Overcoming the Peripheral Bottleneck in Analog Computing

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2385 papers514 conceptsfrozen 2026-10-04

The promise of "beyond von Neumann" computing is seductive: by performing calculations directly within memory or using the speed of light, we could theoretically achieve orders of magnitude better energy efficiency than any modern GPU. Whether using resistive RAM (ReRAM), phase-change memory (PCM), or photonics, the core mechanism is often the same—leveraging physical laws (like Ohm’s Law or Kirchhoff’s Law) to perform massive matrix-vector multiplications in a single step.

However, a persistent "efficiency paradox" haunts the field. While the core computational elements—the memristors or optical neurons—consume almost no power, the system as a whole often struggles to beat digital baselines. The culprit is the peripheral bottleneck: the energy and area required to translate data between the analog world of the processor and the digital world of the computer.

The Mixed-Signal Tax

In a typical analog accelerator, the actual "math" happens in an analog crossbar. But to use this result, the system must convert it back to digital bits. This requires Analog-to-Digital Converters (ADCs) and Digital-to-Analog Converters (DACs).

As Akhilesh R. Jaiswal and Kaushik Roy have noted in their work on 8T SRAM dot-product engines, these converters often account for the vast majority of the total system power. In fact, Hanqing Zhu and Chenghao Feng point out that in hybrid photonic-electronic architectures, these interfaces can consume over 80% of the total energy budget.

The problem isn't just power; it's also physical space. High-precision ADCs are bulky. Linghao Song and Hai Li demonstrated with GraphR that while ReRAM crossbars are incredibly dense, the ADCs required to read them can limit the scalability of the entire chip. This creates a "power wall" where the benefits of analog computing are negated by the cost of talking to the rest of the system.

Mixed-Precision as a Compromise

One of the most successful strategies to bypass this bottleneck is mixed-precision memcomputing. Instead of trying to make the analog hardware perfectly accurate (which would require power-hungry, high-resolution ADCs), researchers use noisy, low-power analog cores for the heavy lifting and a digital processor for error correction.

Manuel Le Gallo and Abu Sebastian pioneered this approach, showing that you can solve complex linear equations or train deep neural networks by accumulating high-precision weights in digital memory while performing the bulk of the multiplications in analog PCM. This "best of both worlds" approach allows the system to tolerate the inherent noise of analog hardware without sacrificing the final result's precision.

Eliminating Active Components

Another frontier involves redesigning the peripheral circuitry itself. Traditional analog designs rely on "active" components like operational amplifiers (op-amps) to integrate signals. Kejie Huang found that these op-amps can consume over 95% of the energy in a computing-in-memory core.

By replacing these with "passive" components—like capacitor-based integrators—researchers have achieved staggering efficiency gains. Similarly, Unhyeon Kang has explored using Ovonic Threshold Switching to mimic biological dendrites, aiming to eliminate the need for external amplification entirely.

Hardware-Software Co-Design

If the ADC is the bottleneck, why not get rid of it? This is the logic behind hardware-aware co-design. By tailoring the neural network to the hardware's limitations, researchers can use much lower resolution converters.

For example, Yulhwa Kim and Jae-Joon Kim showed that by splitting neural network layers into smaller blocks, they could completely eliminate high-resolution ADCs in Binary Neural Network (BNN) accelerators. Mohsen Imani and Farinaz Koushanfar took this further with RAPIDNN, which maps operations into digital lookup tables within memory, bypassing the analog-to-digital conversion struggle entirely.

The Path to All-Analog Pipelines

The ultimate goal is to keep the data in the analog (or optical) domain for as long as possible. Kevin Portner and Till Zellweger recently demonstrated actor-critic reinforcement learning where weight updates and action selections happen "in-memory."

When the entire pipeline—from sensing to processing to acting—stays analog, the "tax" of conversion disappears. As Fuxi Cai and Suhas Kumar showed with their memristor Hopfield networks, treating hardware noise as a computational feature rather than a bug can lead to throughput-per-watt that is four orders of magnitude higher than digital equivalents.

The research direction is clear: the future of efficient AI isn't just about building faster analog "brains," but about minimizing the energy we waste just trying to talk to them.

Papers in this version

Showing 60 of 385 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 286 papers
  2. Version 1 · 2026-10-04 · 385 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_analog_to_digital_conversion_overhead_d97d86da,
  title = {analog-to-digital conversion overhead},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/analog-to-digital-conversion-overhead/rdr_d97d86da98614c8d939d3ce7d9217b30},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_overcoming_the_peripheral_bottleneck_in__d97d86da_v1,
  title = {Overcoming the Peripheral Bottleneck in Analog Computing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/overcoming-the-peripheral-bottleneck-in-analog-computing/rdr_d97d86da98614c8d939d3ce7d9217b30/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}