What is it about?

Modern AI chips perform billions of multiply-accumulate operations every second, and these operations consume a large amount of energy. A common way to save power is to run the hardware at a lower supply voltage, but this often causes timing errors when circuits cannot finish their calculations fast enough. Even a small number of such errors can severely reduce the accuracy of neural-network models. Our work introduces EPIC, a lightweight technique that predicts when timing errors are likely to happen inside a multiply-accumulate (MAC) unit—the main building block of AI accelerators. Instead of monitoring many internal signals, which would add extra hardware cost, EPIC automatically identifies the smallest and most effective set of signals that can be used to predict errors with high accuracy. When a risky situation is detected, the MAC can take corrective action to ensure the output remains reliable. This approach allows AI accelerators to operate safely at much lower voltages, reducing power consumption without sacrificing accuracy. In our experiments, EPIC-enabled MAC arrays preserved correct inference results even when the voltage was reduced by more than 20%. This makes EPIC a practical and energy-efficient solution for future low-power AI hardware in mobile devices, edge computing, and large-scale data centers.

Featured Image

Why is it important?

This work is important because today’s AI systems are rapidly moving into power-limited environments such as mobile devices, edge sensors, and data-center accelerators. As models grow larger and hardware becomes more energy-hungry, lowering the supply voltage is one of the few practical ways to reduce power consumption. However, this normally causes timing errors that can seriously damage the accuracy of AI applications. Our method offers a timely solution by enabling safe operation at much lower voltages without redesigning the entire chip or adding expensive hardware. EPIC predicts and corrects timing errors using only a small set of internal signals, making it both lightweight and highly effective. By allowing AI hardware to maintain high accuracy while consuming significantly less power, this work can benefit a wide range of real-world systems—from battery-powered devices to large-scale cloud infrastructure—helping developers build faster, greener, and more reliable AI technology.

Perspectives

From my perspective, this work reflects a practical challenge I repeatedly observed when designing low-power AI hardware: power budgets are shrinking, but accuracy requirements keep rising. Lowering the supply voltage is an attractive solution, yet even small timing errors can ruin a model’s output. I wanted to find a technique that makes voltage underscaling genuinely safe, not just theoretically possible. Developing EPIC showed me that reliable error prediction does not require heavy or complicated hardware if the right internal signals are chosen. I hope this work encourages more energy-efficient designs that do not compromise correctness, and I believe the ideas here can inspire future accelerators to push even further toward power-aware intelligence.

Tongjing Wu
Shanghai Jiao Tong University

Read the Original

This page is a summary of: EPIC: Error PredIction and Correction for Power-Efficient Voltage Underscaling Multiply-Accumulate Unit, June 2025, Institute of Electrical & Electronics Engineers (IEEE),
DOI: 10.1109/dac63849.2025.11132540.
You can read the full text:

Read

Contributors

The following have contributed to this page