Fourier Optics Multiplication Accelerator for Convolutional Neural Networks and Cryptographic Applications
Abstract
With both MooreÕs law and Dennard scaling ending, desktop HPC performance is predicted with non-Von Neumann architectures. These distributed systems rely on a high-degree of interconnectivity suitable to be performed in the optical domain. The goal and deliverables of this project will model, build, and demonstrate two prototypes of photonic convolution coprocessors capable of performing probabilistic algorithms suitable for fast modular multiplication Montgomery multiplication (MM) with high throughput via parallelism in real-time. The innovative concepts of this optical coprocessor enable (modular)multiplications and convolutions via a) leveraging massive parallelism (106-107 simultaneous channels) using free-space optics, and spatial light modulators (SLM) in Gen1 and digital mirror display technologies (DMD) in Gen1*, b) realizing convolutions as simpler multiplications in the Fourier domain passively performed by a non-power consuming lens, c) favorable O(n) multiplication-scaling rather than O(n2) due to the distributed property of FT optics, and d) provide a testbed to build probability into reasoning- and sensing systems. The approach of this project consists of 3 phases, namely, modeling and sub-module testing (Y1), Gen1&1* tested (Y2&3), and a performance review of an out-year Gen2 system based on GHz-fast photonic integrated circuit (PIC) convolutions or metasurface-based SLMs (Y3). Both delivered prototypes (Gen1&1*) are controlled by a user (PC) with a FPGA as a programmable interface for data I/O, storage of data-feed, such as string of numbers opportunely arranged, images, crypto keys to the optical coprocessor. A PCIe board handles data I/O between the PC and the FPGA (at 64Gbps). An initial Gen-0.5 preliminary prototype of PI Sorger shows convolution functionality. Engineering challenges to be tackled include the phase field, which need to be aligned in the Fourier domain to improve filtering accuracy. To address this, we will analyze the entire system using appropriate commercial and home-built software. The outcomes of this project are to gain fundamental insights into analog probabilistic optical information processing, and to deliver two coprocessors able to perform MM and its related cryptography applications as well as machine learning perception such as in compressed sensing or super-resolution. This project explores efficiently executing large probabilistic modular multiplication and convolutions optically Ôon-the-flyÕ with nanosecond-short delays and high-degree of reconfigurability using high-speed interface with an FPGA to handle high data volumes. A ÔrawÕ performance assessment of Gen1&1* shows a near-HPC like computing capability of nanosecond-short run-time, 250+T MAC/s throughputs with 100 s-short reconfigurable timescales at 50W total system power and footprints comparable to todayÕs GPUs. NSA Relevance: Performing modular multiplications is a critical capability for number of security applications such as public key cryptography. However, security requires the use of large numbers that might not be practical to directly implement in optics. Here we will use multi-word, high radix MM to implement cryptographic operations such as RSA. There is a trade-off between the size of the optical components and the overall performance of the system and its interplay between the probabilistic algorithm used for implementing MM that will be studied. DOD Relevance: Increasing autonomy requires enhanced perception and reasoning capabilities. The rationale is that noise and stochastic processes enable higher robustness and less brittle decision systems. For an optical input signal (e.g. a reconnaissance camera) the proposed system could perform (pre)processing compressed sensing/super-resolution for tracking or security feature filtering for sensitive data.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 04, 2019
- Source ID
- W911NF1910468
Entities
People
- Volker Sorger
Organizations
- Army Contracting Command
- George Washington University
- National Security Agency