Intel builds the world's largest neuromorphic system 'Hala Point' with a processing capacity of 20 quadrillion times per second

Neuromorphic computing is an attempt to build electronic circuits that mimic the systems of the human nervous system and brain. Intel has announced that it has built a large-scale neuromorphic system called ' Hala Point ' equipped with the neuromorphic processor '

https://www.intel.co.jp/content/www/jp/ja/newsroom/news/intel-builds-world-largest-neuromorphic-system.html

Intel and Sandia National Labs Roll Out 1.15B Neuron “Hala Point” Neuromorphic Research System
https://www.anandtech.com/show/21355/intel-and-sandia-national-laborators-roll-out-hala-point-neuromorphic-research-system

'Hala Point' is a neuromorphic system to be introduced at Sandia National Laboratories, and is said to have more than 10 times the number of neurons and up to 12 times the performance of Intel's first-generation large-scale research system ' Pohoiki Springs '. Hala Point features a computing performance of 20 quadrillion operations per second, or 20 peta-ops, and when running conventional deep neural networks, it is said to perform more than 15 trillion 8-bit operations per second per watt.

Hala Point integrates processing, memory, and communication channels into a single massively parallel fabric, delivering 16 petabytes per second (PB/s) of memory bandwidth, 3.5 PB/s of inter-core bandwidth, and 5 TB/s of inter-chip bandwidth across the entire system, capable of processing over 380 trillion 8-bit synapse operations per second and over 240 trillion neuron operations per second.

According to Intel, Hala Point applies a biologically inspired spiking neural network (SNN) model that allows it to run 1.15 billion neurons at full capacity 20 times faster than the human brain, and up to 200 times faster at lower capacity. Although Hala Point is not intended for neuroscience modeling, its neuronal capacity is roughly equivalent to that of an owl brain or the cortex of a capuchin monkey.

This Hala Point is equipped with the neuromorphic chip 'Loihi 2' announced by Intel in 2021. The Loihi2 chip is manufactured using the Intel 4 process, has an area of only ³¹ mm2 per chip, has approximately 2.3 billion transistors per chip, and can simulate 1 million neurons per chip.

Intel announces second-generation neuromorphic chip 'Loihi 2', adopts prototype 'Intel 4' process node - GIGAZINE

By installing 1,152 Loihi 2s, Hala Point can solve optimization problems by performing AI inference at about 50 times faster speeds while reducing power consumption by 1/100 compared to conventional CPU and GPU architectures. At the time of writing, Hala Point is only a research prototype, but in the future, it is expected to be applied to commercial systems and its functions will be expanded. Intel said, 'We hope that these prototyping will lead to practical breakthroughs such as LLM (large-scale language model) that continuously learns from new data.'

'Working with Hala Point, Sandia's research team has gained valuable capabilities to solve scientific modeling problems,' said Craig Vineyard, Hala Point team leader at Sandia National Laboratories. 'Working with Hala Point provides the Sandia team with the ability to perform computations, modeling, simulations and data analysis at the same time as advances in AI.'

Intel commented, 'Sandia's adoption of Hala Point marks the first deployment of a new large-scale neuromorphic research system that Intel plans to share with research collaborators. With further development, we expect applications of neuromorphic computing to address the power efficiency and latency constraints that currently prevent real-time deployment of AI capabilities in the real world.'