IMAX：高能效多级流水线粗粒度线性阵列及应用

The rapid advancement of artificial intelligence (AI) applications has driven a growing need for flexible and highly efficient hardware architectures. To address these demands, we propose IMAX, a novel coarse-grained linear array (CGLA) architecture that alternates cache memory and processing units in a linear structure to absorb irregular memory access latencies. This design achieves exceptional performance and energy efficiency. IMAX3 further enhances the architecture by introducing optimized communication, double buffering, and advanced sparse matrix multiplication (SpGEMM) techniques, delivering significant performance improvements. Real-time evaluations on the Xilinx VPK180 SoC demonstrate IMAX3’s remarkable capabilities: up to 503 times faster execution than GTX 1080Ti in SpGEMM and 10 times the energy efficiency of Jetson AGX Orin in FFT. Additionally, IMAX3 outperforms related architectures in matrix multiplication, achieving speeds approximately 23 times faster than STRELA and 61 times faster than RipTide, with substantial gains in energy efficiency. These results confirm IMAX3 as a cutting-edge, energy-efficient, and flexible hardware platform that surpasses traditional GPUs and other architectures, particularly in real-time processing and low-power environments. IMAX3 sets a new standard in hardware acceleration, making it an ideal solution for the evolving computational needs of AI-driven applications.