Edge AI Chips for Voice-Activated Devices Save Power, Protect Privacy

August 21, 2020 by Luke James

Machine learning startup Syntiant announced in early August that it had secured $35 million to push its processors forward.

Artificial intelligence (AI) and voice-focused chip start-up Syntiant, known for merging machine learning with semiconductor design, announced in early August that it had successfully raised $35 million during a Series C venture funding round led by Apple Ventures and Microsoft’s M12 Ventures.

The company also announced that it has shipped more than one million units of its Syntiant NDP100 and Syntiant NDP101 microwatt-power neural decision processors (NDPs) to customers worldwide.


Processing at the Edge

Syntiant’s NPD100 and NPD101 processors measure roughly 1.4 mm by 1.8 mm and can run models with over 500,000 parameters. They include a general-purpose Arm Cortex-M0 processor paired with 128 KB of RAM. Syntiant says the devices only consume 140 uW while processing.

According to the company, the two processors are designed to operate on the edge without having to send data to the cloud. This makes them more power-efficient in contrast to standard chipsets when it comes to processing voice assistant tasks, such as commands. 


Syntiant's NDP100 measures at 1.8 mm x 1.4 mm

Syntiant's NDP100 measures at 1.8 mm x 1.4 mm. Image used courtesy of Syntiant


By keeping voice processing on the device rather than utilizing the cloud, Syntiant chips are said to help protect end-user privacy. Syntiant notes that this focus on privacy may ease the concerns of consumers who are currently reluctant to purchase voice-activated devices due to existing privacy concerns. 


Neural Decision Processor

Syntiant chips are powered by the company’s neural decision processor (NDP) IP that is capable of processing audio and other sensory data to determine if an activation signal, such as a wake word, has been used. This is done internally, which is what negates the need for cloud server verification.


Block diagram of the NDP101 neural decision processor

Block diagram of the NDP101 neural decision processor. Image used courtesy of Syntiant

The company claims that this internal processing means that its chips can process ten times the data using just 1/100th of the energy required when compared to standard chips. They can achieve this feat despite being roughly half the size of their counterparts, according to Syntiant.


One Million+ Chips Shipped for Voice-Activated Devices

There is clearly a demand for these chips, given that Syntiant has shipped over one million of them already. Last year, the company closed a deal with Amazon in which Amazon certified Syntiant’s chips for integration into Alexa-powered devices. They’re also known to be used in devices running Google Assistant and Siri. 

The closing of Syntiant’s Series C round brings the company’s total funding to around $65 million. This includes previous investments from Intel, M12 Ventures, and the Amazon Alexa Fund. The $35 million Series C funding will go toward scaling up Syntiant’s production capacity so more chips can be put into more devices. 

In a statement, Syntiant’s CEO Kurt Busch said,  “We are especially thrilled that production volumes of applications using our neural decision processors are increasing and expect orders to ramp even higher throughout the remainder of 2020, as our NDPs continues to set the standard for always-on voice as the new interface.”