Tensor operations are a type of superior arithmetic that assist many fashionable applied sciences, particularly synthetic intelligence. These operations go far past the straightforward calculations most individuals encounter. A useful solution to image them is to think about manipulating a Rubik’s dice in a number of dimensions without delay by rotating, slicing, or rearranging its layers. People and conventional computer systems should break these duties into sequences, however mild can carry out all of them on the identical time.
Right this moment, tensor operations are important for AI techniques concerned in picture processing, language understanding, and numerous different duties. As the quantity of information continues to develop, typical digital {hardware} comparable to GPUs faces rising pressure in pace, power use, and scalability.
Researchers Exhibit Single-Shot Tensor Computing With Mild
To handle these challenges, a world workforce led by Dr. Yufeng Zhang from the Photonics Group at Aalto College’s Division of Electronics and Nanoengineering has developed a essentially new strategy. Their methodology permits complicated tensor calculations to be accomplished inside a single motion of sunshine via an optical system. The method, described as single-shot tensor computing, capabilities on the pace of sunshine.
“Our methodology performs the identical sorts of operations that at present’s GPUs deal with, like convolutions and a focus layers, however does all of them on the pace of sunshine,” says Dr. Zhang. “As a substitute of counting on digital circuits, we use the bodily properties of sunshine to carry out many computations concurrently.”
Encoding Info Into Mild for Excessive-Velocity Computation
The workforce achieved this by embedding digital data into the amplitude and part of sunshine waves, remodeling numerical knowledge into bodily variations inside the optical discipline. As these mild waves work together, they routinely perform mathematical procedures comparable to matrix and tensor multiplication, which kind the premise of deep studying. By working with a number of wavelengths of sunshine, the researchers expanded their approach to assist much more complicated, higher-order tensor operations.
“Think about you are a customs officer who should examine each parcel via a number of machines with completely different capabilities after which kind them into the precise bins,” Zhang says. “Usually, you’d course of every parcel one after the other. Our optical computing methodology merges all parcels and all machines collectively — we create a number of ‘optical hooks’ that join every enter to its right output. With only one operation, one move of sunshine, all inspections and sorting occur immediately and in parallel.”
Passive Optical Processing With Broad Compatibility
Some of the placing advantages of this methodology is how little intervention it requires. The mandatory operations happen on their very own as the sunshine travels, so the system doesn’t want energetic management or digital switching throughout computation.
“This strategy might be applied on virtually any optical platform,” says Professor Zhipei Solar, chief of Aalto College’s Photonics Group. “Sooner or later, we plan to combine this computational framework immediately onto photonic chips, enabling light-based processors to carry out complicated AI duties with extraordinarily low energy consumption.”
Path Towards Future Mild-Based mostly AI {Hardware}
Zhang notes that the last word goal is to adapt the approach to present {hardware} and platforms utilized by main know-how corporations. He estimates that the strategy might be included into such techniques inside 3 to five years.
“It will create a brand new technology of optical computing techniques, considerably accelerating complicated AI duties throughout a myriad of fields,” he concludes.
The examine was revealed in Nature Photonics on November 14th, 2025.
Leave a Reply