On April 17th, Intel revealed the Hala Point, a large-scale neuromorphic system designed to advance cutting-edge research in the field of brain-inspired artificial intelligence (AI) and address the current challenges of AI in efficiency and sustainability. Built upon Intel's Loihi 2 neuromorphic processor, the system aims to significantly enhance the capabilities of AI models. Improving upon Intel's first-generation large-scale research system, Pohoiki Springs, Hala Point boasts an enhanced architecture supporting 1.15 billion neurons and 128 billion synapses distributed across 140,544 neuromorphic processing cores, marking a more than tenfold increase in neuron capacity and a twelvefold increase in performance.
Mike Davies, Director of the Neuromorphic Computing Lab at Intel Research, stated, "The computational cost of AI models is continuously rising, and the industry needs novel approaches to scalable computing. To address this, Intel has developed Hala Point, integrating efficient deep learning with innovative brain-like continual learning and optimization capabilities. We hope that research using Hala Point will lead to breakthroughs in the efficiency and adaptability of large-scale AI technologies."
Hala Point, based on the Loihi 2 neuromorphic processor, comprises 1,152 Loihi 2 processors packaged within a six-rack data center enclosure, roughly the size of a microwave oven. This system supports up to 1.15 billion neurons and 128 billion synapses distributed across 140,544 neuromorphic processing cores, with a maximum power consumption of only 2600 watts. Additionally, Hala Point includes over 2300 embedded x86 processors for auxiliary computation.
Featuring an integrated architecture of processors, memory, and communication channels, Hala Point achieves memory bandwidth of 16 petabytes per second, core-to-core communication bandwidth of 3.5 petabytes per second, and chip-to-chip communication bandwidth of 5 terabytes per second. The system is capable of processing over 38 quadrillion 8-bit synaptic operations and over 24 quadrillion neuron operations per second.
Notably, when employed for biologically plausible spiking neural network models, Hala Point can simulate its entire 1.15 billion neurons at speeds up to 20 times faster than real-time and achieve speeds up to 200 times faster than the human brain when simulating lower neuron counts. While Hala Point is not intended for neuroscience modeling, its neuron capacity roughly matches that of an owl's or a macaque monkey's brain cortex.
The neural capacity of Hala Point is equivalent to that of an owl's brain. In mainstream AI workloads, it demonstrates remarkable computational efficiency. Research indicates that when running traditional deep neural networks, the system can perform up to 20 quadrillion operations per second, with an 8-bit computational efficiency reaching 15 TOPS/W, surpassing architectures based on GPUs and CPUs. Hala Point holds promise for driving real-time continuous learning in various AI applications such as scientific research, engineering, logistics, intelligent city infrastructure management, Large Language Models (LLMs), and AI agents.
Currently, Hala Point remains a research prototype aimed at improving future commercial systems. Intel anticipates that this research will bring about practical technological breakthroughs, such as enabling Large Language Models to continuously learn from new data, thereby significantly reducing training energy consumption and enhancing sustainability in the widespread deployment of AI.
