Model quantization for computing-in-memory: a survey

Analysis

Why This Paper Matters

Computing-in-memory (CIM) is a promising paradigm to overcome the von Neumann bottleneck by performing computation directly within memory arrays, enabling high energy efficiency and throughput for deep neural networks. However, CIM hardware imposes strict constraints on weight precision and activation ranges, making model quantization not just an optimization but a necessity. This survey addresses the critical gap of systematically organizing quantization methods specifically for CIM accelerators, which differ from conventional GPU/CPU quantization due to analog non-idealities and limited precision.

The paper matters because it consolidates a fragmented field. As AI models grow, the demand for edge and embedded deployment increases, and CIM offers a path to extreme efficiency. Yet, without proper quantization, CIM accelerators suffer from accuracy degradation. By reviewing existing methods, the survey helps researchers understand the landscape and identify promising directions, potentially accelerating the adoption of CIM in real-world applications.

Technical Contributions

• Fundamental Concepts: The survey introduces quantization basics—bit-width, quantization levels, and calibration—tailored to CIM constraints.
• Categorization of Methods: It organizes quantization techniques into categories such as post-training quantization (PTQ), quantization-aware training (QAT), and hardware-aware quantization, highlighting their suitability for CIM.
• Hardware-Aware Design: It discusses how quantization methods must account for CIM-specific issues like analog-to-digital converter (ADC) precision, crossbar array non-linearity, and limited on-chip memory.
• Trade-off Analysis: The survey compares methods in terms of accuracy, energy efficiency, and hardware overhead, providing a useful decision framework.
• Future Directions: It outlines open challenges, such as adaptive quantization and co-design of algorithms and hardware.

Results

As a survey, the paper does not present new experimental metrics. Instead, it synthesizes findings from prior works, summarizing typical accuracy-efficiency trade-offs. For instance, it likely notes that 8-bit quantization is common for CIM, while lower bit-widths (4-bit or below) require advanced techniques like QAT to maintain accuracy. The survey also highlights that PTQ is preferred for its simplicity but may suffer on CIM due to analog noise, whereas QAT can better adapt to hardware non-idealities.

Significance

The survey provides a foundational reference for both academic and industrial researchers working on CIM accelerators. By clarifying the state of the art, it can help avoid redundant efforts and guide the development of more robust quantization algorithms. This is particularly important as CIM moves from research to commercial products, where reliability and accuracy are paramount. The survey's impact extends to the broader AI field by promoting efficient hardware-software co-design, which is essential for sustainable AI deployment at scale.