Researchers from the University of California, Riverside have proposed a new programming and execution model called Simultaneous and Heterogeneous Multithreading (SHMT) to better utilize the heterogeneous computing resources in modern systems.
As computers integrate more domain-specific accelerators like GPUs and AI chips along with CPUs, the conventional programming models cannot fully tap into the available processing power. They can only delegate code regions to the most efficient processor, leaving other resources idle.
SHMT allows different processor types to concurrently work on tasks from the same code region. To enable this, SHMT uses an abstraction layer and runtime system to divide computation into hardware-independent tasks and schedule them across heterogeneous processors. SHMT also employs mechanisms to ensure result quality from the mix of resources.
| Interested in getting a new PC or upgrading your existing one? Get in touch with us: https://pcx.com.ph/pages/contact-us |
The researchers built a prototype using a multi-core ARM CPU, Nvidia GPU, and Google Edge TPU. Tests on image processing, signal processing, and other applications showed 1.95x speedup on average compared to only using the GPU. SHMT also reduced energy consumption by 51% by leveraging the lower-power Edge TPU.
The paper introduces concepts like Virtual Operations (VOPs) and High-Level Operations (HLOPs) to bridge different hardware programming models. It also proposes scheduling policies like Quality-Aware Work-Stealing (QAWS) to balance workloads and maintain output quality.
Overall, the SHMT model and prototype demonstrate the performance benefits of concurrently utilizing available heterogeneous computing resources compared to conventional approaches. The concept can be extended to systems like smartphones and cloud servers.
Read the research paper: https://dl.acm.org/doi/pdf/10.1145/3613424.3614285
Interested in getting a new PC or upgrading your existing one?
Chat with a Personal Shopper:
- Viber Community Marketplace Sales Chat (8 AM to 10 PM | Mon to Sun):
- FB Chat (8 AM to 10 PM | Mon to Sun):
