Researchers Claim Record-Breaking Speeds in New Optical Neuromorphic Chip
According to the researchers, this new type of optical neuromorphic processor can operate more than 1,000 times faster than any previous type of processor.
A new type of optical neuromorphic processor can operate more than 1,000 times faster than any previous type of processor, according to researchers at the Swinburne University of Technology.
Their newly-developed neuromorphic processor for artificial intelligence (AI) applications is said to function at almost 11 trillion operations per second (TOPS). This would make it the world’s fastest and most powerful optical neuromorphic processor for AI, allowing it to easily and quickly process ultra-large-scale data sets.
Artificial Neural Networks: A Current Key to AI
Artificial neural networks (ANNs) are computing systems that have been inspired by and designed to simulate the network of neurons that make up the human brain. An ANN is based on a collection of connected nodes called artificial neurons, and these are used to help an AI application "learn" things and make decisions in a more human-like manner.
Block diagram of the typical structure of an artificial neural network. Image used courtesy of Harish Rohil
A key form of AI, ANNs can perform complex operations with wide applications in areas like facial recognition, natural language processing, medical diagnosis, and speech translation. They do this by extracting key features of raw data to make predictions with unprecedented accuracy.
Enter Optical Micro-Combs
Optical neural networks, on the other hand, promise a dramatic increase in computing speed by making use of broad optical bandwidths.
In their research published in Nature, the Swinburne team demonstrates a universal optical vector convolutional accelerator operating at more than 10 TOPS, generating convolutions of images with 250,000 pixels, making it good enough for facial image recognition. The same hardware was also used to achieve successful recognition of handwritten digit images with 88 percent accuracy.
The Swinburne team says that this was achieved with so-called optical micro-combs. Micro-combs are a new type of device made up of hundreds of infrared lasers, and these are all housed on a single chip.
The micro-comb chip, which Dr. Xingyuan Xu shows off here, is an integral part of the new optical neuromorphic processor. Image used courtesy of Swinburne University of Technology
When compared to other optical sources, these optical chips are much smaller, lighter, faster, and cheaper. The team’s innovation uses a single processor while simultaneously interleaving the data in time, wavelength, and spatial dimensions via a single micro-comb chip.
Pushing ANNs to the Next Level
According to Dr. Xingyuan Xu, the processor is capable of serving as a universal ultrahigh bandwidth front-end for any neuromorphic hardware, either optical or electronic, enabling massive-data machine learning for real-time ultra-high bandwidth data.
“We’re currently getting a sneak-peak of how the processors of the future will look," Dr. Xu explains in the university’s press release. "It’s really showing us how dramatically we can scale the power of our processors through the innovative use of micro-combs."
Depiction of the operating principle behind the TOPS optical convolutional accelerator. Image used courtesy of Nature
This extra power comes at a lower cost and energy consumption, too. This is because optical technology makes ANNs faster and more efficient in contrast to existing silicon technologies, which are presenting more and more bottlenecks in processing speed and energy efficiency as AI and ANNs become more advanced.