Researchers demonstrate 65 TOPS optical computing chip with 124-channel parallel processing capability

Researchers from the Chinese Academy of Sciences and the University of Ottawa demonstrate a monolithic optical processing unit integrating four analog cores. Published on July 31, 2026, the chip supports 124-channel parallel processing, achieving 65.04 trillion operations per second (TOPS) and a compute density of 5.16 TOPS/mm². An optoelectronic convolutional neural network built on this platform achieves a 95.08% MNIST classification accuracy, addressing speed and integration limits in optical computing.

Optical computing chip architecture

• ▪The proposed optical processing unit simultaneously utilizes coherent interference, wavelength-division multiplexing, and spatial parallelism to process tasks.
• ▪Researchers designed an optical processing unit that integrates four optical analog cores on a single monolithic chip.

Multi-core parallel processing performance

• ▪The optical processing unit achieves a computational speed of 65.04 trillion operations per second.
• ▪The optical processing unit achieves a compute density of 5.16 trillion operations per second per square millimeter.
• ▪The four-core optical processing unit supports 124-channel parallel task processing.

Optoelectronic CNN implementation

• ▪The optoelectronic convolutional neural network achieves a 95.08% MNIST classification accuracy, representing a 9.20% improvement over its single-core counterpart.
• ▪Researchers constructed an optoelectronic convolutional neural network that combines the optical processing unit's parallel operations with electronic operations.

AI hardware scalability challenges

• ▪Artificial intelligence progress requires scalable, energy-efficient hardware capable of massive parallelism and high throughput.
• ▪Current optical computing implementations face speed limitations and multi-core integration challenges despite offering a post-Moore solution.

1 source