Analog In-Memory Computing (IMC) using memristive or Resistive RAM (RRAM) crossbar arrays offers a fundamentally different way to compute neural network workloads. Instead of shuttling weights from memory to a digital processor, crossbars perform Vector-Matrix Multiplication (VMM) directly inside the memory array. They achieve this by leveraging two physical laws: Ohm’s Law ($I = V \cdot G$), where input voltages multiply with device conductances (weights), and Kirchhoff’s Current Law, which sums the resulting currents along a shared column wire.
When a neural network is dense, this architecture is incredibly efficient. Every single memristor at every intersection of the grid participates in the computation, maximizing throughput and minimizing energy overhead.
However, modern deep learning has evolved away from purely dense structures. To run models on edge devices, researchers frequently employ structural innovations like depthwise convolutions, up-sampling deconvolutions, or network pruning. This creates a fundamental conflict: the rigid, two-dimensional physical layout of a memristor crossbar is structurally mismatched with the sparse connectivity patterns of modern neural networks.
The Structural-Coupling Problem
To understand why sparsity breaks down in analog crossbars, we must look at how weights are physically mapped to hardware. In a digital processor, a zero-valued weight means a computation can be skipped or bypassed using gating logic. In an analog crossbar, however, every weight is rigidly tied to a physical crosspoint. This challenge is known as the structural-coupling problem.
If a neural network weight is zero, the corresponding memristor must be programmed to its lowest conductance state. But as Fangxin Liu et al. observe in their work on the Sparse-Multiplication-Engine, these zero-valued cells cannot easily be bypassed or compressed out of the array. They still occupy physical silicon area and can still consume leakage power. Because standard crossbars lack the ability to dynamically aggregate sparse data across different physical rows or columns, unutilized hardware real estate becomes locked up.
Dense Mapping (Efficient) Sparse/Diagonal Mapping (Inefficient)
In1 In2 In3 In1 In2 In3
| | | | | |
V1 --W----W----W--> Out1 V1 --W----0----0--> Out1
| | | | | |
V2 --W----W----W--> Out2 V2 --0----W----0--> Out2
| | | | | |
V3 --W----W----W--> Out3 V3 --0----0----W--> Out3
(Every cell computes) (Zeros waste physical area)
As highlighted by Aayush Ankit et al. in the design of RESPARC, this spatial misalignment leads to severe underutilization of crossbar rows. When mapping convolutional layers to fixed-size crossbars, the local nature of the operations means the connectivity matrices are inherently sparse, diluting the energy efficiency gains that make memristive computing attractive in the first place.
The Worst Offenders: Depthwise Convolutions and Deconvolutions
This structural mismatch is most acute in two common architectural layers: depthwise convolutions and deconvolutions.
Depthwise Convolutions
In a standard convolutional layer, every input channel interacts with every output channel, creating a dense matrix that maps cleanly to a crossbar grid. In contrast, a depthwise convolution applies a single filter per input channel. Mathematically, this decouples the channels, transforming what was a dense matrix into a series of isolated, diagonal-only operations.
When you force this diagonal pattern onto a standard two-dimensional analog crossbar, the result is catastrophic for area efficiency. As Gianmarco Ottavi et al. explain in their analysis of end-to-end IMC inference, the one-to-one channel mapping forces the hardware to allocate an oversized physical footprint relative to the actual parameter count. To maintain the correct mathematical operations, vast regions of the crossbar must be padded with zeros. This structural mismatch forces a harsh trade-off between wasting significant hardware area and suffering from drastically reduced throughput.
Deconvolutions
Deconvolutional (or transposed convolutional) layers, frequently used in generative models and segmentation networks, suffer from a similar mapping penalty. To up-sample an image, deconvolutions insert zeros between the pixels of the input feature maps.
As Zichen Fan et al. demonstrate in their work on RED, feeding these up-sampled, zero-padded inputs into a standard RRAM crossbar causes massive energy and throughput inefficiencies. The hardware blindly applies voltages for these zero inputs, executing redundant computations across the fixed crossbar structure.
Strategies for Bridging the Mismatch
To resolve this tension between sparse software and dense hardware, researchers are attacking the problem from three distinct angles: heterogeneous hardware architectures, hardware-aware pruning, and algorithmic redesign.
1. Heterogeneous and Hybrid Architectures
Rather than forcing analog crossbars to perform operations they are fundamentally unsuited for, one prominent direction is to build hybrid computing clusters.
In architectures proposed by both Gianmarco Ottavi et al. and Angelo Garofalo et al., the system is split into specialized units. Standard, dense matrix-vector multiplications are offloaded to the analog IMC cores where they thrive. Meanwhile, highly sparse or structurally awkward operations—like depthwise convolutions—are routed to a dedicated digital accelerator or a cluster of programmable RISC-V cores. By utilizing digital hardware for these specific layers, the chip preserves its overall area efficiency while bypassing the Amdahl bottleneck that plagues pure analog accelerators.
2. Hardware-Aware and Pattern Pruning
If a network must be pruned to reduce its size, the sparsity must be introduced in a way that the physical crossbar can exploit. Standard, irregular pruning removes weights randomly, leaving a scattered distribution of zeros that provides no benefit to a rigid crossbar layout.
To fix this, Arash Fayyazi et al. introduced CSrram, a framework that enforces "structured predefined sparsity" on the network before training begins. By constraining the sparse connections to specific, automated layouts, the physical manufacturing and mapping processes can realize actual area and power savings.
Similarly, Songming Yu et al. proposed a pattern-pruning weight mapping scheme. By forcing sparse weights into regular, predictable patterns, they eliminate the need for complex digital indexing circuitry and allow the hardware to pack weights more densely into the RRAM cells.
Irregular Pruning (Bad for RRAM) Pattern/Structured Pruning (Good)
X . . X X X . .
. X . . X X . .
. . X X . . X X
X . . . . . X X
(Scattered zeros prevent (Blocks of zeros allow
crossbar compression) orderly hardware optimization)
3. Algorithmic Redesign
A third approach is to rethink the neural network architecture entirely, designing models from the ground up to be "RRAM-friendly."
Recognizing that depthwise convolutions are ill-suited for RRAM crossbars due to padding requirements, Wenyong Zhou et al. proposed a novel CNN architecture that avoids them altogether. Instead of relying on channel-expanding layers or depthwise operations to lower FLOPs, their architecture intelligently concatenates feature maps from earlier layers. This maintains a dense weight-to-cell mapping, ensuring high hardware utilization and competitive accuracy without wasting chip area on padding.
The Reliability Trade-off
As designers navigate these mapping schemes, they must also contend with a hidden physical trade-off: reliability.
In a comprehensive design space exploration, Corey Lammie et al. demonstrated that while sparse mapping schemes can significantly reduce power consumption by minimizing active devices, they simultaneously increase the network's sensitivity to hardware noise, device variability, and stuck-at-faults. Because a sparse mapping distributes the model's logic across fewer active components, each individual memristor becomes a critical point of failure. If a few devices suffer from manufacturing defects, there are fewer redundant paths to absorb the error, causing accuracy to degrade more rapidly than in a densely packed configuration.
Ultimately, overcoming the crossbar utilization mismatch is not just about compressing matrices or skipping zeros. It requires a careful balancing act between algorithmic structure, hardware topology, and the underlying physics of analog devices.
Key Foundational Work
RESPARC: A Reconfigurable and Energy-Efficient Architecture with Memristive Crossbars for Deep Spiking Neural Networks (2017) — contributed 3 concept elements to this direction.
RED: A ReRAM-based Deconvolution Accelerator (2019) — contributed 1 concept elements to this direction.
CSrram: Area-Efficient Low-Power Ex-Situ Training Framework for Memristive Neuromorphic Circuits Based on Clustered Sparsity (2019) — contributed 1 concept elements to this direction.