Intel “Reverse Engineers the Brain” With Loihi 2 Neuromorphic Chip and Lava Framework
Packing one million neurons, the Loihi 2, along with an open-source software framework Lava, is pushing the limits of neuromorphic computing.
One of the major research areas in computing is neuromorphic computing, a field that aims to mimic the neural plasticity of the human brain.
Intel is one of the organizations at the forefront of this field, and today the company releases its Loihi 2, a second-generation neuromorphic computing chip, along with a new software framework for neuromorphic computing.
The Loihi 2 chip.
Mike Davies, director of Intel’s neuromorphic computing lab, debriefed All About Circuits on the details of this new brain-like chip and associated API.
Reverse Engineering the Brain
Billions of years of evolution have created the most powerful, energy-efficient computing machine known to man: the brain. For instance, while a cockatiel parrot's brain only requires 50 mW of power to perform somewhat complex tasks, like manipulating cups to drink or learning English words, an autonomous drone requires 18,000 mW of power to fly between known gates at 5.6 mph.
Electronic computing and intelligence are a far cry from what exists in nature.
Human-made artificial “intelligence," while in its infancy, lacks the plasticity of the human brain. Our artificial intelligence agents are unable to learn things online once deployed and are only taught to do singular tasks, like identifying a specific object. Compare this to the human brain, which can learn information online and can extrapolate what it has learned to other tasks.
Neuromorphic computing seeks to bolster learning by mimicking the brain.
The field of neuromorphic computing seeks to bridge this gap to create computing architectures that mimic the biological processes of the brain.
"What we’re trying to do here is effectively reverse engineer the brain,” Davies comments.
A Look at Loihi 1, the Forerunner of Loihi 2
To pursue this goal of neuromorphic computing, Intel released its first Loihi chip back in 2018—a device that represented “a new type of computer architecture that differs dramatically from conventional architectures,” according to Davies.
The Loihi processor incorporated a 128-core design based on a specialized architecture optimized for spiking neural networks (SNN) algorithms. Loihi is unlike standard computing architectures in two ways. First, all memory in the system is tightly embedded in the individual cores; there are no separate memory chips or cache hierarchies at play.
Architecture of the Loihi 1 chip, Intel’s first neuromorphic chip. Image used courtesy of WikiChip
Second, unlike standard computing, Loihi performs all of its computation in an event-based, asynchronous fashion.
“Information coding and processing are happening in a highly-sparse, asynchronous way—similar to that of the brain, where pulses of voltage (spikes) are transferring information between neurons," Davies explains. "In an event-based way, when there’s important computation to be performed, we activate the circuit and there is a pathway of spikes that trigger computation. It’s all asynchronous.”
Intel found that Loihi was able to support accelerated learning and perform difficult tasks at well under 1 W, which Intel estimates to be 10 times to 100 times less than CPU and GPU solutions.
Loihi 2 Scales Down the Size—But Not Power—of Loihi 1
Now, building off Loihi 1, Intel is releasing both its Loihi 2 chip and Lava, a new software framework to support neuromorphic computing.
Loihi 2 keeps the same architecture as Loihi 1 but implements it in a die size that is twice as small.
Loihi 2 leverages the pre-production version of Intel 4 to take nearly the same architecture as Loihi 1 but integrates it into a die that is twice as small. This scaling has importantly allowed for eight times more neurons, twice as many processors, four times more bandwidth per chip-to-chip link, and a decrease in chip-to-chip congestion by more than 10 times.
"Scaling has given us a chip with way more programmability. The neuron models were previously fixed-functioned," Davies says. "Now with Loihi 2, we are providing an instruction set . . . which allows users to program almost arbitrary neuron models.”
Lava: An Open-source API for Neuromorphic Computing
In addition to the Loihi 2 chip, Intel is releasing a new framework, Lava. This open-source software framework can be used to develop neuromorphic computing applications.
Intel says the new framework addresses a need for a common software framework in the neuromorphic research community in the same way that TensorFlow or Pytorch once did for machine learning. Intel hopes that Lava will allow researchers and application developers to build on each other’s progress and converge on a common set of tools, methods, and libraries.
All images used courtesy of Intel.
Neuromorphic Computing Research Heats Up
Intel hopes these two releases will help foster a community for neuromorphic research and provide people with the necessary tools to this end. Here are a few other recent advances in the neuromorphic computing space to check out.