Four optical computing cores hit 65 TOPS and boost MNIST accuracy

A four-core optical chip processed 124 channels in parallel and helped an AI image classifier reach 95.08% accuracy on MNIST.

Editorial illustration — not from the study.

The study proposes a high-throughput optical processing unit (OPU) that combines coherent interference, wavelength-division multiplexing, and spatial parallelism. On a monolithic chip, the researchers integrated four optical analog cores to support 124-channel parallel task processing, reporting 65.04 trillion operations per second (TOPS) and 5.16 TOPS/mm2 compute density.

To test the hardware in an AI setting, the team built an optoelectronic convolutional neural network (OE-CNN). The design fused the chip’s parallel 4-kernel convolution and average pooling with electronic nonlinear activation and fully connected layers, reporting 95.08% MNIST classification accuracy—9.20% higher than a single-core counterpart.

Four-core chip and MNIST test

The researchers integrated four optical analog cores on a monolithic chip to create an OPU capable of 124-channel parallel task processing. They reported a computational speed of 65.04 TOPS and 5.16 TOPS/mm2 compute density. Using this platform, their OE-CNN achieved 95.08% accuracy on MNIST, which they state is a 9.20% improvement over a single-core version.

Parallel optical computing in an AI pipeline

The work aims to show that optical hardware can be arranged to run many channels in parallel and still be used inside a neural-network pipeline. By pairing the multi-core optical processing unit with convolution and pooling in a neural network, the authors report an accuracy gain on MNIST alongside the higher-throughput hardware demonstration.

Hardware demo and MNIST results

This is a journal article describing a hardware demonstration and a neural-network test on MNIST. The results reported in the abstract are specific to the system and configuration the authors built; the abstract does not say how performance would scale to other tasks, larger datasets, different network sizes, or real-world deployments, nor does it provide details on power use, training setup, or broader comparisons.

Reliable AI partnerships depend on people, not just stronger models

A survey of human-AI collaboration argues that more capable AI systems do not automatically produce reliable partnerships. It identifies human-centered design, governance and attention to high-stakes uses as central to making these systems trustworthy and beneficial.

9 hours ago·6 views

AI & Computing

RNA splicing tools can follow irrelevant clues and miss RNA’s shape

Deep-learning tools for predicting RNA splicing can rely on genomic clues unrelated to splicing and miss important effects of RNA structure, a study finds. These weaknesses were linked to systematic errors, especially when the tools evaluated sequences that differed from those used for training.

20 hours ago·5 views

AI & Computing

Better-trained AI finds more flaws in unfamiliar software projects

AI vulnerability detectors often perform less well on software projects outside their training data. A study finds that cleaner, more varied training data and an encoder-based design improved detection, including a reported 6.8% gain in recall on the BigVul benchmark.

2 days ago·34 views