A groundbreaking computing paradigm shift promises to enhance the performance of devices such as smartphones and laptops twofold, all without necessitating the replacement of any existing components. By enabling various processing units, including GPUs, NPUs, and hardware accelerators, to operate concurrently rather than sequentially, systems can achieve up to double the processing speed while consuming 50% less energy.
Contemporary devices are equipped with diverse chips dedicated to different types of processing tasks. In addition to the central processing unit (CPU), devices feature graphics processing units (GPUs), hardware accelerators tailored for artificial intelligence (AI) tasks, and digital signal processing units for audio signal processing.
However, the conventional execution models for programs result in these components processing data from individual programs separately and sequentially, leading to slowed processing times. Information traverses from one unit to the next based on their efficiency in handling specific segments of code within a program, thus creating a bottleneck. With one processor completing its task before passing it on to the subsequent processor in line, efficiency diminishes.
To address this issue, scientists have introduced a novel framework for program execution known as “simultaneous and heterogeneous multithreading (SHMT).” In this approach, processing units operate in parallel, allowing them to tackle the same code region simultaneously instead of waiting for processors to address different code regions sequentially, depending on the workload and the component’s efficiency.
The team detailed this innovative approach in a paper published in December 2023 on the preprint server arXiv. By leveraging SHMT, devices can harness the collective power of their processing units more effectively, resulting in significantly improved performance without the need for hardware upgrades.