Intel today is announcing and making available to the broader research community an 8 million neuron system comprising 64 of its Loihi neuromorphic chips, codenamed Pohoiki Beach. Intel is aiming to scale Loihi further to 100 million neurons and 100 billion synapses by the end of the year. This marks a new milestone towards the eventual commercialization of the technology, as Intel has stated for the first time explicitly.
As a brain-inspired neuromorphic chip, Loihi is one of the rare designs that is not based on the von Neumann computing model. Its other unusual feature is that it is an asynchronous circuit, meaning that it does not have a global clock signal. More specifically, it implements an asynchronous spiking neural network (SNN).
Other specs of the 128-core 14nm chip include 130,000 neurons and a thousand times as many synapses, three Quark x86 cores, 2.1 billion transistors and a die size of 60mm2. The Loihi programming toolchain includes a Python API. Intel says it is up to 1,000 faster and 10,000 times more efficient than CPUs for target applications that include sparse coding, graph search, path planning, simultaneous localization and mapping (SLAM) and constraint-satisfaction problems.
While the previous, USB form factor Kapoho Bay was still a single chip, Intel’s roadmap consisted of scaling up Loihi to a multi-chip system. Theoretically, Loihi can be scaled up to 16,384 chips – over two billion neurons (humans have ~86 billion neurons). While Intel has no plans to do such, in May 2018 Intel said that it saw a path to reach 100 million neurons in a single system by the end of 2019. If we assume that means a 768-chip system; it would contain over 1.5 trillion transistors.
With the current announcement of Pohoiki Beach today, Intel has scaled up Loihi to a 16-chiplet system, resulting in 8 million neurons, 3,840mm2 total die area, and 132 billion transistors. Pohoiki Beach is now available to the over 60 INRC ecosystem partners.
Intel has also provided several examples of how researchers use Loihi to prove the viability of neuromorphic computing. Projects include providing adaptation capabilities to a prosthetic leg, object tracking, automating a foosball table, learning to control an inverted pendulum and inferring tactile input to the skin of a robot.
Providing some benchmark numbers, Chris Eliasmith, co-CEO of Applied Brain Research and professor at the University of Waterloo, said: “With the Loihi chip we've been able to demonstrate 109x lower power consumption running a real-time deep learning benchmark compared to a GPU, and 5x lower power consumption compared to specialized IoT inference hardware. Even better, as we scale the network up by 50x, Loihi maintains real-time performance results and uses only 30 percent more power, whereas the IoT hardware uses 500 percent more power and is no longer real-time.”
The system that Intel will use to reach 100 million neurons by the end of the year is codenamed Pohoiki Springs. The goal of these research systems is to clarify what application areas neuromorphic computing is most suitable for, to quantify how large the achievable gains are and pave the way for commercial use.
Intel announced Loihi in September 2017 and has since passed several milestones towards eventual commercialization. The chip came back from the fab in November 2017 and ran a test program. In March 2018, Intel created the Intel Neuromorphic Research Community (INRC) to spur innovation in neuromorphic computing with research into areas such as neuromorphic theory, programming models, SNN algorithms and applications to solve real-world problems. The company also provided community member access to a Loihi test system via the cloud. Last October, Intel held its inaugural INRC symposium, announcing fifty projects from over a dozen universities, and also made the chip available to its INRC partners via the USB form factor device Kapoho Bay.
