Aspinity’s Analog Neural Net Wake-Up Call

Startup’s analog neural network chip and ‘analyze-first’ architecture may lower power use and data congestion in always-on voice systems.


Putting an analog chip in front of an always-on system for digitizing speech and having the analog chip listen for sounds of interest may help avoid huge power waste and data congestion in current voice-recognition systems. Aspinity, an analog neuromorphic semiconductor startup, has worked the problem and just announced its Reconfigurable Analog Modular Processor (RAMP) platform yesterday. RAMP is an analog chip with a neural network that is more efficient than digital for listening and waking up a system.

Always-on applications, such as digital assistants, have a ‘digitize-first’ structure where the system is on constantly listening and sending data to the cloud for analysis even as it waits for the wake word, such as “Hey Computer,” from a human user. Not only is a lot of power used in keeping the system always on, but system is always transmitting data back to the cloud even when no sound of interest exists. The privacy concerns are hard to stomach.

In the RAMP’s ‘analyze-first’ structure, the analog neural network chip can listen without sending data to the cloud for processing. The initial processing and data cleaning is done on the analog neural network. “There’s still a need to do digital work,” said Tom Doyle, founder and CEO, Aspinity Inc., in an interview with Semiconductor Engineering, “but if there are things that should be done in analog and done more efficiently in analog, then we can wake up the rest of the system for the higher level.”

The analog blocks in RAMP are modular, parallel and continuously operating, said Aspinity’s press release. The RAMP platform analog circuitry supports event detection and classification from raw, unstructured analog sensor data. Its analog blocks are reprogrammable.

More efficient?
Is analog more efficient for processing sound waves, which are of course, analog? “There are definitely things that we can do in analog that are more efficient,” said Doyle. Some of the efficiencies are the nature of analog and some are Aspinity’s innovations, such as smaller transistors. “The core of a neural network is a multiply and accumulate functions. We can actually build those given our innovation with much smaller transistor. We’re able to actually do that in some ways more efficiently in analog.”

“Voice is a great one because while we’re not focused on commands and things of that nature in our chip—that’s better left for a high-powered digital processor. What we can do is say ‘hey we’re looking at all the sound data and we recognize speech is present. So go ahead, start processing for commands,’ versus trying to do that stuff when there’s no speech at all. Those things need a lot of sense,” said Doyle.

Other innovations have helped Aspinity store information more efficiently in analog. “Typically with mobile networks and with analog, you have wave biasing and so we’re able to store them values given another one of our innovations much more efficiently. We’re able to build this modular process or this configurable processor in a much smaller footprint than standard analog. A few different innovations allow us to get to this level where we can do things more efficiently.”

Same path
Using Aspinity’s analog neural network chip is intended to feel familiar to machine learning people. “It’s very similar to what happens in the digital world but we’re able to do it in analog,” said Doyle. “We follow a very similar path.”

“You can train a DSP, you can train an Nvidia core, or you can train these other cores in the digital world. Many of those other chips have a neural network that they used for decision-making and classification of information,” said Doyle. “We follow the same path. We built an analog neural network on our chips, so it can have the same characteristics that you would find in the digital world, so a machine learning person or someone who is building an application can follow the same path. They would keep training data, they would use our environment to go ahead and build a model that would subsequently run on our RAMP chip that would be geared toward detecting a voice, detecting breaking glass, detecting alarms, so detecting these same events that you could you could detect an digital but we do it much sooner in the signal chain.”

The RAMP platform also does some data cleaning in the chip using an algorithm that takes the raw data. “We will do other things that are important to the data. Everybody believes that analog is raw unstructured—it may not be clean data, it may need conditioning or you may need some sensor interfacing to boost up the gain—things of that nature. We do those things inside our chip. When we look at our chip, we build an algorithm that we can train and build an algorithm that goes in our chip; that algorithm takes the raw data. We do the necessary cleaning of it [the data], we will do some feature extraction and then we go into a neural network that makes a decision. It’s very similar to what happens in the digital world but we’re able to do it in analog.”

Fig. 01: Aspinity’s analyze-first system architecture. (Source: Aspinity)

Longer battery life
How the data is handled directly affects power consumption. “We really see data and battery life as hand-to-hand,” said Tom Doyle, founder and CEO, Aspinity Inc. “Where you analyze the data, what you do with the data in relation to the sensor and the signal path plays a big role in the battery life.”

“We’ve seen a lot of movement over the last decades putting a lot of computing and processing in the cloud where we have an abundance of processing power and capacity, but now we’re moving this back to the edge.” said Doyle. “The edge brings about some big challenges and trade-offs not only battery life but features, accuracy, and we’re reducing what we do at the edge so that we can last longer on battery. That has an impact as well.”

The idea of digitizing all the raw data is a big problem, said Doyle, although he admits to some benefits to a ‘digitize first’ approach.  “But we can do things better and earlier in the signal chain so that we are not digitizing all that data and not pushing all this raw, irrelevant data forward. That’s what we look at and want to help with—that tradeoff.”

“It’s the signal chain approach. All the data is analog. A lot of what we collect is what we call raw unstructured analog. And so we spent a lot of effort pushing all that needed to digital processors to gather insights,” said Doyle. “But we find that the vast majority of the time that the data that we’re looking for is something such as speech, such as a wake word in data that doesn’t include speech. [‘Digitize first’] is totally inefficient in that respect.”

The analog is efficient that Aspinity claims RAMP uses ten times less power and compressing the quantity of vibration data by 100x.

Fig 02: The DSP wake word engine gets extra help from a RAMP analog chip. (Source: Aspinity)

RAMP analyzes the incoming sound at the microphone edge to keep the wake-word engine and other digital processors in a low-power sleep state for the 80% of the time that no voice is present.

RAMP is aimed at battery-operated, always-on sensing devices for consumer, smart home, Internet of Things (IoT), industrial and other markets. Aspinity, founded in 2015 from ideas formulated at West Virginia University, is now just coming out of stealth mode. Amazon is an investor, coming through two rounds. The RAMP chip is in silicon now but will be offered more broadly to customers next year. Made on a mature process node at Taiwan foundries, including TSMC, the cost may be lower for these analog chips. “We’re in the lower cost analog process-mature node technologies,” said Doyle. Time will tell if it pays to be older, more mature and analog.

Related Story

Using Analog For AI

Leave a Reply

(Note: This name will be displayed publicly)