The existence of the ``NorthPole'' chip developed by IBM over a total of 20 years is revealed, and the brain-inspired design can speed up AI

The existence of a new type of digital chip `` NorthPole '' that IBM Research Center in Aldeman, California has been researching for 20 years together with its predecessor chip has been revealed. According to development team member Dharmendra Moda, the basic structure of the computer chip was changed by taking inspiration from the way the brain calculates.

Neural inference at the frontier of energy, space, and time | Science

IBM Research's new NorthPole AI chip | IBM Research Blog

'Mind-blowing' IBM chip speeds up AI

IBM has made a new, highly efficient AI processor | Ars Technica

Computer chips have the same basic von Neumann structure that has been around since the advent of semiconductors, with separate processing units and separate memories for storing the information being processed. While this has scaled well for decades with a simple structure, it has instead created a 'Neumann bottleneck' in which data is continually moved back and forth between memory, processors, and other devices. It took time and energy.

Meanwhile, in 2014, IBM's research team led by Moda developed a chip called ``TrueNorth'' that was inspired by the brain.

IBM's ultra-power-saving chip 'TrueNorth' imitating a 'brain' steadily evolves and reaches the brain level of a mouse - GIGAZINE

For another eight years, Moda and his colleagues have been researching and developing a new type of digital AI chip, NorthPole, for neural inference.

When we tested the popular image recognition model 'ResNet-50' and the object detection model 'YOLOv4' on a prototype device equipped with NorthPole, we found that the energy efficiency was average in terms of the number of joules of power required per frame. It is said to be 25 times better than 12nm process GPUs and 14nm process GPUs, and has the lowest delay of any chip.

The biggest difference between NorthPole and conventional products is that each of its 256 cores is equipped with its own memory, which allows it to perform AI inference faster than any existing chip.

'Architecturally, NorthPole blurs the line between compute and memory,' Moda said. 'At the level of the individual core, NorthPole looks like memory-near-compute; From the outside and at the input/output level, it appears to be active memory.'

However, NorthPole's biggest advantage is also its limitation, which is that it can only be easily drawn from onboard memory. Although each core is equipped with sufficient memory, this memory is not intended for operation as a ``jack of all trades,'' and ``NorthPole is only for inference,'' Moda says.

in Hardware, Posted by logc_nt