Dual-domain architecture shows almost 40 times higher energy efficiency for running neural networks

Many conventional computer architectures are ill-equipped to meet the computational demands of machine learning-based models. In recent years, some engineers have thus been trying to design alternative architectures that could be better suited for running these models.

Some of the most promising solutions for running artificial neural networks (ANNs) are so-called compute-in-memory (CIM) systems. These systems are designed to process and store data in a single hardware system, which can reduce power consumption and boost the performance of ANN-based models.

CIM systems are divided into two broad categories: digital CIM (DCIM) and analog CIM (ACIM) systems. In a paper published in Nature Electronics, researchers at Tsinghua University introduce a new dual-domain ACIM system that could run ANNs more efficiently, while also boosting their performance in general inference tasks.

“Research on ACIM for neural networks computations has gained significant attention, but progress has mostly been in classification tasks (e.g., image classification),” Ze Wang, first author of the paper, told Tech Xplore.

“To date, ACIM still faces significant challenges in handling complex regression tasks (e.g., YOLO for object detection), due to two key limitations. The first is high computational noise and the second consists of floating-point (FP) data compatibility issues, which make it unsuitable for regression tasks typically requiring high-precision FP computation.”

To overcome the challenges typically associated with ACIM systems, Wang and his colleagues designed a new hybrid architecture that combines high-precision FP-compatible digital computing with good energy efficiency. Their proposed architecture was found to be promising for running neural networks, also allowing these networks to complete general inference tasks.

“ACIM excels in performing highly parallel and energy-efficient matrix multiplications, a fundamental operation extensively used in neural networks, offering significant advantages over digital computing,” explained Wang.

“ACIM operates on Kirchhoff’s current law and Ohm’s law to perform voltage-based multiplication and current-based summation directly in the analog domain. However, fabrication variations introduce computational noise, affecting accuracy in neural network computing.”

In addition to being prone to the introduction of noise, which impairs the accuracy of ANNs, ACIM systems were so far primarily applied to tasks that entail linear multiplication and addition. In contrast, they were so far found to be incompatible with FP computations, more complex mathematical operations that are performed on so-called FP numbers (i.e., computational representations of numbers).

“For the first time, we demonstrate a fully-hardware-implemented multi-target and multi-class object detection task, YOLO (You Only Look Once), on a real ACIM system,” wrote Wang. “This represents a significant milestone, extending ACIM’s capabilities beyond merely supporting classification tasks to a full support of general neural network inference with FP dataflow.”

Wang and his colleagues evaluated their architecture in a series of tests and found that it exhibited a remarkable energy efficiency, which was 39.2 times higher than that achieved by common FP-32 multipliers running ANN-based models. The researchers also created a prototype memristor-based computing system based on their design, which was found to attain a high precision (2.7 times higher on average than that of pure ACIM systems).

This recent study could inspire the development of other hybrid ACIM architectures, which could be better suited for running machine learning-based models on complex computing tasks. Meanwhile, Wang and his colleagues plan to continue building on their architecture, to further improve its precision and energy efficiency.

“There is still ample room for further exploration and improvement of our current work,” added Wang.

“Our future research will focus on the co-design and optimization of architecture, algorithms, and hardware to enhance the energy efficiency and accuracy of hybrid analog-digital computing systems. Ultimately, we aim to support a broader range of neural network computations and unlock new application scenarios.”

© 2025 Science X Network