Researchers at Tsinghua University Beijing have reported an all-analog optoelectronic chip aimed at visual processing that they claim is 3,000 times faster than conventional electronic processors.
The system is named “all-analog chip combining electronic and light computing,” or ACCEL.
The layout is 99 percent optical processing in the analog domain and this is the source of the speed and extreme low-power of the system, according to a paper published in Nature: All-analog photoelectronic chip for high-speed vision tasks.
The paper reports that ACCEL is able to recognize and classify objects with an accuracy that is comparable with digital neural networks. It classifies high-resolution images of daily life more than 3,000 times faster and with 4,000,000 times less energy consumption than a top-of-the-line GPU, the researchers said.
A diffractive optical analog computing (OAC) module processes the input image in the optical domain for feature extraction, and its output light field is used to generate photocurrents in a photodiode array for analog electronic computing. This electronic analog computing (EAC) chip outputs sequential pulses corresponding to multiple output nodes of the equivalent network.
The EAC is an array of 32 by 32 pixel circuits here corresponding to the calculation matrix of 1,024 weighted outputs.The binary weights in EAC are reconfigured during each pulse by SRAM, by switching the connection of photodiodes to either V+ or V− lines. The comparator outputs the pulse with the maximum voltage as the predicted result of ACCEL.
The researchers report that the ACCEL system achieves a 4.6 peta-operations per second and an energy efficiency of 74.8 peta-operations per second per watt.
One of the key breakthroughs is using an optical encoder for feature extraction with the subsequent light-induced photocurrents used directly for further calculation in an integrated EAC chip without the requirement for ADCs. ADCs are big source of energy consumption and latency in digitally based hybrid optoelectronic systems. As a result has a frame latency of 72ns.
ACCEL can be used across a broad range of applications such as wearable devices, autonomous driving and industrial inspections the researchers said.
The researchers point out that ACCEL is a low-cost solution as it make use of 180nm CMOS manufacturing process for the EAC and low-cost etched SiO2 panels for the OAC compared with state-of-the-art GPUs and tensor processing units that require more advanced processes.
Advanced CMOS technology could be used in ACCEL to reduce the power consumption of the control unit and allowing it to operate at a higher clock frequency.