Heterogeneous Integration of Sensing & Processing

For many applications that require analog frontend sensing, such as CMOS image sensors and medical transducers, the deluge of generated raw data to improve sensing resolution is causing heavy burdens on the backend data processing and buffering. We explore advanced 2.5-D/3-D integration technologies to improve the latency and energy cost of processing sensor data.

Research Directions:

• Novel sensing and memory circuits for in-/near-sensor processing
• AI-driven heterogeneously integrated sensor fusion
• Modeling and benchmarking of interconnect technologies for sensor data delivery

We have demonstrated a 3-D integrated CIS with near-pixel buffering and computing. The footprint and signaling distance of the frontend CIS can be significantly reduced by placing buffer memories and near-pixel compute engines directly beneath the photodiodes. The PWM-based pixels transmit data via Cu-Cu hybrid bonding to the lower tier memory, which is designed with BEOL IWO 2T gain cell for buffering the frame data.