Smart Speakers Digitize All Input, Zap Battery. A New ML Analog Chip Has “Selective Hearing”
Always-on devices zap battery when digital processors analyze 100% of input. A machine-learning analog chip can save power by only activating for selective input.
The world is filled with millions of sensors that are on duty 24/7, 365 days a year. But, the events they report can be exceedingly rare—perhaps a dangerous spike in heat, a pipeline leak, or an intruder. That sensor might be in a remote or even dangerous location with no power source except a battery.
When the sensor does receive a stimulus, it expends its battery current analyzing and reporting the anomaly. But the problem is that even when there is no reportable event, the sensor is still digitizing irrelevant data and drawing current. All too soon, the battery needs to be replaced or recharged.
Aspinity’s Reconfigurable Analog Modular Processing (RAMP) chips act as ultra-low-power gatekeepers, alerting members of Infineon’s XENSIV family of sensors when to turn on and begin their power intensive digital analytic tasks. Otherwise, the sensor stays in low-power sleep mode. Far less power is consumed, resulting in an extended life for batteries.
The analog RAMP chip uses machine learning to only recognize trigger sounds without wasting power digitizing irrelevant sounds. Screenshot used courtesy of Aspinity
This Aspinity-Infineon partnership are aimed at "always-on" systems, which Cadence describes as a system in which some compute resources are "always on to process audio, visual, or other sensor data while the more powerful compute resources in the system are turned off." We can see demand for "always-on" efficiency most prevalently in wearable, mobile devices, voice-activated smart speakers, and any number of IoT devices.
What Is RAMP Technology?
The idea behind RAMP technology is to not wake up the power-consuming digital sensor until it needs to wake up. Most digital processors "listen to" 100% of input, even if they don't analyze most of it, which represents a major strain on battery power. The RAMP chip, on the other hand, is an extremely low-power analog chip that only alerts the sensor to an event it must analyze before processing it digitally. The sensor can “rest” and be woken up only when it’s time to get to work. The analog chip is trained with neural networks to recognize its wake-up call.
Think of it like a cat, sprawled out and half-asleep on the couch, completely ignoring the noisy world surrounding it. But, when somebody opens the refrigerator door, the cat is instantly “notified” and springs into action. The basic idea (of the RAMP chip, not the cat) is illustrated below.
Aspinity's analyze-first architecture. Image used courtesy of Aspinity
On the left, the sensor, with its relatively large appetite for power is always only, digitizing and analyzing even when there’s nothing relevant for it. On the right, the RAMP analog processor, using very little power, stands guard, listening for that “refrigerator door" to open and only waking up the somnolent sensor when it’s needed.
The result is an extremely large saving in power with the battery lasting far longer than it would have otherwise.
The All-Analog Chip Employs Machine Learning
RAMP is an all-analog chip that enables a new approach to machine learning, allowing sophisticated digital signal processing tasks to be replicated in analog. It features an analog neural network and a flexible sensor interface to accommodate a range of sensor types. The RAMP chip, fully alert at all times, draws as little as 10 µA of current.
RAMP is an ultra-low processing technology that incorporates machine learning into an analog processor. Screenshot used courtesy of Aspinity
Aspinity offers the RAMP development environment that has been specifically designed for engineers without analog expertise, allowing them to develop RAMP algorithms through the types of interfaces that they are familiar with.
Infineon Touts the Broadest Portfolio of Sensor Types
Where does Infineon come into this collaboration? Infineon claims to offer the broadest portfolio of sensor types on the market, some of which include magnetic sensors, pressure sensors, acoustic sensors, 3D image sensors, and radar sensor MMICs.
One of Infineon's XENSIV hall-effect sensors for high-precision automotive/industrial applications. Image used courtesy of Infineon
According to the press release, the RAMP chip assesses analog data derived from Infineon's XENSIV MEMs sensor, identifying only relevant data. Aspinity notes that designers can "easily program a RAMP chip for application-specific inferencing" and that combining the RAMP chip with the XENSIV sensors give rise to a "power-efficient analyze-first architecture" for always-on devices.
Tom Doyle, founder and CEO of Aspinity states that “We look forward to collaborating with Infineon to enable customers to overcome the power challenges associated with integrating high-performance always-on sensing into a growing array of battery-operated always-on products.”
What are some design methods you've tried for conserving power in always-on devices? Share your thoughts in the comments below.
That is very interesting and important subject. Thank you for sharing.