WAGONN: Weight Bit Agglomeration in Crossbar Arrays for Reduced Impact of Interconnect Resistance on DNN Inference Accuracy

Jeffry Victor, Dong Eun Kim, Chunguang Wang, Kaushik Roy, Sumeet Kumar GuptaIEEE Trans. Circuits Syst. I Regul. Pap 2025 · DOI

TL;DR

Deep neural network (DNN) accelerators using in-memory computing (IMC) with crossbar arrays are promising but suffer from accuracy degradation due to interconnect resistance in scaled technologies. This paper introduces WAGONN, a training-free technique that re-maps weight bits within crossbar arrays to mitigate these resistance effects, significantly boosting DNN inference accuracy with minimal hardware overhead.

What Problem Does This Paper Address?

Deep neural networks (DNNs) are revolutionizing many fields, but their computational demands are immense. Traditional Von Neumann architectures struggle with the power and performance bottlenecks of moving data between memory and processing units. In-memory computing (IMC) offers a solution by performing computations directly within memory arrays, particularly using crossbar structures. These crossbars compute matrix-vector multiplications (MVMs) by leveraging the relationship between input voltages and memory element conductances to generate currents that sum up on sense lines.

However, as semiconductor technologies scale to smaller nodes (like 7nm), interconnects become thinner and more resistive. This increased resistance leads to significant voltage drops (IR drops) within the crossbar array. These IR drops distort the currents generated by the memory cells, causing computational errors and a substantial drop in DNN inference accuracy. Existing solutions often involve complex training procedures or limited improvements. This paper tackles the critical challenge of maintaining DNN accuracy in deeply scaled IMC systems by addressing the detrimental impact of interconnect resistance.

Methods

The core innovation of this paper is WAGONN (Weight Agglomeration), a training-free technique designed to combat interconnect resistance in crossbar arrays. The fundamental idea is to re-map the weights stored in the crossbar array.

Here's how it works:

  • Understanding the Problem: In a crossbar array, cells further away from the Analog-to-Digital Converter (ADC) experience larger IR drops due to longer interconnect paths. This leads to a greater deviation in their output current from the ideal value.
  • Weight Agglomeration: WAGONN strategically groups "ON" weight bits (representing low-resistance states, LRS) together in rows closer to the ADC and "OFF" weight bits (high-resistance states, HRS) in rows further away. This agglomeration aims to:
  • Increase ON-cell currents: By placing LRS cells closer to the ADC, the impact of IR drops on their output current is reduced, bringing them closer to their expected values.
  • Decrease OFF-cell currents: For HRS cells (which still have some finite resistance and can produce non-zero currents due to leakage), WAGONN induces stronger source degeneration by placing them further from the ADC. This effectively reduces their non-zero current, further mitigating errors.
  • Input Re-mapping: To maintain the correct MVM functionality after weight re-mapping, the input activations also need to be dynamically re-mapped. This is achieved by calculating a "row-sum" (sum of weights in a row) and sorting rows based on this sum. A tracking vector stores the mapping information, allowing hardware to dynamically adjust input application to the re-arranged weight rows.
  • Synergy with PWA: The paper also explores combining WAGONN with Partial Wordline Activation (PWA), a technique that activates only a subset of rows per cycle. They propose a "Distributed PWA" (DPWA) approach that synergizes with WAGONN's weight distribution for even better accuracy.

Datasets

The paper evaluates the proposed WAGONN technique using the ResNet-20 model on the CIFAR-10 dataset. No new benchmarks or datasets were introduced.

Key Insights and Ideas

  • Interconnect resistance is a major bottleneck for scaled IMC: The paper strongly emphasizes that as technology nodes shrink, interconnect resistance becomes a dominant factor limiting the accuracy of IMC systems, far more than just geometric scaling alone.
  • Strategic weight placement is key: The core insight is that by intelligently rearranging weights, we can exploit the inherent non-uniformity of IR drops across the crossbar array to our advantage, rather than being a victim of it.
  • WAGONN benefits both ON and OFF states: The technique not only improves the accuracy of active computations (ON cells) but also helps suppress erroneous currents from inactive or high-resistance states (OFF cells).
  • Minimal hardware overhead: A significant finding is that WAGONN achieves substantial accuracy gains with very little added hardware complexity, making it a practical solution.
  • Synergy with existing techniques: The paper demonstrates that WAGONN can be effectively combined with other IMC optimization techniques like PWA, leading to further improvements.

Key Results

  • SRAM Accuracy Improvement: For 128x128 8T-SRAM crossbar arrays at 7nm technology, WAGONN boosts inference accuracy from a baseline of 47.78% to 83.5%. This is a remarkable 35.72% increase.
  • FeFET Accuracy Improvement: For 128x128 FeFET crossbar arrays, WAGONN improves accuracy from 84.85% to 87.53%.
  • Combined with DPWA: The combination of WAGONN and Distributed PWA (DPWA) on a 128x128 SRAM array achieves an accuracy of 88.83%, significantly outperforming PWA alone.
  • Hardware Overheads:
    • Energy: Minimal, at most 0.6% increase.
  • Latency: Approximately 1% when using 1 ADC per crossbar.
  • Area: Approximately 16% when using 1 ADC per crossbar. These overheads are largely dominated by ADCs, making the additional circuitry for WAGONN relatively small.
  • Robustness to Variations: WAGONN consistently provides higher inference accuracy than the baseline design even in the presence of device-to-device variations.

Limitations and Future Work

  • Dynamic Input Re-mapping Complexity: While the weight re-mapping is a one-time cost, the dynamic input re-mapping requires additional hardware (Input Re-mapping Unit - IRU) which adds to the overhead. The paper acknowledges that using all-to-all muxes for this would incur significant latency and routing overheads.
  • Granularity of Re-mapping: The current work focuses on independent re-mapping for each crossbar. The authors suggest that performing swaps at a higher granularity (across multiple crossbars) could reduce hardware overhead and simplify dynamic input re-mapping, though this might reduce effectiveness. This is identified as a future research direction.
  • Future Work: The paper suggests exploring the benefits of higher granularity re-mapping and analyzing its impact on parallelism. Further investigation into the effectiveness of WAGONN in conjunction with other techniques and for different DNN architectures is also implied. The authors also note that ADC conversion time and energy increase non-linearly with precision, suggesting that WAGONN's overheads might be even lower for larger FeFET arrays due to this effect.

Summary updated

BibTeX
@article{victor2025wagonn,
  author = {Jeffry Victor and Dong Eun Kim and Chunguang Wang and Kaushik Roy and Sumeet Kumar Gupta},
  title = {{WAGONN: Weight Bit Agglomeration in Crossbar Arrays for Reduced Impact of Interconnect Resistance on DNN Inference Accuracy}},
  journal = {IEEE Trans. Circuits Syst. I Regul. Pap.},
  year = {2025},
  doi = {10.1109/tcsi.2025.3531758},
}
Figure