Tensor operations are the kind of arithmetic that form the backbone of nearly all modern technologies, especially artificial intelligence, yet they extend beyond the simple math we’re familiar with. Imagine the mathematics behind rotating, slicing, or rearranging a Rubik’s cube along multiple dimensions. While humans and classical computers must perform these operations step by step, light can do them all at once.
Today, every task in AI, from image recognition to natural language processing, relies on tensor operations. However, the explosion of data has pushed conventional digital computing platforms, such as GPUs, to their limits in terms of speed, scalability and energy consumption.
Motivated by this pressing problem, international research collaboration led by Dr. Yufeng Zhang from the Photonics Group at Aalto University’s Department of Electronics and Nanoengineering has unlocked a new approach that performs complex tensor computations using a single propagation of light. The result is single-shot tensor computing, achieved at the speed of light itself.
“Our method performs the same kinds of operations that today’s GPUs handle, like convolutions and attention layers, but does them all at the speed of light,” said Zhang. “Instead of relying on electronic circuits, we use the physical properties of light to perform many computations simultaneously.”
To achieve this, the researchers encoded digital data into the amplitude and phase of light waves, effectively turning numbers into physical properties of the optical field. When these light fields interact and combine, they naturally carry out mathematical operations such as matrix and tensor multiplications, which form the core of deep learning algorithms. By introducing multiple wavelengths of light, the team extended this approach to handle even higher-order tensor operations.
“Imagine you’re a customs officer who must inspect every parcel through multiple machines with different functions and then sort them into the right bins,” Zhang explained. “Normally, you’d process each parcel one by one. Our optical computing method merges all parcels and all machines together — we create multiple ‘optical hooks’ that connect each input to its correct output. With just one operation, one pass of light, all inspections and sorting happen instantly and in parallel.”
Another key advantage of this method is its simplicity. The optical operations occur passively as the light propagates, so no active control or electronic switching is needed during computation.
“This approach can be implemented on almost any optical platform,” said Professor Zhipei Sun, Leader, Aalto University’s Photonics Group. “In the future, we plan to integrate this computational framework directly onto photonic chips, enabling light-based processors to perform complex AI tasks with extremely low power consumption.”
Here is an exclusive Tech Briefs interview, edited for length and clarity, with Zhang and Sun.
Tech Briefs: What was the biggest technical challenge you faced while developing this single-shot tensor computing?
Zhang and Sun: The biggest technical challenge is experimental implementation: achieving pixel-level registration between spatial light modulators (SLMs) while mitigating noise and environmental disturbances in a laboratory-scale setup. We addressed this by developing a pixel-level calibration procedure and performing extensive calibration and optimization. Although the resulting performance is not perfect, it is sufficient to validate the proposed method.
Tech Briefs: Can you explain in simple terms how it works please?
Z&S: We encode the input matrix A directly onto the spatial amplitude of an optical field. Leveraging the phase–position duality between the spatial phase and spatial frequency domains, we assign each row of A a unique linear phase gradient. After optical Fourier transformation and imaging, the light fields corresponding to all encoded rows are spatially superposed, so that every row of A performs a dot product with every column of matrix B in parallel. The linear phase serves as a positional tag, causing each product to appear at its correct location in the output matrix within a single optical propagation. Because linear phase is wavelength dependent, multi-wavelength multiplexing naturally extends this scheme from matrix–matrix multiplication to tensor–matrix multiplication.
Tech Briefs: Do you have any set plans for such further research/work/etc.? If not, what are your next steps?
Z&S: Yes, we have completed preliminary proof-of-concept and chip framework design and will work with our cooperator to complete tape-out and testing. In addition, we are also exploring applications of optical computing technology in other fields.
Tech Briefs: Do you have any advice for researchers aiming to bring their ideas to fruition?
Z&S: Because this work lies within classical applied physics rather than cutting-edge fundamental physics, the recommendations focus on the following: Conduct thorough theoretical analysis and numerical simulations before starting experiments. In wave optics and related applied fields, simulation results are typically highly reliable. During platform setup and experimentation, analyze each issue carefully considering theory and simulation outcomes; avoid attributing problems solely to “systematic errors.”
