Eta Compute is trying to make chips that run on so little power that they act almost like neurons in the human brain, running artificial intelligence locally in sensors and other devices.
Now the start-up has hired Narayan Srinivasa, Intel's chief scientist in charge of neuromorphic computing, as its chief technology officer.
The company, funded by $3.5 million from venture capital firm Walden International and others, is targeting chips that operate with subthreshold voltages typically too small to turn on transistors, allowing electricity to flow through them and the circuit to carry out computations. That lowers power consumption, which is proportional to the square of the voltage.
Eta Compute, whose founders came from optical networking chipmaker Inphi, instead takes advantage of the current leakage that afflicts all computer chips. The company’s microcontrollers and other chips rouse themselves with little more than the current that leaks out of transistors technically turned off.
Based down the road from Inphi’s headquarters in Westlake, California, the company plans to sell the blueprints to chips based on its proprietary delay insensitive asynchronous logic (Dial) technology. With it, Eta Compute can build chips whose transistors switch with as little as 0.25 volts whereas traditional circuits switch with more than 1.2 volts.
Eta Compute’s processors can consume as little as two microwatts with the voltage throttled low enough. That means they can be powered by energy harvested from sunlight or vibrations instead of batteries. The tradeoff is that these microcontrollers and other chips are not as powerful, making them ideal for computing over long periods or in quick bursts.
Eta Compute claims that its technology can be used to create multicore processors that consume so little power that individual cores act like neurons in the human brain. These mosaics of silicon, called neuromorphic chips, imitate how humans learn new concepts, creating synaptic highways between neurons in the brain. Intel recently revealed it had built such an experimental chip.
“When you have such a low power core – let’s say a microwatt – well, ten thousand of these cores would consume only 10 milliwatts,” said Paul Washkewicz, vice president of marketing for Eta Compute, in an interview in March. That could enable self-learning chips that learn to recognize images or compose music by flipping through lots of previous examples.
Gopal Raghavan, chief executive of Eta Compute, said in a statement that Srinivasa was “the natural choice." Before he started as Intel’s chief scientist and senior engineer, Srinivasa led HRL Laboratories' Center for Neural and Emergent Systems and was principal investigator for Darpa’s Synapse program, which produced IBM’s neuromorphic chip True North.
The hiring could also boost Eta Compute in the cottage industry for subthreshold voltage chips. Its rivals include Psikick and Ambiq Micro, which has raised almost $80 million in funding from Arm Holdings and other investors since it was founded by chip designer Scott Hanson in 2010. Other competition includes Minima Processor, based in Oulu, Finland.