On April 18th, Intel officially unveiled its next-generation large-scale neuromorphic system, codenamed "Hala Point," designed for cutting-edge research in brain-inspired AI, aiming to enhance AI efficiency and sustainability.
Built upon the Intel Loihi 2 neuromorphic processor, the system represents a significant advancement over its predecessor, the large-scale neuromorphic research system "Pohoiki Springs." The architecture has been further refined, increasing the neuron capacity by over tenfold, reaching an unprecedented 1.15 billion neurons, roughly equivalent to the scale of an owl's or a macaque monkey's brain cortex. Additionally, performance has surged by up to 12 times.
The Loihi 2 processor, released as early as 2021, initially employed Intel's 4-nanometer technology. It integrates 23 billion transistors, six low-power x86 cores, and 128 neuromorphic cores. With a single chip boasting 1 million neurons and 120 million synapses, it surpasses the previous generation in scale by eightfold and boosts performance by tenfold.
Loihi 2 leverages various neuromorphic computing principles, such as asynchronous operation, event-based spiking neural networks (SNN), and dynamic sparse connectivity. Moreover, neurons can communicate directly with each other, bypassing the need for memory.
Especially for emerging edge workloads on a small scale, it achieves orders of magnitude improvement in efficiency, speed, and adaptability.
For instance, when handling AI inference loads and tackling optimization problems, Loihi 2's speed outpaces conventional CPU and GPU architectures by up to 50 times, while consuming only one percent of the power.
The Hala Point system takes the form of a six-rack data center enclosure, roughly the size of a microwave oven, housing 1152 Loihi 2 processors. It boasts a total of 140,544 neural morphology processing cores, 1.15 billion neurons, and 128 billion synapses, all while consuming a maximum of just 2600 watts.
Additionally, the system incorporates over 2300 embedded x86 processors for auxiliary computation tasks.
With a memory bandwidth of 16 PB/s (16,000 TB/s), inter-core communication bandwidth of 3.5 PB/s (3,500 TB/s), and chip-to-chip communication bandwidth of 5 TB/s, it can handle over 3.8 quadrillion 8-bit synaptic operations and over 2.4 quadrillion neuron operations per second.
Hala Point demonstrates exceptional computational efficiency in mainstream AI workloads. For instance, when running traditional deep neural networks, it achieves a staggering 20 petaflops (20PFlops), with an energy efficiency of 15 trillion operations per watt (15TOPS/W) for 8-bit operations, surpassing architectures based on GPUs and CPUs.
In applications utilizing spiking neural network models, Hala Point operates all of its 1.15 billion neurons at speeds up to 20 times faster than the human brain in real-time.
Notably, even with a lower number of neurons, its speed can exceed that of the human brain by up to 200 times!
Early research indicates that by leveraging sparse connections up to 10:1 and event-driven activity, Hala Point achieves an energy efficiency of up to 15TOPS/W when running deep neural networks, without the need for batch processing of input data.
The Hala Point system holds promise for driving real-time continuous learning in various AI applications across multiple domains such as scientific research, engineering, logistics, smart city infrastructure management, large language models, AI assistants, and more.
