Artificial Intelligence promises huge benefits to the world but lacks the proper infrastructure for implementation. AI infrastructure today is either too expensive, too slow, too hot or far too power hungry to be sustainable and commercially available to many potential customers. This is especially salient for edge solutions where accurate, split-second responses are required to mitigate rapidly evolving situations, like accidents, crime and violence, virtual frauds, cyber hacking and even viral plague spread. These require massive computing power which today comes with an unacceptably high price tag.
Photonic Computing, can solve these issues and take the world into a faster, cleaner, better 21st century computing. CogniFiber is leading these efforts with DeepLight™ Technology.
Using light within silicon chips, may seem appealing, but these photonic chips present serious barriers in their interface, power consumption, size, scaling, thermal stability, yield and noise, especially due to the fact that computations rely on coherent signals.
DeepLight technology is about harnessing the tremendous capabilities of optical communication platforms together with robust, stable and resilient fiber optics manufacturing and products, to generate phase-independent computing modules and systems that are:
These advantages make DeepLight a superior solution within the newly growing photonic computing industry, and a far more suitable solution for the growing needs of the global information economy.
Following the completion of a successful POC during our Seed stage, CogniFiber is now on a fast track to complete its first full system prototype (expected April 2022). Our first products, expected Q4 2023, implement a trainable photonic Auto-Encoder neural network system with expected inference performance of >400 million tasks per second (>100X acceleration) and a modest power consumption of less than 500 Watts (<20% of any competing technology). This represents a compute efficiency game-changer, enabling fast, affordable and sustainable AI inference applications at the Edge and Datacenter.
Our first line of products is most suitable for applications in Industrial IoT and Cybersecurity. Both verticals have a massive use of real-time, on-premise, auto-encoder based functions such as anomaly detection, transformation, de-noising and compression. Datacenter acceleration can also benefit from a high performance, low-power auto-encoder as recently was exemplified in NLP and Computer Vision. Stay tuned for the display of our first-of-its-kind prototype in CLEO international conference (Sun Jose, California, May 2022).
The physics behind DeepLight Technology is based on turning a flaw into a feature. In optical communication, transmission of light via multi-core fibers brings tremendous speeds and capacity. Traditional methods required maximal separation between cores to prevent crosstalk noise. Computing, however, requires inter-channel interaction in the same manner that “Source” and “Drain” channels interact in semiconductor transistors with the “Gate” as interaction control.
The principal of in-fiber computing is exactly the same: allowing controllable, programable interactions between input data channels such that the output light transmits the exact desired mathematical function. In this manner both linear operations (i.e. Multiply-Accumulate) and nonlinear operations (i.e. sigmoid, RelU etc.) can be implemented. The fact that the control of these interactions can be done in a configurable manner, results in an “FPGA-Like” photonic device that can implement many interchangeable functions by simply re-programing their control parameters. Our publication demonstrates that light routing between cores can be controlled by amplification patterns that are injected optically into our processors allowing for direct photonic compute from start to end without memory read/write operations.
The advantages of Fiber-Based Processing Units:
Our careful calculations and POC results demonstrate that DeepLight Technology will reach >100,000,000 TOPS (>100 exa-operations per second) with efficiency of >1,000 TOPs/Watt in just several years (2026) and bring sustainable supercomputing capabilities to the edge.
Company IP includes eleven patent applications, of which three are accepted and four pending.
Inspired by the great promise involved in combining optical communication platforms with the revolution of deep learning networks, Eyal set out to explore and map a brand-new field of in-fiber photonic computing. In early 2019, leveraging his broad technical experience in neuroscience (Weizmann Institute), HW design (Mellanox, Saifun), algorithms development (Oren, MS Tech) and photonic computing research (Technion, Bar Ilan University, Hebrew University), Eyal co-founded CogniFiber with world-class photonics expert Prof. Zeev Zalevsky, CogniFiber’s CTO. He recruited an elite group of engineers, implementing the theory from the whiteboards into the lab. During its first two years, CogniFiber developed and registered significant IP enabling a POC that dramatically supersedes the most powerful current AI acceleration technologies in terms of performance and power consumption. Eyal is truly energized and excited to lead CogniFiber’s journey into the next technological revolution.
Prof. Zalevsky is a Full-professor and Dean of the Faculty of Engineering at Bar Ilan University. His research expertise involves biomedical optics, super-resolution, and electro-optical devices. He has published more than 450 peer-reviewed papers, 10 books, 30 book chapters, and around 100 patents. For his work and great service to science, he received many national and international scientific prizes. He is also a co-founder and technological leader in many successful start-up companies that originated from technologies commercialized from his lab. Together with Dr. Cohen in 2012, Prof. Zavlevsky began the promising research endeavor which explored the implementation of in-fiber neural networks, enabling de-facto light speed inference computing.
Additionally to our founders, CogniFiber relies on an extraordinary team of experts across a wide variety of fields. Optics, AI and advanced electronics are just the tip of the iceberg. Developing and creating DeepLight™ Technology is a multidisciplinary effort, that wouldn’t be possible without the great and diverse team at CogniFiber.
We are looking for an electro-optics R&D engineer that will be a part of the team that invents, designs, assembles, tests and delivers future photonic computational systems, parts and devices. The job requires in-depth knowledge and experience in fiber optics and photonics, specific to CogniFiber unique technology.
Contact us for more information and send relevant CV.
