University of Pennsylvania engineers have developed a novel chip that utilizes light waves, as opposed to electricity, for conducting the intricate mathematical operations crucial for AI training. This innovative chip holds the promise of significantly enhancing the computational speed of computers while concurrently reducing their energy consumption.
The design of the silicon-photonic (SiPh) chip represents a pioneering integration of the research conducted by Benjamin Franklin Medal Laureate and H. Nedwill Ramsey Professor Nader Engheta in manipulating nanoscale materials to execute mathematical computations using light—the fastest mode of communication—with the SiPh platform that employs silicon, an abundant and cost-effective element utilized in the mass production of computer chips.
The utilization of light waves interacting with matter presents a potential pathway for developing computers that surpass the constraints of current chips, which are fundamentally rooted in the principles established during the early days of the computing revolution in the 1960s.
In a publication featured in Nature Photonics, Engheta’s team, in collaboration with Firooz Aflatouni, Associate Professor in Electrical and Systems Engineering, delineates the creation of this groundbreaking chip.
Engheta states, “We opted to collaborate,” capitalizing on Aflatouni’s group’s expertise in nanoscale silicon devices.
Their objective was to establish a platform for executing vector-matrix multiplication, a fundamental mathematical operation in the advancement and operation of neural networks, the foundational computer architecture driving contemporary AI applications.
Engheta elaborates on the chip’s design, highlighting the strategic thinning of the silicon to around 150 nanometers in specific regions rather than maintaining a uniform thickness across the wafer. These height variations—devoid of additional materials—enable the regulation of light propagation within the chip. The controlled scattering of light induced by these height discrepancies facilitates the chip’s ability to perform mathematical computations at the speed of light.
Aflatouni underscores that due to the limitations imposed by the commercial foundry responsible for chip production, this design is already primed for commercial utilization and could potentially be adapted for integration into graphics processing units (GPUs), which are in high demand amidst the burgeoning interest in developing new AI systems.
Aflatouni envisions, “They can integrate the Silicon Photonics platform as an extension,” thereby expediting training and classification processes.
Apart from the advantages of increased speed and reduced energy consumption, Engheta and Aflatouni’s chip boasts privacy benefits: With multiple computations occurring concurrently, there is no necessity to store sensitive data in a computer’s active memory, rendering a future computer powered by this technology nearly impregnable to hacking attempts.
Aflatouni emphasizes, “No unauthorized access to your data by breaching non-existent memory.”
The co-authors of this groundbreaking research include Vahid Nikkhah, Ali Pirmoradi, Farshid Ashtiani, and Brian Edwards from Penn Engineering.
More information:
“Inverse-designed low-index-contrast structures on silicon photonics platform for vector-matrix multiplication,” published in Nature Photonics (2024). DOI: 10.1038/s41566-024-01394-2. www.nature.com/articles/s41566-024-01394-2
Provided by University of Pennsylvania
Citation:
“New chip opens door to AI computing at light speed” (2024, February 16)
Retrieved on February 16, 2024
from https://phys.org/news/2024-02-chip-door-ai.html
Copyright restrictions apply to this document. Any usage beyond fair dealing for private study or research requires written permission. The content is shared for informational purposes only.
