New chip opens door to AI computing at light speed

This approach has the potential to significantly boost the processing speeds of computers while simultaneously reducing their energy consumption, marking a monumental step forward in computing technology.

The chip, rooted in silicon-photonic (SiPh) technology, is hailed as the first of its kind to merge the pioneering work of Benjamin Franklin Medal Laureate and H. Nedwill Ramsey Professor Nader Engheta in nanoscale material manipulation for mathematical computations using light, with the cost-effective and widely available silicon used in mass computer chip production. This fusion of cutting-edge research with practical materials could herald a new era of computing power and efficiency.

Light's interaction with matter is identified as a key pathway to transcending the limitations of contemporary chips, which, despite decades of advancements, still fundamentally operate on principles established in the 1960s. This new chip signifies a leap towards surpassing these limitations by harnessing the unparalleled speed of light for communication.

In their collaborative research published in Nature Photonics, Engheta's team, along with Associate Professor in Electrical and Systems Engineering Firooz Aflatouni's group, outlines the development of this chip. Engheta reveals: “We decided to join forces,” taking advantage of Aflatouni’s pioneering efforts in developing nanoscale silicon devices.

The core aim of their collaboration was to create a platform capable of executing vector-matrix multiplication, an essential mathematical operation in neural networks and, by extension, the AI tools powered by these networks. This operation is crucial for both the development and functionality of AI applications.

Explaining the technological innovation, Engheta said: “Instead of using a silicon wafer of uniform height, you make the silicon thinner, say 150 nanometres,” but this thinning is applied selectively. These height variations, achieved without incorporating additional materials, manipulate light's propagation through the chip. This manipulation allows the chip to scatter light in precise patterns, enabling it to perform calculations at the speed of light.

Aflatouni notes that, despite the manufacturing constraints of the commercial foundry where the chips were produced, the design is ready for commercial application. He suggests its potential adaptation for graphics processing units (GPUs), which have seen surging demand with the growing interest in AI development. “They can adopt the Silicon Photonics platform as an add-on,” Aflatouni proposes, “and then you could speed up training and classification.”

Beyond the promises of accelerated processing speeds and reduced energy consumption, the chip also offers significant privacy benefits. With the capability for numerous computations to occur simultaneously, there's no necessity to store sensitive data in a computer's working memory. Aflatouni emphasises this advantage, stating: “No one can hack into a non-existing memory to access your information,” envisaging a future where computers powered by such technology are virtually impervious to hacking.

This pioneering development not only represents a significant technological leap in computing but also underscores the potential for SiPh technology to redefine the landscape of AI development and application, merging unparalleled speed with enhanced privacy and efficiency.