AIS. AI-in-Sensor. No digitizing. No latency. No distortion. No bus. No power -- almost. No problem.
AIS. AI-in-Sensor. No digitizing. No latency . No distortion. No bus. No power -- almost. No problem.
A storm is coming. AIStorm presents its AI-in-Sensor SoCs…
Sensors are not digital devices. To convert sensor information into a digital representation takes time, power, silicon, and adds distortion & noise. AIS solutions can be finished processing before the competition even has enough digital information to begin. So why not skip the middle man and consider our SoCs for your next project?
LOWEST POWER
Because we don’t have to digitize sensor information, we offer the lowest power usage for performance in the industry.
HIGHEST PERFORMANCE
Our patented charge domain technology guarantees a high performance, fully integrated SoC that you can’t find anywhere else.
LOW COST
We pass cost savings onto you by eliminating energy-intensive GPUs and DSPs so that you can focus on making your next project even better.
Technology
Digital systems typically struggle to provide an “always on“ operation because of the need to continuously poll, digitize and analyze data quickly enough not to miss a transaction or seem slow to function, especially when they contain an analog sensor.
The sensors are the system's eyes, ears and fingers. Sensors can focus on what we care about and ignore the rest – a face, another car, an intruder, a sound, a change...
AIS technology is analog and mixed signal technology, but it's more than that. For the first time the sensor itself is smart enough to deal with events, make decisions and can perform complex analysis…often before the competition even receives a digital sensor representation to start processing.
Our solution "skips the middle steps" of digital processing and polling so that you can get straight to the results, reactions and outcomes you're looking for.
The core of our technology is our patented charge domain MAC.
The benefits of this solution are lower noise, power, speed and size over similar technologies including memristors, resistive RAM, NVM-controlled resistors and other competing technologies.
Our architecture is based around four core ideas:
Event-driven charge processing using 10’s to 1000’s of electrons each cycle
Virtual neurons taking advantage of >1GHz equivalent digital processing capability
Eliminating digitization through on chip sensors coupling directly to the AI
Scalable, software driven, on chip networking driven by standard tools
There are several companies that are offering analog AI solutions, but these products are challenged by fundamental physical noise and bandwidth limitations that cannot be overcome.
AIStorm has taken a different path –charge domain processing. Working with charge, this method opens up opportunities for AI-in-Sensor processing for image or audio enhancement that just cannot be implemented any other way.
Charge domain processing is a revolution in processing that overcomes the noise and bandwidth limitations of other analog solutions. Charge domain solutions can accept IoT sensor information directly without the cost, power and latency of digitization and therefore represent a better solution over both analog and digital solutions.
Our Metrics Matter
- Up to 1000 TOPS/W possible (highest performing MAC)
- 10x Lower Cost – In trailing lithography or >100x faster in same node
- >8 Bit Operation – Exceeds theoretical accuracy limit of memristor, NVM or analog multiplier-base
- No input digitization – Integrates sensor, such as pixels or audio MEMs membrane into pre-processing or neural network input layer
- Fastest Imaging – Best in class imaging frame rate. > 50,000 frames per second
- <5uA “always on” with no polling delay – Capable of performing high speed extraction only of items of interest based upon changing pixels or smart regional pattern matching
- Smartest event driven capability – Smart event driven capability such as “wake up on word” or “wakeup on face” without digital polling
- Flexible – Compatible with standard neural network configurations or neuromorphic spiking approaches
- Standard Tool Flow – Compatible with standard tools such as TensorFlow Lite
