Among the post-Moore’s Law chip architectures entering the test phase are neuromorphic chip designs that go beyond today’s CPUs and programmable logic chips to achieve near-term goals such as real-time learning and adaptive control.
Intel Labs’ entry into the neuromorphic sweepstakes is a demonstrator dubbed Loihi, which was turned over to an industrial partner earlier this year for testing on an adaptive robotic controller. The test chip is among a number of approaches designed to mimic more directly how the human brain works. “We have some data coming out soon in December regarding the efficiency of this [Loihi test] chip,” Gabriella Cruz Thompson of Intel Labs told the IEEE Rebooting Computing conference earlier this month.
Meanwhile, Intel Labs has been scaling deployment of Loihi as the chipmaker expands its neuromorphic computing research beyond integrating sensors and actuators to include new algorithms, applications and programming models. The robotic arm used a test board with four Loihi chips. That was expanded to an Arria 10 FPGA expansion board that can be scaled up to 32 neuromorphic chips and accessed via cloud connections.
Intel released the test chip in a USB form factor to university researchers this fall, and expects by the end of this year to release remote access to a platform called Pohoiki Springs incorporating 768 Loihi neuromorphic chips. That’s the equivalent of more than 100 million neurons in a single stack, Cruz Thompson said.
The chipmaker is providing access to or providing actual hardware to industrial startups and a list of university researchers in the U.S. and Europe to test the Loihi chip and “stretch its legs,” she added. Among the goals is developing a neuromorphic chip ecosystem as Intel and other chip makers transition away from standard CMOS silicon architectures.
“So far, we are getting really good feedback” on the Loihi test chip, Cruz Thompson told the computing conference. “Part of the work [testers] are doing is making sure that we have the entire software stack optimized and that we’ve figure out what are the best applications for this Loihi test chip.”
Intel expects to release more data on the results of its Loihi testing early next year.
The challenge for Intel and other neuromorphic researchers is transforming test chips and demonstrations into working computers that seek to emulate human brain functions.
For its part, Intel stressed that its neuromorphic research is part of a broader effort to connect devices to humans and manage the data those interactions generate in the cloud. “We’re working on devices that humans can interact with easily and we are working on all the data that then gets collected as the cloud and ideally making it an efficient way of managing the data, not just in the cloud but also at the device level,” Cruz Thompson added.