High-performance near-memory processing architecture for data-intensive applications

Near-memory processing (NMP) offers a transformative approach to computing architecture by positioning processing units close to memory, thereby reducing data transfer delays and minimizing energy consumption. This paper introduces an innovative NMP architecture designed to bridge the gap between data storage and computation, significantly enhancing system performance for data-intensive applications. By embedding processing elements directly within memory arrays, the proposed design achieves notable improvements in key performance metrics. Experimental results demonstrate a latency reduction of over 40%, with matrix multiplication latency decreasing from 120.5 ms in traditional architectures to 70.4 ms. Energy consumption is reduced by nearly 50%, with workloads like matrix multiplication showing a drop from 25.6 J to 12.8 J. Additionally, the architecture achieves throughput gains of up to 100%, as seen in data analytics workloads where throughput increases from 48.1 GOPS to 96.5 GOPS. These results emphasize the architecture's ability to enhance computational efficiency and scalability, making it particularly advantageous for applications in artificial intelligence, scientific research, and big data analytics. This study underscores the potential of NMP to redefine high-performance computing by restructuring traditional data processing paradigms.