Heterogenous Convolution ASIC for AI Acceleration (HelCat)
Abstract
Convolution Neural Networks are an essential component in modern electronic warfare. In an information dominance age, an increasingnumber of mission tasks are augmented via automation including machine meaning such as in autonomous vehicles, object recognition,target tracking, and many more. Indeed, large data-sets and long training sessions are required to perform machine learning, both in a centralized data center and near future edge-computing. However, data-set are increasing in size, making those algorithms to be more and more demanding, in terms of speed, power consumption, and latency. Nonetheless, a large supply-demand gap exists between current CMOS-based machine learning accelerators and demand, especially in terms of energy consumption and latency, for applications where a few timing is critical. Heterogenous Convolution ASIC for AI Acceleration (HelCat) will develop two advanced versions of a PhotoFourier Computational Unit (PFCU) based on the full-optical Joint Transform Correlation (JTC) approach that is capable to speed up Fourier-based convolution processing by several orders of magnitude. Generation 1 will explore the integration of a series of JTC processors into a unique system, capable of reducing the computational cost of any convolution operation from O(N^4) to just O(N^2),leveraging the full integration of lasers, modulators, on-chip lens, and photodetectors in a single packaged silicon photonic chip.This processor will be able to handle vast data at gigabits per second datarates and nanosecond short latency. This first system will be capable of performing 10#s of TOPS/W, fully capable of addressing centralized high-demanding Convolution Neural Networks, for both training and inferences. The implementation and characterization of this first JTC-based architecture will pave the way for a second-generation (Gen-2) PFCU processor leveraging CMOS+Photonic monolithic integration into the same chip, using trusted foundry processes. The integration of several CMOS functions right next to the photonic components will increase performance and the energy efficiency by exploiting the lower parasitic capacitance, and so higher bandwidth leveraged by monolithic integration of heternogenouschip capability. This enables increasing the number of channels per single PFCU, and the number of PFCUs per system, while simultaneously increasing the bandwidth at the same time, having all the electrical components such as TIA, ADC, and DAC, directly in the same chip as all the optical components. With these features, the system will be able to integrate 16 PFCUs on the same board, reaching a top performance of 100#s TOPS/W. The team working on this project is made up of professionals from the business and academic worlds. They have a strong working knowledge of the device, circuit, and system-level R&D, including prototyping in optoelectronics, integrated photonics, RF circuits, computer engineering and architecture, neural networks, and data analytics.With Gen-2, we will push the boundaries of Neural Networks, as we will be able to perform training and inference at high throughput and low latency. Applications with demanding computation and low latency requirements, such as navigation, ranging, and edge-processing at high-SWAP performance, object detection and tracking, image processing, while providing a high-level of zero-trust assurance. Key applications for DOD include synthetic aperture radar, target recognition, classification of electromagnetic signatures, hyper-spectral filtering, to name a few. If successful, the project will deliver a whole new architecture to perform Convolution Neural Networks, with orders of magnitude improvements in computational cost and latency than the current electronics processors for DOD potential for missions withdual-use in civil applications.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 24, 2023
- Source ID
- N000142312687
Entities
People
- Volker Sorger
Organizations
- Office of Naval Research
- United States Navy
- University of Florida