TReCiM: Lower Power and Temperature-Resilient Multibit 2FeFET-1T Compute-in-Memory Design

Basic Information

• Paper ID: 2501.01052
• Title: TReCiM: Lower Power and Temperature-Resilient Multibit 2FeFET-1T Compute-in-Memory Design
• Authors: Yifei Zhou, Thomas Kämpfe, Kai Ni, Hussam Amrouch, Cheng Zhuo, Xunzhao Yin
• Category: cs.ET (Emerging Technologies)
• Publication Date: January 2025
• Paper Link: https://arxiv.org/abs/2501.01052

Abstract

Compute-in-Memory (CiM) represents a promising solution to address hardware challenges in artificial intelligence and Internet of Things applications, particularly in mitigating the "memory wall" problem. By employing non-volatile memory (NVM) devices within crossbar array architectures, CiM can efficiently accelerate critical operations in neural networks—multiply-accumulate (MAC) operations. Among various NVM technologies, ferroelectric field-effect transistors (FeFETs) are particularly well-suited for ultra-low-power CiM arrays due to their CMOS compatibility, voltage-driven write/read mechanisms, and high ION/IOFF ratios. Subthreshold operation of FeFETs can further minimize power consumption; however, it is susceptible to temperature drift, which degrades computational accuracy. This paper proposes TReCiM, an ultra-low-power, temperature-resilient multibit 2FeFET-1T CiM design that reliably executes MAC operations across a temperature range of 0°C to 85°C. Benchmark evaluations using the NeuroSim framework on VGG-8 neural networks and the CIFAR-10 dataset demonstrate that TReCiM achieves 91.31% accuracy when accounting for temperature drift effects, representing a 1.86% improvement over existing 1-bit 2T-1FeFET CiM arrays. Furthermore, the design achieves 48.03 TOPS/W energy efficiency at the system level, comparable to existing designs with smaller technology nodes.

Research Background and Motivation

Problem Background

1. Memory Wall Problem: Traditional von Neumann architectures face high power consumption and performance bottlenecks due to frequent data transfers between processing and storage units
2. Edge AI Requirements: AI and IoT applications demand substantial MAC operations with extremely stringent energy efficiency requirements
3. Limitations of Existing CiM: While CiM technology addresses the memory wall problem, existing designs fall short in temperature resilience and multibit storage capabilities

Research Motivation

1. FeFET Advantages: FeFETs offer CMOS compatibility, low leakage current, and high ION/IOFF ratios, making them suitable for ultra-low-power applications
2. Subthreshold Operation: Operating in the subthreshold region significantly reduces power consumption but increases temperature sensitivity
3. Limitations of Existing Solutions:
   • Existing temperature-resilient designs perform poorly at large array scales
   • Support only 1-bit storage, limiting storage density
   • 2T-1FeFET designs require additional discharge time, increasing latency

Core Contributions

1. Proposes TReCiM Architecture: The first temperature-resilient multibit 2FeFET-1T CiM design supporting 0°C-85°C temperature range
2. Innovative 2FeFET Clamping Structure: Implements temperature compensation utilizing complementary absolute temperature (CTAT) characteristics
3. Multibit Storage Capability: Leverages FeFET multi-level cell (MLC) characteristics for 2-bit and higher storage
4. System-Level Verification: Complete system-level evaluation and benchmarking based on NeuroSim framework
5. Performance Enhancement: Achieves 3× improvement in temperature resilience compared to existing solutions, with energy efficiency reaching 48.03 TOPS/W

Methodology Details

Task Definition

Design an ultra-low-power multibit CiM array capable of stable operation across a wide temperature range (0°C-85°C), supporting neural network MAC operations while maintaining high accuracy and energy efficiency.

Model Architecture

2FeFET-1T Cell Design


Core Components:

• M1, M2: Two FeFET devices forming a clamping structure
• M3: NMOS transistor serving as output control
• Control Signals: WL1, WL2 (word lines), DL (data line), BL (bit line), SL (source line)

Operating Principles:

1. Write Operation: Different voltages (±4V) applied via WL1 and WL2 to set FeFET states
2. Read Operation: MAC operations realized through WL voltage control
3. Temperature Compensation: Utilizes MOSFET CTAT characteristics and feedback mechanisms

Multibit Storage Implementation

• State '0': M2 in VTH1 state, M1 in VTH0 state
• State '1': Both M1 and M2 in VTH1 state (clamping configuration)
• State '2' and above: M1 in different VTH states, M2 off

Technical Innovations

1. 2FeFET Clamping Structure

When storing state '1':
- M1 and M2 form a voltage divider
- Intermediate node VS voltage is stably clamped
- Significantly reduces temperature drift impact

2. CTAT Temperature Compensation Mechanism

Subthreshold region MOSFET leakage current formula:

ID = I0 * exp(Vgs / (ξVT))
where VT = kT/q

Temperature feedback mechanism:

• Temperature increase → M1 leakage current increases → VS voltage rises → M3 output current increases
• However, due to CTAT characteristics, current increase is suppressed, minimizing output fluctuations

3. Array-Level Design

• 8-row × multi-column structure: Supports parallel MAC operations
• Flash ADC: Uses current-sensing amplifiers to reduce sensing delay
• Shared ADC: 8 columns share a single 3-bit ADC, balancing area and performance

Experimental Setup

Simulation Environment

• SPICE Simulation: Utilizes Intel FinFET model and Preisach FeFET compact model
• NeuroSim Framework: Modified to support subthreshold FeFET and temperature effects
• Technology Node: 45nm technology
• Supply Voltage: Subthreshold design Vdd=0.8V, saturation design Vdd=1.0V

Datasets

• Neural Network: VGG-8 architecture
• Dataset: CIFAR-10
• Network Structure: 6 convolutional layers + 2 fully connected layers
• Quantization: Hardware quantization using WAGE model

Evaluation Metrics

1. Temperature Resilience: Noise margin ratio (NMR) and minimum NMR values
2. Accuracy: Neural network inference accuracy
3. Energy Efficiency: TOPS/W (tera operations per watt)
4. Area: Chip area utilization
5. Throughput: Computation speed

Comparison Methods

• 1FeFET-1R: Basic single FeFET design
• 2T-1FeFET: Existing temperature-resilient design
• Other NVM: RRAM, PCM, and other technologies

Experimental Results

Main Results

Temperature Resilience Verification

• 1-bit TReCiM: NMRmin = 0.291 (0-85°C), NMRmin = 2.6 (20-85°C)
• Temperature Resilience Improvement: 3× improvement over 1FeFET-1R design
• Compared to 2T-1FeFET: 1.06× improvement

Neural Network Performance

Ablation Studies

Process Variation Impact

• Monte Carlo Simulation: 500 runs, σVT = 54mV
• State '1': 100% accuracy, variation only 3.89%
• State '2': Variation 20.8%
• State '3': Variation 17.1%

Temperature Characteristics Analysis

Temperature sensitivity of different storage states:

• State '1': Negligible temperature drift (clamping effect)
• State '2': Maximum fluctuation 32.9%
• State '3' and above: Reduced temperature sensitivity as VTH decreases

Case Study

In VGG-8 network, weight distribution:

• Weight '0': 27.2%
• Weight '1': 24.1%
• Weight '2': 23.5%
• Weight '3': 25.2%

Combined variation rate of 13.9%, achieving final inference accuracy of 91.31%.

Related Work

CiM Technology Development

1. NVM Technologies: Applications of ReRAM, PCM, FeFET in CiM
2. Subthreshold Computing: Power reduction with increased temperature sensitivity
3. Temperature-Resilient Design: Existing solutions primarily targeting TCAM and simple CiM structures

Current FeFET Research

1. Device Modeling: NCFET, Preisach, Monte Carlo models
2. Circuit Applications: Primarily focused on memory and simple compute units
3. Temperature Effects: Limited research addressing temperature impact on FeFET performance

Advantages of This Work

Compared to existing work, this paper is the first to achieve:

• Temperature-resilient CiM design with multibit storage
• System-level temperature effect modeling and verification
• Complete evaluation in neural network applications

Conclusions and Discussion

Main Conclusions

1. Technical Feasibility: 2FeFET-1T structure successfully achieves temperature resilience and multibit storage
2. Performance Advantages: Significantly improves temperature resilience while maintaining low power consumption
3. System Value: Validates design effectiveness in practical neural network applications

Limitations

1. Partial State Sensitivity: Storage states '2' and above exhibit certain temperature sensitivity
2. Process Dependence: Performance depends on FeFET device process maturity
3. ADC Overhead: Multibit designs require higher-precision ADCs, increasing area and power
4. Temperature Range: While covering 0-85°C, performance degradation occurs at extreme temperatures

Future Directions

1. Device Optimization: Further optimize FeFET device characteristics to reduce temperature sensitivity
2. Circuit Improvements: Explore temperature-resilient designs for higher bit-width storage
3. System Integration: Combine with on-chip temperature sensors for dynamic compensation
4. Application Extension: Validate design effectiveness in more AI application scenarios

In-Depth Evaluation

Strengths

1. Strong Innovation: First multibit temperature-resilient FeFET CiM design with novel technical approach
2. Solid Theory: Temperature compensation mechanism based on CTAT characteristics has firm physical foundation
3. Comprehensive Verification: Complete verification chain from device to system level
4. Practical Value: Demonstrates design value in practical neural network applications
5. Excellent Performance: Both energy efficiency and accuracy reach advanced levels

Weaknesses

1. Limited Temperature Compensation: Temperature compensation effectiveness is limited for high bit-width storage states
2. Increased Complexity: 2FeFET structure adds design complexity compared to single FeFET
3. Process Requirements: Demands high consistency in FeFET device characteristics
4. Scalability: Performance at larger array scales requires further verification

Impact

1. Academic Value: Provides new technical pathway for temperature-resilient CiM design
2. Industrial Significance: Important reference for edge AI chip design
3. Technology Advancement: Promotes FeFET technology development in CiM applications

Applicable Scenarios

1. Edge AI: Power-sensitive edge inference applications
2. IoT Devices: IoT environments with significant temperature variations
3. Mobile Computing: Mobile devices with extreme energy efficiency requirements
4. Industrial Control: Industrial applications requiring wide temperature range operation

References

The paper cites 42 relevant references covering CiM technology, FeFET devices, temperature effects, and neural network accelerators, providing solid theoretical foundation. Key references include NeuroSim framework, FeFET modeling, and related CiM design work.

Overall Assessment: This is a high-quality technical paper making important contributions to temperature-resilient CiM design. The paper features clear technical approach, comprehensive experimental verification, and significant implications for advancing FeFET applications in AI accelerators. While improvements are possible in certain aspects, the work overall represents important progress in the field.