Intel Unveils Neuromorphic, Self-Learning Chip Codenamed Loihi

There's been a huge surge of interest in topics like AI, machine learning, and deep learning over the last few years. Thus far, we've seen much of the market flow towards either GPUs (almost entirely Nvidia, though AMD might be tipping a toe into those waters), or to specialized architectures designed by specific companies. Google has TensorFlow, Fujitsu is working on its own platform, Microsoft uses FPGAs to accelerate web searches, and multiple companies are designing self-driving car hardware around various commercial solutions. Intel, in contrast, has been a smaller player. While it owns Movidius and that company's Myriad processors, it hasn't commanded the same mind-share as some of its competitors.

That may change in the future, if Intel's latest AI bet takes off. The company has announced(Opens in a new window) a new neuromorphic chip, codenamed Loihi, designed for AI and deep learning workloads. Intel's Dr. Michael Mayberry claims that Loihi does not need to be trained in the traditional way and that it takes a new approach to this type of computing by using asynchronous spiking. Unlike a transistor, neurons do not constantly flip back and forth between a 0 and a 1. They trigger when signal thresholds are reached, and continue to fire so long as the number of spikes exceeds a given threshold. The strength of a muscle flex, for example, is based on the average number of spikes the muscle receives over a given unit of time.

The brain’s neural networks relay information with pulses or spikes, modulate the synaptic strengths or weight of the interconnections based on timing of these spikes, and store these changes locally at the interconnections. Intelligent behaviors emerge from the cooperative and competitive interactions between multiple regions within the brain’s neural networks and its environment.

While neural spike models have been a useful way to study and understand biological cells, they have yet to be deployed as solutions for real-world computing or engineering problems. There's tremendous potential for cutting-edge applications, but it's less clear how these capabilities will be practically deployed.

[embed width="640" height="360"]https://www.youtube.com/embed/EgCRwZw4p8c[/embed]

Intel claims that Loihi is up to a million times faster than other "typical" spiking neural nets when solving MNIST digit recognition problems, though it doesn't say what those typical nets consist of, or how they are constructed. It also claims that Loihi is much more efficient when used for convolutional neural networks or deep learning tasks.

Each neuron on the chip is capable of communicating with thousands of other neurons, while each neuromorphic "core" includes what Intel is calling a "learning engine." The total number of neurons onboard is 130,000, with 130 million synapses. That's markedly less than IBM's TrueNorth, which debuted three years ago with one million programmable neurons and 256 million synapses across 4,096 neurosynaptic cores.

We must stress, however, the two chips can't really be compared until we know more about Loihi's capabilities. Neural network performance isn't simply a function of the total number of neurons or synapses on a processor, and Intel hasn't revealed nearly enough information to allow for an apples-to-apples comparison of the two solutions. Intel clearly sees Loihi as part of a stable of products that range from the HPC-focused Xeon Phi, to the deep learning products it bought when it acquired Nervana, to its own FPGAs, to low-power Movidius solutions. Long-term, the company wants to field deep learning and AI resources that can stretch to cover any market segment or available power envelope.