PROPHET: Predictive On-Chip Power Meter in Hardware Accelerator for DNN

• Published in: 2023 60th ACM/IEEE Design Automation Conference (DAC) pp. 1 - 6
• Main Authors: Peng, Jian, Liang, Tingyuan, Xie, Zhiyao, Zhang, Wei
• Format: Conference Proceeding
• Language: English
• Published: IEEE 09.07.2023
• Subjects: DNN accelerator; Memory management; Meters; on-chip power meter; pattern-based; Power demand; power prediction; Power system management; Runtime; System-on-chip; Voltage
• DOI: 10.1109/DAC56929.2023.10247979

On-chip power meters play a critical role in power management by generating timely and accurate power traces at runtime. However, both performance-counter-based and existing RTL-based on-chip power meters have difficulty in providing sufficient response time for fast power and voltage management scenarios. Additionally, they can be costly to implement for large-scale DNN accelerators with many homogeneous process elements. To address these limitations, this paper proposes PROPHET, a data-pattern-based predictive on-chip power meter targeting multiply-accumulate-based DNN accelerators. By sampling pre-defined data patterns during memory access, PROPHET can predict power consumption before it actually happens. In our experiments, PROPHET predicts power consumption dozens of clock cycles in advance, with a temporal resolution of 4 clock cycles and NMAE < 7% and area overhead < 2% for various systolic-array-based DNN accelerators. PROPHET has the potential to enable fine-grained power management and optimization for large-scale DNN accelerators, improving their energy efficiency.