Embedded Charge Domain Neural Network for Sensors to Enable Low-Power Edge Computing

Session details:

ecent advances in neural-network hardware architectures have seen the emergence of a transformative paradigm: Charge-Domain AI-in-Sensor Processing, which fundamentally redefines how imaging, audio, and and sensing applications perform on-device inference. This presentation introduces the underlying theory and implementation of this novel neural-network approach, emphasizing its architecture, operational advantages, and potential to reshape low-power for always-on edge and mobile applications.

At the heart of this innovation is the integration of sensor element which output directly into the neural-network architecture. Traditional digital and Process-in-Memory (PIM) frameworks rely on separate conversion and memory stages—requiring significant analog-to-digital conversion, volatile memory access, and extensive data movement. In contrast, charge-domain processing eliminates these inefficiencies by enabling sensor inputs to feed directly into an analog charge-multiplying neural layer, thus drastically reducing latency, silicon area, power consumption, and system complexity

In AIStorm’s charge-domain neural network architecture, electronic charge—rather than voltage, current, or digital bits—serves as both the information carrier and terms which are computated. Charges can be precisely manipulated across storage nodes in a manner that is inherently robust against manufacturing, process, voltage, and temperature variations due to AIStorm’s patented techniques. Charge-domain neurons leverage rapid, pulse-based switch matrices and allow dynamic weight substitution each computation cycle. Coupled with recursion, this enables a modest set of neurons to simulate the behavior of much larger networks through time-multiplexing, improving computational density as well as flexibility.

The result is an amazing energy and area efficiency. AIStorm’s Charge-domain implementations can reach over 300 to 600x higher TOPS/W as compared to to traditional PIM or GPU networks (comparing the same models and technology nodes).  Memory requirements shrink dramatically due to fewer cycles and localized computation. The architecture’s always-on capability stems from the sensor itself being part of the neural network, eliminating idle trigger dependence and making continuous inference both feasible and energy efficient. 

AIStorm’s charge-domain neural network delivers low latency by performing inference directly on the sensor without requiring a transfer to an external processor, this is accomplished by bypassing the analog to digital conversion which contributes to both speed and energy efficiency.  The elimination of analog-to-digital converters and separate memory banks leads to significant manufacturing and BOM cost reductions, offering solutions up to 5-30x smaller and less expensive than PIM or digital alternatives.

Real-time applications; such as biometrics, gesture recognition, inventory tracking, or environmental monitoring; stand to benefit enormously from this model. The inherent robustness to process and temperature variations, combined with the capacity for real-time adaptability, creates resilient architectures ideal for unpredictable environments.

AIStorm’s Charge domain neural processing is an emerging and compelling direction for next-generation in sensor compute. It unites sensors and computation, enabling high-throughput, low-power, minimal-latency inference directly at the edge. The resulting systems offer transformative benefits—energy savings, cost reduction, compact form factors, and agility—poised to redefine real-time AI deployment in consumer, industrial, and smart-environment contexts.

As this technology matures, its combination of analog precision, direct sensor integration, and neural flexibility positions it as a potential standard for embedded intelligence. Researchers and practitioners in the field should explore this architecture further, investigating applications ranging from interactive wearables to always-on environmental sensing, with an eye toward both performance and sustainability.