Syntiant Corp, an AI semiconductor firm, has unveiled its ultra-low-power analog neural network technology at Intel Capital's Global Summit in California.
The Irvine, CA-based startup claims its Neural Decision Processors (NDPs) utilise custom analog neural networks to optimise deep learning functions at the transistor level and eliminates data movement penalties, enabling smart devices to perform neural computation locally, faster, and more efficiently than traditional CPU, GPU and DSP options.
The semiconductor firm says NDPs are focused on always-on applications for battery-powered devices, such as keyword spotting, speaker identification, wake word, event detection, image recognition and sensor synthesis. “Analog neural networks and deep learning are a match made in heaven, delivering more than 50 times improvement in efficiency versus traditional digital stored-program architectures,” said Kurt Busch, CEO of Syntiant.
“Machine learning and AI in the cloud typically demands a great deal of power," said Wendell Brooks, president of Intel Capital. “For machine learning to be deployed in edge devices, it has to become much more power efficient. We believe ultra-low-power, analog neural networks ... could dramatically boost the adoption of distributed AI.”
According to Busch, Syntiant's technology combines the latest in deep learning research to deliver transformative functionality for battery powered devices – from ear buds to cell phones. "Ultimately, this will enable OEMs to bring machine learning and AI functionality to smart devices free from cloud connection, size, and power consumption constraints,” Busch noted.
Syntiant also announced a development agreement with Infineon Technologies to complement its NDPs with the company’s high-performance microphone technology. The firm believes its chip solutions merge deep learning with semiconductor design to produce highly efficient ultra-low power analog neural computation for always-on applications in battery powered devices.