Thrust 3 Closed-Loop Attention Control
Goals
Add adaptivity to:
•Feature extraction
Sensor front end so that we can extract more informative insight while consuming less energy and do so robustly across input and environmental condition changes.
Theme 3 Simulation Platforms
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Task 3.1 Sensor Trust Scoring
Task. 3.1 Measurement of Aging Effect of an Analog Computing -In-Memory Macro in 28 nm CMOS
3.2 Adaptive Feature Extraction and Reconstruction for Single-Photon LiDAR
Timestamp Measurement
Depth Reconstruction
Reflectivity Reconstruction
Ultrafast Modeling of Single-Photon LiDAR
Needs:
(1)Efficient data processing to handle large amounts of image sensor data.
(2)Low-power image compression and digitization techniques.
We Did:
•Designed an SC-based analog front end to enhance pixel voltage readout.
•Low-power data compression with an in-sensor analog DCT processor.
•Proposed a bidirectional ramp (BDR) ADC for DCT results.
Impact:
i.Achieved energy-efficient in-sensor compression with fine image quality.
ii.Reduced power consumption in image digitization using a BDR ADC.
iii.Up to 21x data compression with a high frame rate (252fps).
Task 3.3. Adaptive Feature Extraction – Hardware Design Study
Task 3.4 Hardware Acceleration
•We prototyped a sub-6mW IMC-based microcontroller that can support inference and training at high efficiency and low latency.
•The existing in-memory computing (IMC) macros suffer from low hardware utilization and degrades latency and energy efficiency for depthwise (DW) convolution layers.
•We develop a DW-IMC macro to improve the efficiency by 8X.
•We support an arbitrary gradient descent-based training model using existing IMC hardware
Our microcontroller achieves 4X to 22X EDP improvement over the state-of-the-art.
