energy efficiency. Moving data between different parts of the chip not only slows down processing, but also consumes power. NorthPole’s architecture saves energy by minimizing the distance data has to travel between memory circuits and the core.

IBM measured the chip’s power consumption by running an open-source image recognition model called ResNet-50. This model is often used to compare the capabilities of different AI processors. According to IBM, NorthPole ran his ResNet-50 with 25 times more power efficiency than graphics cards and central processing units based on 12-nanometer and 14-nanometer nodes, respectively.

NorthPole’s power efficiency is constrained by the fact that it is manufactured using a 12-nanometer node, several generations behind the latest 3-nanometer processes available to chipmakers. However, processors can be upgraded to new manufacturing technologies. IBM researchers reportedly It is estimated that such upgrades could allow NorthPole to be 25 times more energy efficient than competing chips based on “current designs.”

A processor’s power consumption directly affects the amount of heat it generates. IBM says NorthPole’s energy efficiency limits the maximum temperature, meaning the chip only requires simple external cooling to operate. As a result, it could be implemented in space-constrained systems, such as self-driving cars, where space for cooling equipment is limited.

Image: IBM