Many artificial intelligence systems rely on neural networks, but they don’t work in the same way that human and animal brains do. Intel has been working with computers that think like a brain for years, and their Loihi neuromorphic technology has produced some spectacular, if strange, results. Loihi is having its first update right now, and it’s a huge one. The team packed up to eight times as many artificial neurons into a chip half the size of Loihi using an Intel 4 manufacturing technique that isn’t currently available for commercial processors. According to Mike Davies, director of Intel’s neuromorphic computing unit, this, along with a slew of other enhancements inspired by the past few years of experimentation, makes the Loihi 2 quicker and more versatile.
Unlike conventional AI’s artificial neurons, which store information as weights that indicate the strength of connections between neurons, Loihi’s neurons carry information in the timing of digitally-represented spikes, which is more similar to what happens in your brain. Because these pulses drive neural computing, there’s no need for a central clock to keep things in sync. When there is no event to observe, much of the chip will be idle, saving power.
Loihi systems have been used by more than 250 research partners for tasks such as piloting drones or robot arms, improving railway schedules, searching databases, and learning to recognize different aromas. “The results have been very promising,” Davies says. There were increases in energy efficiency and the amount of data required for the system to learn that were “orders of magnitude” in some cases. (The results will be published in the IEEE Proceedings in May 2021.)
However, these tests revealed a set of constraints that Davies and his team wished to solve in the following generation. For one reason, Davies claims that Loihi’s neural network model wasn’t flexible enough to achieve what Intel and its partners desired. They also discovered that Loihi’s calculations were limited by the binary spike model of brain activity, in which just the timing information was communicated from one neuron to the next. Congestion between many Loihi processors and difficulties integrating the system with conventional computers were also obstacles.
A close-up of a gold rectangular chip with closely spaced dots covering its surface. According to Davies, INTEL Loihi 2 addresses these concerns utilizing the same basic architecture, but with a set of circuits that were “redesigned from the ground up.” The following is a rundown of what’s new:
To improve precision, new spike messages have been added: Only temporal information was contained in the spike signals in the original-flavor Loihi. A binary-value spike message is a message that contains either yes or no information. Spikes with both timing and magnitude parameters are allowed in Loihi 2 with little energy or performance impact. “If we have this magnitude that can be provided alongside” the timing information, we can solve the same problems with fewer resources,” Davies says.
Improved programmability: The old Loihi was built around a specific spiking neural network (SNN) model. The Loihi 2 neuromorphic cores now support arithmetic, comparison, program control flow, and other activities, allowing the chip to run a larger number of SNN models.
Loihi essentially supported a single set of brain-inspired learning rules. Loihi 2 alters this in such a way that it can now conduct some of the most cutting-edge learning methods, as well as a close approximation of the deep learning backpropagation technique. Because of the modification, algorithms that could only be used as a proof-of-concept on Loihi 2 may now be scaled up, allowing it to learn faster.
More of everything: As previously stated, the Loihi 2 is constructed utilizing a manufacturing technique so advanced that the business has yet to use its own commercial chips. That’s a million neurons per chip against 125,000 for Loihi. Furthermore, depending on the type of network running, the circuits and memory that implement those neurons and their memory were optimized, resulting in between 2x and 160x more resources for neuromorphic computing.
Faster circuits: The revamped circuits in Loihi 2 result in a doubling of processing speeds while updating the state of neurons, a five-fold increase in synaptic operations, and a 10-fold increase in spike generation speed. Overall, the chip is now 5000 times faster than organic neurons at processing neuromorphic networks.
New chip interfaces: Loihi 2 chips offer 4x quicker asynchronous chip-to-chip signaling, as well as a 10-fold reduction in interchip bandwidth requirements. It also contains an Ethernet interface and one for developing event-based sensors, such as Prophesee’s camera chips, and is ready to handle communications in a 3D chip-stacking structure.
Software is the key to getting any use out of new chip architectures, as is the case with most new semiconductor architectures. “Software is still holding the field back,” Davies argues. “In the deep learning field, there hasn’t been the establishment of a single software framework.”
“We’re now delivering something ourselves since the emergence of a single framework hasn’t happened,” he explains. Lava, a new software framework, tries to provide a single platform that supports all of the preceding three-and-a-half years of research initiatives. Lava, not simply Loihi or Loihi 2, is an open-source framework that supports systems that do event-based, asynchronous message passing.
According to Davies, there are no intentions to market Loihi 2. “We’re still going to offer this as a research chip to research partners.” According to Davies, Loihi’s technology will likely first appear as an acceleration core on a system-on-chip running a specific algorithm, rather than as a general-purpose processor.
Even if Intel isn’t ready to make a profit from neuromorphic processors, that doesn’t imply others aren’t. In August, Sydney-based Brainchip received its first shipment of finalized event-based neural processor chips, with the goal of assisting clients in developing low-power systems that benefit from incremental and one-shot learning.
