Thursday 06 March 2025
Scientists have made a significant breakthrough in developing a new compiler framework that can efficiently accelerate deep neural networks (DNNs) on resource-constrained crossbar-array-based processing-in-memory (PIM) accelerators. The framework, called COMPASS, is designed to optimize the performance of DNNs by minimizing memory access and energy consumption while ensuring high throughput.
Traditionally, PIM accelerators have been limited in their ability to process large-scale neural networks due to their restricted memory capacity. To address this challenge, researchers have developed various partitioning schemes that divide the network into smaller segments, allowing each segment to be processed within the limited memory footprint of the accelerator. However, these approaches often result in suboptimal performance and energy efficiency.
COMPASS tackles this problem by introducing a novel algorithm that dynamically optimizes the partitioning of DNNs based on their characteristics, such as layer topology and data dependencies. The framework takes into account the unique constraints of PIM accelerators, including limited memory capacity and high write costs, to minimize memory access and energy consumption.
The COMPASS compiler uses a genetic algorithm (GA) to search for an optimal partitioning scheme that balances the trade-off between throughput and energy efficiency. The GA iteratively refines the partitioning scheme over multiple generations, adapting to changing network conditions and hardware constraints.
Experimental results demonstrate the effectiveness of COMPASS in accelerating DNNs on PIM accelerators. Compared to traditional partitioning schemes, COMPASS achieves higher throughput and better energy-delay product (EDP) in diverse workload settings. The framework is particularly well-suited for resource-constrained PIM accelerators, such as those using SRAM or ReRAM technologies.
The development of COMPASS has significant implications for the widespread adoption of DNNs in edge AI applications, where power efficiency and scalability are crucial. By enabling efficient acceleration of large-scale neural networks on PIM accelerators, COMPASS can help reduce energy consumption and improve performance in real-world applications.
In a nutshell, COMPASS is a powerful compiler framework that optimizes the performance of deep neural networks on resource-constrained crossbar-array-based processing-in-memory accelerators. Its ability to dynamically adapt to changing network conditions and hardware constraints makes it an attractive solution for edge AI applications where power efficiency and scalability are paramount.
Cite this article: “COMPASS: A Compiler Framework for Efficiently Accelerating Deep Neural Networks on Resource-Constrained PIM Accelerators”, The Science Archive, 2025.
Compiler, Deep Neural Networks, Processing-In-Memory, Accelerators, Genetic Algorithm, Memory Access, Energy Consumption, Edge Ai, Resource-Constrained, Optimization Framework
