Intel builds world’s largest neuromorphic system

Quantum computing is billed as a transformative computer architecture that’s capable of tackling difficult optimization problems and making AI faster and more efficient. But quantum computers can’t be scaled yet to the point where they can outperform even classical computers, and a full ecosystem of platforms, programming languages and applications is even farther away.

Meanwhile, another new technology is poised to make a much more immediate difference: neuromorphic computing.

Neuromorphic computing looks to redesign how computer chips are built by looking at human brains for inspiration. For example, our neurons handle both processing and memory storage, whereas in traditional computers the two are kept separate. Sending data back and forth takes time and energy.

In addition, neurons only fire when needed, reducing energy consumption even further. As a result, neuromorphic computing offers massive parallel computing capabilities far beyond traditional GPU architecture, says Omdia analyst Lian Jye Su. “In addition, it is better at energy consumption and efficiency.”

According to Gartner, neuromorphic computing is one of the technologies with the most potential to disrupt a broad cross-section of markets, as “a critical enabler,” however, it is still three to six years away from making an impact.

Intel has achieved a key milestone, however. Today, Intel announced the deployment of the world’s largest neuromorphic computer yet, deployed at Sandia National Laboratories.

The computer, which uses Intel’s Loihi 2 processor, is code named Hala Point, and it supports up to 20 quadrillion operations per second with an efficiency exceeding 15 trillion 8-bit operations per second per watt – all in a package about the size of a microwave oven. It supports up to 1.15 billion neurons and 128 billion synapses, or about the level of an owl’s brain.

According to Intel, this is the first large-scale neuromorphic system that surpasses the efficiency and performance of CPU- and GPU-based architectures for real-time AI workloads. Loihi-based systems can perform AI inference and solve optimization problems 50 times faster than CPU and GPU architectures, the company said, while using 100 times less energy.

And the technology is available now, for free, to enterprises interested in researching its potential, says Mike Davies, director of Intel’s Neuromorphic Computing Lab.

To get started, companies should first join the Intel Neuromorphic Research Community, whose members include GE, Hitachi, Airbus, Accenture, Logitech, as well as many research organizations and universities – more than 200 participants as of this writing. There is a waiting list, Davies says. But participation doesn’t cost anything, he adds.

“The only requirement is that they agree to share their results and findings so that we can continue improving the hardware,” Davies says. Membership includes free access to cloud-based neuromorphic computing resources, and, if the project is interesting enough, free on-site hardware, as well.

“Right now, there’s only one Hala Point, and Sandia has it,” he says. “But we are building more. And there are other systems that are not as big. We give accounts on Intel’s virtual cloud, and they log in and access the systems remotely.”

Intel was able to build a practical, usable, neuromorphic computer by sticking with traditional manufacturing technology and digital circuits, he says. Some alternate approaches, such as analog circuits, are more difficult to build.

Intel Corporation

But the Loihi 2 processor does use many core neuromorphic computing principles, including combining memory and processing. “We do really embrace all the architectural features that we find in the brain,” Davies says.

The system can even continue to learn in real time, he says. “That’s something that we see brains doing all the time.”

Traditional AI systems train on a particular data set and then don’t change once they’ve been trained. In Loihi 2, however, the communications between the neurons are configurable, meaning that they can change over time.

The way that this works is that an AI model is trained – by traditional means – then loaded into the neuromorphic computer. Each chip contains just a part of the full model. Then, when the model is used to analyze, say, streaming video, the chip already has the model weights in memory so it processes things quickly – and only if it is needed. “If one pixel changes, or one region of the image changes from frame to frame, we don’t recompute the entire image,” says Davies.

The original training does happen elsewhere, he admits. And while the neuromorphic computer can update specific weights over time, it’s not retraining the entire network from scratch.

This approach is particularly useful for edge computing, he says, and for processing streaming video, audio, or wireless signals. But it could also find a home in data centers and high-performance computing applications, he says.

“The best class of workloads that we found that work very well are solving optimization problems,” Davies says. “Things like finding the shortest path through a map or graph. Scheduling, logistics – these tend to run very well on the architecture.”

The fact that these use cases overlap with those of quantum computing was a surprise, he says. “But we have a billion-neuron system shipped today and running, instead of a couple of qubits.”

Intel isn’t the only player in this space. According to Omdia’s Su, a handful of vendors, including IBM, have developed neuromorphic chips for cloud AI compute, while companies like BrainChip and Prophesee are starting to offer neuromorphic chips for devices and edge applications.

However, there are several major hurdles to adoption, he adds. To start with, neuromorphic computing is based on event-based spikes, which requires a complete change in programming languages.

There are also very few event-driven AI models, Su adds. “At the moment, most of them are based on conventional neural networks that are designed for traditional computing architecture.”

Finally, these new programming languages and computing architectures aren’t compatible with existing technologies, he says. “The technology is too immature at the moment,” he says. “It is not backwardly compatible with legacy architecture. At the same time, the developer and software ecosystem are still very small with lack of tools and model choices.”