APPLICATIONS OF TECHNOLOGY:
Applications that require high performance computing using convolutional neural networks (CNNs), like:
- Data mining
- Medical imaging
- Autonomous vehicles
- National defense/security
- Scientific research
- Control circuits for quantum computing
BENEFITS:
- More energy efficient than existing CNN hardware accelerators
- Smaller form factor than existing binary SFQ accelerators, resulting in higher-scale CNNs in the same chip area
- High computational throughput compared to today’s state of the art
BACKGROUND:
Superconducting digital computing is a promising technology for scaling high performance computing systems in applications including scientific research, defense, medical imaging, and more. However, new approaches in computer architecture and computational models are needed to make efficient use of superconducting computing technology. Additionally, superconducting chips have strict area constraints, meaning that current computing architectures may not be able to achieve the necessary performance levels with the limited amount of space available.
TECHNOLOGY OVERVIEW:
Berkeley lab researchers have developed a new way to implement a hardware accelerator for convolutional neural networks (CNNs) efficiently using superconducting chips. By using a combination of race logic and pulse streams as the basis of the computational model, this new data representation scheme can perform the computations needed for CNNs much more efficiently than with existing methods. They also developed new ways to store information in memory that allows more efficient area consumption on superconducting chips, something that is a grand challenge for Rapid Single Flux Quantum (RSFQ) superconducting digital computing.
The accelerator improves peak performance per area by 3x to 23x compared to a state of the art binary SFQ accelerator. It improves peak performance per Watt by 639x compared to a CMOS state-of-the-art CNN hardware accelerator.
Overall, this invention has the potential to make computers that can perform pattern recognition and other tasks much more quickly and efficiently, which could have important applications in fields like medicine, transportation, and security. This invention’s area efficiency makes realistic-size CNNs given today’s stringent area constraints of superconducting chips.
The invention could also be applied to quantum computing due to similarly low operating temperatures.
DEVELOPMENT STAGE:
Proof of concept
PRINCIPAL INVESTIGATORS:
Patricia Gonzalez-Guerrero, Georgios Michelogiannakis
IP Status:
Patent pending
OPPORTUNITIES:
Available for licensing or collaborative research