Electrical Efficiency

The Massachusetts Institute of Technology (MIT) will develop a unified optical communication technology for use in datacenter optical interconnects. Compared to existing interconnect solutions, the proposed approach exhibits high energy efficiency and large bandwidth density, as well as a low-cost packaging design. Specifically, the team aims to develop novel photonic material, device, and heterogeneously integrated interconnection technologies that are scalable across chip-, board-, and rack-interconnect hierarchy levels.

Columbia University will develop a new datacenter architecture co-designed with state-of-the-art silicon photonic technologies to reduce system-wide energy consumption. The team’s approach will improve data movement between processor/memory and will optimize resource allocation throughout the network to minimize idle times and wasted energy. Data transfer in datacenters occurs over a series of interconnects that link different server racks of the datacenter together. Networks in modern mega-scale datacenters are becoming increasingly complicated.

Ayar Labs will develop new intra-rack configurations using silicon-based photonic (optical) transceivers, optical devices that transmit and receive information. The team will additionally develop methods to package their photonic transceiver with an electronic processor chip. Marrying these two components will reduce the size and cost of the chip system. Integrated packaging also moves the photonics closer to the chip, which increases energy efficiency by reducing the amount of "hops" between components.

The IBM T.J. Watson Research Center will develop datacenter networking technology incorporating extremely fast switching devices that operate on the nanosecond scale. At the heart of the process is the development of a new type of photonic switch. The dominant switching technology today are electronic switches that toggle connections between two wires, each wire providing a different communication channel. A photonic switch toggles connections between two optical fibers, where each individual fiber themselves can carry many communication channels allowing immense numbers of data transfers.

The University of Southern California (USC) will develop a framework and testbed for evaluating proposed photonic and optical-electronic interconnect technologies, such as those developed under the ARPA-E ENLITENED program. These new approaches will develop novel network topologies enabled by integrated photonics technologies, which use light instead of electricity to transmit information. USC’s effort aims to offer an impartial assessment of these emerging datacenter concepts and architectures and their ability to reduce overall power consumption in a meaningful way.

The United Technologies Research Center (UTRC) will develop an AI-accelerated search technique, LENS, to quickly discover new design concepts for energy applications. The project will combine the strengths of the two pillars of AI—logical inference and statistical learning—to achieve this task by using constraint programming, generative models, reduced order models, active learning, and rule discovery. The end goal is to accelerate the design of power converters, which have a significant impact on energy savings.

Northwestern University will develop a machine learning-enhanced mixed-variable conceptual design optimization framework to construct new functional materials for energy savings. The team will use natural language processing (NLP) and physics-based machine ML to more efficiently guide the autonomous search for materials. The project will deliver a series of new ML techniques using NLP, conditional variational autoencoders, active learning, latent-variable Gaussian processes, and reinforcement learning in Bayesian optimization.

IBM Research will develop a reinforcement learning (RL)-based electrical power converter design tool. Such converters are widely used and critically important in many applications. Designing a specific converter is a lengthy and expensive process that involves multiple manual steps—selecting and configuring the correct components and topologies; evaluating the design performance via simulations; and iteratively optimizing the design while satisfying resource, technology, and cost constraints.

The Princeton University team will use machine learning-enabled methods to transform the modeling and design methods of power magnetics and catalyze disruptive improvements to power electronics design tools. They will develop a highly automated, open-source, machine learning-based magnetics design platform to greatly accelerate the design process, cut the error rate in half, and provide new insights to magnetic material and geometry design.

The University of Michigan-Dearborn will develop a machine learning-enhanced design tool for the automated architectural configuration and performance evaluation of electrical power converters. This tool will help engineers consider a wider range of innovative concepts when developing new converters than would be possible via traditional approaches.