AIEAnalyzer: An Analytical Framework for Versal AI Engine Performance Analysis

Versal Adaptive Compute Acceleration Platform (ACAP) is a novel heterogeneous architecture characterized by powerful computation units, i.e., AI Engines (AIEs). Developing high-performance ACAP applications requires both accurate and efficient evaluation of various design points. Although the traditional cycle-accurate simulator provides highly precise performance evaluation, its considerable runtime overhead has become a major bottleneck in the design and optimization process. While existing solutions offer modeling frameworks from both analytical and learning-based perspectives, they still fall short of jointly satisfying the demands of accuracy, transferability, and ease of use. To settle this issue, this paper proposes AIEAnalyzer, an analytical performance modeling framework for Vitis ADF graph–based AIE designs. Specifically, AIEAnalyzer introduces a two-stage analytical method from the perspectives of intra-AIE computation and inter-AIE communication, which respectively capture the fine-grained execution patterns of the single AIE kernel and model the communication mechanisms among interconnected AIEs. Experimental results demonstrate that AIEAnalyzer achieves a mean error rate of 3.255% and a mean speedup factor of 594.4 \(×\) compared with simulation-based methods.