Deep random forest with ferroelectric analog content addressable memory

Area of Science:

• Artificial Intelligence
• Computer Engineering
• Materials Science

Background:

• Deep Random Forest (DRF) models offer competitive accuracy and interpretability for edge intelligence tasks, similar to Deep Neural Networks (DNNs).
• Existing hardware accelerators for DRF lag behind DNN counterparts, particularly in efficiently handling the crucial branch-split operations at decision nodes.
• Ferroelectric materials offer unique properties for novel memory and computing architectures.

Purpose of the Study:

• To propose and demonstrate an efficient hardware accelerator for Deep Random Forest (DRF) algorithms.
• To leverage ferroelectric analog content addressable memory (ACAM) for accelerating DRF's decision-making processes.
• To evaluate the performance, energy efficiency, and scalability of the proposed DRF accelerator architecture.

Main Methods:

• Implementation of DRF using associative searches within ferroelectric analog content addressable memory (ACAM) cells.
• Design of an ultra-compact ACAM cell utilizing two ferroelectric field-effect transistors (FeFETs) to store decision boundaries as analog polarization states.
• Development of a DRF accelerator architecture and a methodology for mapping DRF models onto ACAM arrays.
• Experimental and simulation-based validation of the FeFET ACAM DRF's functionality, characteristics, scalability, and robustness against device non-idealities.

Main Results:

• The proposed FeFET ACAM DRF accelerator achieves significant improvements in energy efficiency (∼10^6×/10×) and latency (∼10^6×/2.5×) compared to state-of-the-art CPU and ReRAM-based DRF hardware implementations.
• The ACAM cell enables energy-efficient branch-split operations by utilizing analog polarization states in FeFETs.
• The architecture demonstrates robustness against FeFET device non-idealities, ensuring reliable operation.

Conclusions:

• Ferroelectric ACAM provides a promising pathway for realizing ultra-compact and energy-efficient hardware accelerators for Deep Random Forest (DRF) algorithms.
• The proposed FeFET ACAM DRF architecture offers substantial performance gains, addressing the current limitations in DRF hardware acceleration for edge intelligence.
• This work paves the way for next-generation edge AI systems with enhanced computational capabilities and reduced power consumption.