Labs Team’s Paper Proposes Advances in Non-Volatile Memory

[Labs_Editorial]

Over the years, silicon photonics has provided bandwidth and energy efficiency to drive applications in many aspects of communications. Hewlett Packard Labs has been a leader in this area – and now a Labs team is applying the technology to novel optical computing accelerators for AI.

A team in the Large-Scale Integrated Photonics Lab, led by Senior Research Scientist Bassem Tossoun, published a paper in Nature Communications proposing a way to help microprocessors act more efficiently by improving their memory capabilities. The paper, “High-Speed and Energy-Efficient Non-Volatile Silicon Photonic Memory Based on Heterogeneously Integrated Memresonator,” was published on Jan. 16.

The work Labs is doing could impact compute-hungry applications like deep neural networks, quantum computing, and field programmable arrays. Today, photonic integrated circuits are doing a good job powering these apps, but the circuits are constrained by limitations in the tuning speed and large power consumption of the phase shifters used to transmit light.

Tossoun’s team introduced the “memresonator,” an optical resonator which stores its resonant wavelength within a memory device, called a memristor. Tests the team conducted showed the memresonator providing a smooth pathway for reconfigurable, non-volatile silicon photonic phase shifters for programmable photonic integrated circuits.

“With current photonic technologies, the challenge is getting good performance in retention time and endurance in the nonvolatile memory,” Tossoun said. “There’s also the issue of speed and energy that it takes to switch or rewrite the data on memory. Those things can limit what type of application you can run. If it’s slow or power inefficient, it’s difficult to train neural networks on the hardware. It takes more time and more power.”

Bassem TossounTossoun’s team proposed two main innovations. By changing the resistance state of the memristor, they were able to tune the optical phase within the waveguide and alter the resonant wavelength of the device. And by integrating multiple memristors on the same chip as photonic neural networks, they can save significant amounts of energy and latency by avoiding energy lost in the transfer of data from the processor to an external memory chip.

Tests the team conducted showed the new circuits being able to retain memory for 12 hours, conduct switching voltages lower than 5 volts, and endure up to 1000 switching cycles. For perspective, these devices displayed an improved endurance from 300 to 1,000 switching cycles. Also, the switching speed and energy are both an order of magnitude lower than other previous implementations of these devices due to optimizing the material choice and design of the device.

Tossoun said the team is continuing the research project by designing variations of the devices and building large-scale in-memory photonic integrated circuits for AI accelerators.

“We’re going to explore the use of applications in circuits and ways to best operate these devices in a larger circuit,” he said. “The big picture problem is to develop novel circuits that can much more efficiently process AI workloads such as transformers and deep neural networks. This is one technology that could make a difference.”

The paper was authored by Bassem Tossoun, Di Liang, Stanley Cheung, Zhuoran Fang, Xia Sheng, John Paul Strachan and Raymond G. Beausoleil.