A groundbreaking research from the Advanced Computer Architecture Laboratory at Shanghai Jiao Tong University, led by Professor Jiang Li and Assistant Professor Liu Fangxin, has introduced a novel SRAM-based in-memory computing architecture named ‘COMPASS’ – an innovation that promises to revolutionize brain-inspired computing. This development has been announced in a recent report by Machine Heart, a prominent platform for disseminating academic and technological content.
The IEEE/ACM International Symposium on Microarchitecture, commonly known as MICRO, alongside ISCA, HPCA, and ASPLOS, stands as one of the ‘Big Four’ conferences in the field of computer architecture. These conferences showcase the most advanced architectural achievements, serving as a beacon for cutting-edge international research. Notably, they have been the stage for numerous groundbreaking innovations from tech giants like Google, Intel, and NVIDIA in the semiconductor industry.
This year’s symposium received a total of 497 submissions, with only 113 papers, representing a 22% acceptance rate, being selected for inclusion. Among these, the introduction of COMPASS highlights the growing importance of brain-inspired computing, particularly with the rise of Spiking Neural Networks (SNNs) in the realm of artificial intelligence.
SNNs, known for their energy efficiency and high performance, have emerged as a promising alternative to traditional Deep Neural Networks (DNNs) in recent years. However, the pursuit of higher accuracy often comes at the cost of increased energy consumption and computational latency, posing challenges for their deployment on edge devices. The introduction of COMPASS addresses these issues by offering a new approach for efficient SNN hardware acceleration.
The Challenge and Breakthrough in Brain-Inspired Computing
While DNNs have excelled in computer vision, natural language processing, and speech recognition, their massive computational requirements lead to less-than-ideal energy efficiency. In contrast, SNNs leverage binary spike events, reducing computational demands through event-driven information processing. However, SNNs face performance degradation in time-sensitive tasks, necessitating hardware solutions that maintain energy efficiency while minimizing computation delay and memory usage.
COMPASS, the innovative SRAM-based in-memory computing architecture, capitalizes on the explicit sparsity of input spikes and the implicit sparsity of output spikes. By introducing a dynamic pulse speculation mechanism, the architecture significantly reduces redundant computations and enhances hardware resource utilization. Additionally, COMPASS employs a time-domain compression technique for both input and output spikes, further minimizing memory occupancy and enabling efficient parallel execution.
Details of the Innovative Architecture
At the heart of the COMPASS architecture lies its high-sparse utilization of the SRAM-based in-memory design. The irregularity and time dependency of spikes present a significant challenge for efficient parallel processing in traditional CIM architectures. The COMPASS team overcomes these obstacles by designing an architecture that efficiently exploits the irregular sparsity in both input and output spikes of SNNs.
By leveraging SNN’s pulse sparsity, COMPASS introduces dynamic pulse speculation patterns, which compress inference latency logarithmically, thus reducing unnecessary computations and hardware overhead. Adaptive speculation window scheduling and time pulse sparse representation further optimize the processing of input and output spikes, reducing memory usage and facilitating parallel execution.
Impressive Performance and Results
A comprehensive performance evaluation of COMPASS reveals remarkable results. Compared to existing SNN hardware accelerators, the architecture achieves a 24.4 times speedup in end-to-end acceleration, with a staggering 386.7 times reduction in energy consumption per inference. These outcomes not only validate COMPASS’s superiority in handling SNN tasks but also highlight its immense potential for real-world applications.
In a series of comparative experiments, COMPASS outperforms conventional DNNs and other SNN models, particularly in terms of energy efficiency when tackling complex tasks. This groundbreaking work underscores the significant potential of the COMPASS architecture in enabling efficient and energy-conscious computing.
In conclusion, the advent of the COMPASS architecture signifies a new chapter in brain-inspired computing, offering a promising solution to the challenges posed by SNNs. With its innovative use of SRAM-based in-memory computing and efficient spike processing, COMPASS paves the way for more energy-efficient and high-performance computing in the rapidly evolving field of artificial intelligence.
【source】https://www.jiqizhixin.com/articles/2024-08-26-2
Views: 2