IMPACT: In-Memory ComPuting Architecture based on Y-FlAsh Technology for Coalesced Tsetlin machine inference.

Area of Science:

• Computer Engineering
• Artificial Intelligence
• Materials Science

Background:

• Traditional von Neumann architecture faces limitations with increasing data demands for machine learning (ML).
• In-memory computing (IMC) offers a promising alternative by integrating data storage and processing, reducing latency and energy consumption.
• Y-Flash memory devices provide a high-yield, non-volatile, low-power solution for advanced computing applications.

Purpose of the Study:

• To introduce the In-Memory comPuting architecture based on Y-FlAsh technology for Coalesced Tsetlin machine inference (IMPACT).
• To leverage Y-Flash devices for efficient implementation of the novel coalesced Tsetlin machine (CoTM) ML algorithm.
• To demonstrate the performance and energy efficiency of the IMPACT architecture for ML inference.

Main Methods:

• Fabrication of Y-Flash memory devices on a 180 nm CMOS process.
• Development of the IMPACT architecture utilizing Y-Flash arrays for computational crossbars.
• Implementation of the CoTM algorithm, which employs Tsetlin automata for stochastic Boolean feature selection.
• Validation of the IMPACT architecture on the MNIST dataset.

Main Results:

• The IMPACT architecture achieved [Formula: see text] accuracy on the MNIST dataset.
• Demonstrated significant energy efficiency improvements compared to existing technologies: 2.23x over CNN-based ReRAM, 2.46x over neuromorphic NOR-Flash, and 2.06x over DNN-based PCM.
• Y-Flash devices exhibited high yield, non-volatility, and low power consumption.

Conclusions:

• IMPACT provides an energy-efficient and high-performance solution for ML inference, addressing the limitations of traditional architectures.
• The Y-Flash based IMC architecture is well-suited for modern ML applications requiring substantial data processing.
• This work highlights the potential of novel memory technologies and algorithms for future secure computing platforms.