Analog In-Memory Computing (IMC) promises to bypass the traditional von Neumann bottleneck by performing matrix-vector multiplication (VMM) directly within memory arrays. In an ideal crossbar array, computation happens instantaneously at the location of data: input vectors are encoded as voltages ($V_i$), weights are stored as device conductances ($G_{ij}$), and according to Ohm’s Law, the current through each cell is the product of the two ($I_{ij} = V_i \cdot G_{ij}$). Kirchhoff’s Current Law then naturally sums these currents along shared columns to output the final vector product ($\sum I_{ij}$).
However, physical hardware departs from this ideal. The metal wires connecting these devices are not perfect conductors; they possess finite, unintended parasitic interconnect resistance. As analog arrays scale up to handle deep neural networks and high-resolution inputs, this parasitic resistance introduces severe signal degradation, creating a fundamental physical barrier to scalability.
The Mechanism of Interconnect Voltage Drops
To understand why parasitic resistance corrupts analog computation, we must look at how voltage propagates across a physical crossbar. A long metal wire acts as a distributed network of tiny resistors. As current flows down a row or column, it encounters this wire resistance, causing a cumulative voltage drop ($V = I \cdot R_{wire}$) at every successive node.
Consequently, devices located far from the voltage drivers do not receive the intended input voltage $V_i$. Instead, they experience a degraded, position-dependent voltage. As single-author Runze Han explains in A Novel Convolution Computing Paradigm Based on NOR Flash Array With High Computing Speed and Energy Efficiency, this cumulative resistance leads to parasitic voltage drops that directly degrade the precision and accuracy of analog calculations.
The further a signal must travel across an array, the more distorted it becomes. This spatial dependency was observed early on by Christof Koch et al. in Object-Based Analog VLSI Vision Circuits, where the authors noted that maintaining consistent voltage gradients becomes increasingly difficult as the physical distance between nodes grows.
Ideal Crossbar:
V_in ------> [G] ------> [G] ------> (Constant Voltage)
Real Crossbar:
V_in --(r)-- [G] --(r)-- [G] --(r)-> (Degraded Voltage due to r_wire)
The High-Conductance Paradox
A highly counterintuitive result of parasitic interconnect resistance is that choosing "better," lower-resistance (high-conductance) memory devices can actually destroy system accuracy.
Intuitively, one might assume that highly conductive memory elements (like Magnetic Random-Access Memory, or MRAM) would improve signal propagation. However, Ohm's law dictates that higher conductance ($G$) draws more current ($I$) for a given voltage. Because the parasitic voltage drop along the wire is directly proportional to the total current flowing through it ($V_{drop} = I_{total} \cdot R_{wire}$), high-conductance devices exponentially worsen the interconnect voltage drop.
Single-author Saion K. Roy details this phenomenon in Energy-Accuracy Trade-Offs for Resistive In-Memory Computing Architectures. The author demonstrates that high-conductance technologies suffer from severe signal degradation in high-density configurations. In contrast, lower-conductance devices (such as Ferroelectric FETs or Resistive RAM) inherently limit the total current, thereby preserving the Signal-to-Noise-plus-Distortion Ratio (SNDR) across the array.
Indranil Chakraborty et al. reached a similar conclusion in GENIEx: A Generalized Approach to Emulating Non-Ideality in Memristive Xbars using Neural Networks, confirming that using devices with lower "ON" resistance paradoxically degrades system accuracy by drawing higher currents that exacerbate parasitic effects.
The Non-Linear Scaling Wall
Why can we not simply calibrate or calculate away these voltage drops? The challenge lies in the fact that parasitic resistance effects do not scale linearly with array size.
As an array expands to accommodate larger workloads, two things happen simultaneously: the interconnect wires grow longer (increasing $R_{wire}$), and more devices are placed in parallel (increasing $I_{total}$). This compounding effect causes the total signal distortion to rise non-linearly.
The authors of Examining the Role and Limits of Batchnorm Optimization to Mitigate Diverse Hardware-noise in In-memory Computing, Abhiroop Bhattacharjee et al., showed that this non-linear rise in parasitic resistance creates hardware errors that quickly exceed the corrective capacity of standard architectural components, such as batch normalization layers. This creates what Suhas Kumar et al. describe in Improved Hopfield Network Optimization using Manufacturable Three-terminal Electronic Synapses as a "scaling wall," where increasing the array size leads to prohibitive computational errors.
Furthermore, this degradation restricts energy efficiency. As argued by Mrigank Sharad et al. in Ultra Low Power Associative Computing with Spin Neurons and Resistive Crossbar Memory, parasitic voltage drops constrain our ability to scale down operating voltages, because lower signal levels become completely swallowed by the noise induced by wire resistance.
Mitigating Interconnect Non-Idealities
To build viable, large-scale analog accelerators, researchers employ a mix of hardware-aware algorithms and alternative computing domains to bypass the constraints of wire resistance.
Hardware-Aware Training and Mapping
Instead of treating the crossbar as an ideal mathematical matrix, compilers can model the physical layout of the chip during training or deployment. For instance, Abhiroop Bhattacharjee et al. in Examining and Mitigating the Impact of Crossbar Non-idealities for Accurate Implementation of Sparse Deep Neural Networks proposed crossbar-column rearrangement and Weight-Constrained-Training (WCT). By strategically mapping weights and rearranging columns, these techniques redistribute the current draw across the array, preventing localized current bottlenecks and minimizing the worst-case parasitic voltage drops.
Shifting to the Charge Domain
Another paradigm shift involves moving away from current-based accumulation entirely. Because current-based architectures require continuous charge flow (which triggers Ohmic voltage drops), they are inherently vulnerable to wire resistance.
To circumvent this, Zuopu Zhou et al. developed a charge-domain architecture detailed in Charge-domain content addressable memory based on ferroelectric capacitive memory for reliable and energy-efficient one-shot learning. By performing computations using capacitive charge sharing rather than steady-state currents, the system minimizes Ohmic losses. However, the authors note that this shifts the physical bottleneck from parasitic resistance to parasitic capacitance along long interconnect lines, proving that physical scaling limits remain a primary challenge in neuromorphic engineering.
Key Foundational Work
A Charge-Based Parallel Analog Vector Quantizer (1994) — contributed 1 concept elements to this direction.
Object-Based Analog VLSI Vision Circuits (1992) — contributed 1 concept elements to this direction.