RRAM-based CAM combined with time-domain circuits for hyperdimensional computing.

Area of Science:

• Utilizes in-memory computing with Resistive RAM (RRAM) for efficient Content Addressable Memory (CAM) applications.
• Focuses on hardware acceleration for artificial intelligence (AI) tasks, particularly on resource-constrained platforms.

Background:

• Content addressable memory (CAM) is crucial for high-speed search and match operations in critical domains, demanding low power consumption.
• Resistive RAM (RRAM)-based in-memory computing offers a promising approach for efficient static CAM, especially for AI workloads.

Purpose of the Study:

• To present an XNOR-based RRAM-CAM integrated with a time-domain analog adder for efficient winning class computation.
• To demonstrate the effectiveness of the proposed design for hyperdimensional computing in MNIST classification.

Main Methods:

• Developed an XNOR-based RRAM-CAM that compares voltage and resistance operands to output similarity voltages.
• Employed time-domain analog addition to process similarity voltages, mitigating issues like voltage saturation, variation, and noise.
• Utilized a digital realization to identify the winning class based on the longest pulse width.

Main Results:

• The proposed RRAM-CAM design, implemented using 65 nm CMOS technology, achieved a total area of 0.0077 mm².
• Demonstrated low energy consumption of 13.6 pJ per 1k query within a 10 ns clock cycle.
• Achieved significant reductions in area (~31x) and energy consumption (~3x) compared to fully digital ASIC implementations.

Conclusions:

• The XNOR-based RRAM-CAM with a time-domain analog adder offers a highly efficient solution for AI tasks.
• The design presents a remarkable reduction in area and energy consumption compared to existing state-of-the-art RRAM designs.
• This approach is particularly beneficial for resource-constrained platforms requiring near real-time decision-making.