Towards Unified Acceleration: Weight-Stationary GEMM on HPC-oriented Elastic CGRAs

Elastic coarse-grained reconfigurable arrays (CGRAs) are emerging accelerators that combine performance efficiency with computational flexibility, making them well suited for high-performance computing (HPC) applications. Recent studies have also explored their role in AI acceleration: some works integrate CGRAs into AI accelerators to handle nonlinear functions such as softmax and layer normalization, while others attempt to map general matrix multiplication (GEMM) kernels directly onto CGRAs. However, these approaches are fundamentally limited, focusing either on nonlinear/HPC tasks without efficient GEMM support, or on GEMM acceleration at the cost of sacrificing the computational capabilities for general purpose computing. In this paper, we present a double-buffer architecture with dedicated weight-uploading chains that significantly enhance the AI computational capabilities of elastic CGRAs. Our experimental results show that our approach achieves 3 × higher performance than existing methods for AI tasks while maintaining HPC application processing capabilities. This work demonstrates the feasibility of unified acceleration for both HPC and AI workloads, thereby expanding the applicability of elastic CGRAs.