On April 17th, local time, the chip manufacturing giant Intel announced it has created the world’s largest neuromorphic system, featuring a staggering 11.5 billion neurons and 128 trillion synapses, with speeds reaching up to 200 times faster than a human brain.
Nicknamed “Hala Point,” this monumental neuromorphic system was initially deployed at Sandia National Laboratories. It uses Intel’s Loihi 2 processors and is designed to support research in future brain-like artificial intelligence (AI) and address challenges related to the efficiency and sustainability of current AI technologies.
Hala Point advances Intel’s first-generation large-scale research system, Pohoiki Springs, with architectural improvements leading to over a tenfold increase in neuron capacity and an up to twelvefold improvement in performance.
Mike Davies, director of Intel’s Neuromorphic Computing Lab, commented: “The computational cost of today’s AI models is growing at an unsustainable pace. The industry needs scalable new methods. For this reason, we developed Hala Point, combining deep learning efficiency with novel, brain-like learning and optimization capabilities. We hope that research with Hala Point will enhance the efficiency and adaptability of large-scale AI technology.”
Hala Point: Integrating 1152 Loihi 2 Chips and Over 2300 Embedded x86 Processors
Under the hood, Hala Point is powered by the Loihi 2 neuromorphic processor, which applies principles of brain-like computing such as asynchronous, event-driven spiking neural networks (SNNs), integrated memory and computation, as well as sparse and dynamic connectivity to achieve orders of magnitude improvements in power efficiency and performance. Neurons communicate directly with each other, bypassing memory, and hence reducing overall power consumption.
△Loihi 2 Chip
Loihi 2 is crafted using Intel’s 4-process technology, has a core area of 31mm², and integrates 128 Neuromorphic Cores (each with a 192KB cache) and six low-power Intel x86 cores. Thanks to significant advancements in processing technology, Loihi 2 supports up to one million neurons, 7.8 times more than its predecessor, albeit with a slight reduction in the number of synapses to 1.2 trillion. It can allocate up to 4096 state variables, depending on the neural model requirements. These enhancements have enabled Loihi to process information ten times faster than the first-generation chip.
Hala Point houses 1152 Loihi 2 processors, built on Intel’s 4 node process, in a six-rack unit data center cabinet about the size of a microwave oven. The system supports up to 11.5 billion neurons and 128 trillion synapses distributed across 140,544 neuromorphic processing cores and maxes out at an energy consumption of 2,600 watts. It also includes over 2,300 embedded x86 processors for auxiliary computations.
Hala Point integrates processing, memory, and communication channels into a massively parallelized structure, delivering a total of 16 PB/s of memory bandwidth, 3.5 PB/s of core-to-core communication bandwidth, and 5 TB/s of interchip communication bandwidth. This system can handle over 3.8 quadrillion 8-bit synaptic operations per second and over 2.4 quadrillion neuron operations per second.
For biomimetic spiking neural network models, the Hala Point system can operate at its full 11.5 billion neuron capacity at speeds 20 times faster than the human brain, and at lower capacities, it can reach speeds up to 200 times faster. While not intended for neuroscience modeling, the neuron count of Hala Point roughly equates to that of an owl monkey’s cortical brain.
Loihi-based systems can perform AI inference and solve optimization problems with 100 times more energy efficiency and 50 times the speed compared to traditional CPU and GPU architectures. Early results suggest that by leveraging up to a 10:1 ratio of sparse connections and event-driven activity, Hala Point can achieve up to 15 TOPS/W of deep neural network efficiency without needing to batch input data – a common optimization in GPUs that can significantly delay the processing of real-time arriving data, such as video from cameras. While still under research, future continuously learning neuromorphic LLMs could conserve gigawatt-hours of energy by eliminating the need for periodic retraining on ever-growing datasets.
Intel expresses that Hala Point is the first large-scale neuromorphic system demonstrating cutting-edge computational efficiency on mainstream AI workloads. In traditional deep neural networks, Hala Point can support up to 20 petaops per second (20 trillion operations per second) with efficiency exceeding 15 TOPS/W (trillion 8-bit operations per second per watt), matching and surpassing levels achieved by GPU and CPU-based architectures. Hala Point’s unique capabilities can provide real-time continuous learning for future AI applications in areas such as problem-solving in science and engineering, logistics, smart city infrastructure management, large language models (LLMs), and AI agents.
The Role and Importance of Hala Point
Researchers at Sandia National Laboratories plan to use Hala Point for advanced brain-scale computational studies, focusing on solving computational and scientific modeling problems in the realms of device physics, computer architecture, computer science, and informatics.
“Collaborating with Hala Point has enhanced the ability of our Sandia team to tackle computational and scientific modeling issues. Using a system of this scale for research will allow us to keep pace with AI advancements in everything from business to defense to fundamental science,” said Craig Vineyard, team leader for the Hala Point team at Sandia National Laboratories.
Currently a research prototype, Hala Point is expected to improve the capabilities of future commercial systems. Intel anticipates that these learnings will bring tangible advancements, such as LLMs that continuously learn from new data, hopefully alleviating the unsustainable training burden brought by widespread AI deployment.
The recent trend of extending deep learning models to trillions of parameters has exposed the daunting sustainability challenges faced by AI and underscored the need for innovation at the lowest hardware architecture level. Neuromorphic computing is an innovative approach that draws insights from neuroscience to integrate memory and computation with high-level parallelism, minimizing data movement. At this month’s International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Loihi 2 showcased efficiency, speed, and adaptability improvements by several orders of magnitude for emerging small-scale edge workloads.
Building on its predecessor, Pohoiki Springs, Hala Point has seen significant improvements and now brings neuromorphic performance and efficiency enhancements to mainstream traditional deep learning models, particularly those processing real-time workloads like video, voice, and wireless communications. For instance, Ericsson Research is applying Loihi 2 to optimize the efficiency of telecom infrastructure, as highlighted at this year’s World Mobile Congress.
As described by Intel, moving forward, Hala Point’s delivery to Sandia National Laboratories marks the first deployment in a new series of large-scale neuromorphic research systems that Intel plans to share with its research collaborators. Further development will enable neuromorphic computing applications to overcome the power and latency limitations that currently constrain real-world, real-time deployment of AI capabilities.
In collaboration with an ecosystem comprising over 200 members of the Intel Neuromorphic Research Community (INRC), including leading global academic groups, government labs, research institutions, and companies, Intel aims to push the boundaries of brain-like AI and evolve this technology from a research prototype to an industry-leading commercial product in the coming years.
Editor: Xin Zhi Xun – Lang Ke Jian