Taking the Fast Bridge between Neural Networks

How do neural networks scale up to 20,000 processors and beyond? How does machine learning scale massively if connections are a potential weak link in the race for acceleration?

Editor’s Note: Neural networks have been around since the 1950s. The advent of fast, massively parallel processors like the Graphics Processing Unit (GPU) have made neural network applications like object recognition feasible. Neural networks are one means used to create Artificial Intelligence (AI). The latest iPhones now have an AI chip, primarily to offload face recognition tasks.[i] Voice translation tasks would also benefit from an AI chip. Google provides voice translation as long as there is access to a cloud. The ability to translate directly from a phone without requiring Internet access to Google engines would be advantageous, and it’s possible that the iPhone is headed in that direction.

AI often needs real-time operation. China’s answer to policing during the New Year’s travel crush involves providing police officers with smart glasses for rapid facial recognition. If using traditional video cameras, the suspect has left the scene by the time policemen arrive. China’s police force is equipped with smart facial-recognition glasses connected to a pocket-sized data module for identifying up to 10,000 individuals. The module’s database dramatically reduces latency versus a cloud access system. Recognizing a face in as little as 100 ms, police can immediately act upon the information.[ii] At the other end of the spectrum, enormous banks of GPUs in server farms can require low latency. Scaling up to 20,000 processors and beyond needs relatively high bandwidth between GPUs to communicate quickly. Connectivity can be the slow link in a system with the fastest GPUs. A company located in Quebec, Canada, called Reflex Photonics, has a connectivity solution that fits the bill, enabling a machine learning or AI infrastructure to scale up with more processors. Reflex Photonics provides tiny optical transceiver chips that can move a tremendous amount of data, which reduces the latency between GPUs so that they can appear as though working seamlessly, in parallel.

Figure 1: The Mercedes Freightliner Inspiration truck is the first road-approved truck for autonomous operation. (Source: Daimler)

One area that Reflex Photonics serves includes VPX systems. According to the VMEbus International Trade Association (VITA.com), “VPX is a broadly defined technology utilizing the latest in a variety of switch fabric technologies in 3U and 6U format blades.” VITA technologies are well known in military and aerospace. Interconnects for serial switch fabric in venerated technologies like XMC, VPX, and VXS, as well as new standards that include Gigabit Ethernet, PCI Express, and Serial RapidIO are in the VITA ecosystem. Form factors in this technology include credit card-sized processing platforms up to the 6U Eurocard.

Technology for optical interconnects is vital in the industry, since processor speeds are outpacing copper wires for bandwidth and latency among devices, including VPX. During our recent interview, Gerald Persaud, VP Business Development at Reflex Photonics, told me this is a solvable problem, even in the harsh environments of military and outer space. Edited excerpts follow:

Lynnette Reese (LR): What is Reflex Photonics currently developing?

Gerald Persaud, Reflex Photonics

Gerald Persaud (GP), Reflex Photonics: Today, we are focused on aerospace and defense, and industrial markets. Our expertise is delivering chip sized rugged high bandwidth optical transceivers that work in the harshest environments, such as space. For example, we were recently selected for a major satellite program because our parts could meet the required 20 years lifetime in space. Many optical transceiver suppliers claim high bandwidth operation at 25Gbps per channel but only for an operating temperature of 0 to 70oC. All of Reflex Photonics’ rugged transceivers operate error-free over a temperature range of -50 to 100oC while also meeting severe shock, vibration, damp heat, and thermal cycling requirements.

LR: Reflex Photonics’ expertise is in ruggedized optical communications. How did your process for dealing with the challenges of harsh environments evolve?
GP: In 2002 when we started the company our goal was to create a chip-size optical module that could be solder reflowed to support low-cost board assembly. This was much harder than we had imagined due to differences in material properties such as thermal expansion, thermal conductivity, and curing processes. Over the years we were able to incrementally improve our manufacturing processes from a commercial offering to a full space-qualified part. An excellent understanding of materials and processing is critical to the successful production of high-bandwidth rugged optical modules.

Figure 2: Optical interconnect for high-speed, high- bandwidth 10GigE and 40GigE cameras used in machine vision.

LR: What is on your roadmap?
GP: We plan to release higher channel speeds up to 56Gbps, more I/O density such as 24 transmitters or receivers in a chip size optical module. As well, we will continue to harden our parts to meet even wider temperature extremes of -65 to 125oC. Another product we recently released is active blind-mate optical connectors called LightCONEX®. We have gained a great deal of interest in this solution from the VPX community, as it frees up a lot of board space and simplifies field upgrades.

LR: Can you give an example where Reflex Photonics has a play in VPX for machine learning?
GP: One example of this is in unmanned vehicles where machine learning is critical for autonomous operation. Many sensors are interconnected to machine learning VPX compute farms via an optical switch. Optical interconnect, with its long reach, high bandwidth and light weight, is the only viable solution for advanced Autonomous Vehicles (AVs). From the start, Reflex set out to make the smallest rugged optical modules capable of supplying enormous bandwidth (BW) and optical channels. Today, Reflex Photonics’ rugged technologies are field proven and well positioned to take advantage of the trend for smarter, smaller, and robust systems.

LR: How are you dealing with power challenges in a Small Form Factor (SFF)?
GP: Power is indeed a challenge for mobile vehicles, which have a limited amount of power to supply onboard electronics. Today a 150Gbps chip consumes about 1.3W. However, as bandwidth demand grows from 150Gbps to 2400Gbps over the next five to 10 years we cannot scale power linearly or the same chip will consume 21W. And there are multiple chips per board!  We will need to introduce techniques to improve optical coupling efficiency and lower laser bias currents. As well, laser drivers and amplifier will need to operate at lower voltages. Closer integration of the drive electronics with optical transceivers could save a lot of power as the need for Clock and Data Recovery (CDR), equalizer, or pre-emphasis could be eliminated.

LR: What are your competitors doing? How is Reflex Photonics any different?
GP: Everyone including Reflex is racing to increase BW and interconnect density. However, in the aerospace and defense sector, suppliers must also meet the challenges of operating in a very harsh environment while keeping space, weight, and power [SWaP] to a minimum. Reflex is different in that we were the first to deliver a 150Gbps chip-size optical module that could operate from -50 to 100oC while consuming 1.2W. Most recently Reflex launched the first radiation-hardened parallel optical chip for space applications. These chips passed extreme environmental test conditions that our competitors were unable to meet. This is excellent news for the space industry, where size and weight are critical and smallsats are expected to do far more than their predecessors.

LR: I have always considered price to be a specification. How is your pricing affected by ultra-hardening for space?
GP: The price differential is not as significant as most would expect. In the old days when you said “space,” it meant 10 times the price. Those days are gone. There might be 30% increase in price for space grade over a military grade device. One grade down from military is the industrial device, which has similar operating temperatures but is not expected to have as long a life as Space and MIL grade parts.

Figure 3: LightCONEX® 50G and 150G is a rugged blind mate optical interconnect for VPX embedded computing systems. Used in military systems, optical interconnects are faster and more resistant to noise than copper-based connection systems. The potential for use in automotive reflects the need for high bandwidth communications at real-time speeds for modern automotive AI systems. (Image Source: Reflex Photonics)

LR: Can you detail some of the challenges for optics at extreme operating temperatures?
GP: Optical transceivers require exact alignment (less than five micrometers) of the laser or photodetector to the optical coupler. One challenge is maintaining this alignment over a wide temperature range. Reflex developed a patented approach using materials with low coefficient of thermal expansion and a simple coupling structure with no intermediate lens to maintain alignment over a wide temperature range of -57 to 125oC. Another challenge is having a cost-effective sealing method (for moisture resistance in the optical path) that will withstand many thermal cycles without compromising the mechanical integrity of the module. Of course, there are other challenges like radiation hardening, solder reflow temperature survival, low power, optical sensitivity, and signal integrity.

LR: What are the different grades of products that you have for harsh environments?
GP: Most of our sales are for MIL, Space and Industrial grade parts. We offer some commercial grades such as QSFP and CFP for Telecom/Datacom markets. Our industrial components are used in many applications such as commercial aircraft, semiconductor wafer inspection, and instrumentation and tests. Most recently, we have had a number of automotive applications for our industrial parts.

LR: Where would the automotive or transportation industry need rugged optical transceivers?
GP: The automotive industry is quite large and includes cars, city buses, transport trucks, and other vehicles. We expect as self-driving or assisted driving goes mainstream fiber-optics will interconnect all systems in the vehicle. Compact AI engines will connect many sensors to automate driving. The vehicles of tomorrow will provide great energy efficiencies, less pollution, and a comfortable and productive driving experience. NVidia is now offering small form factor AI engines that are already deployed in Unmanned Aerial Vehicles.

LR: Any optical transceiver is still going to need fiber to transport the signal in a system. Isn’t vibration a real problem for this kind of signaling in a vehicle?
GP: No. Our parts have been tested to MIL-STD-883, Method 2007.3 for vibration and Method 2002.4 for shock. Vibration is 20 to 2000Hz, 20g, 16 minutes per axis and shock is 500g, 0.5ms pulse, 5 repetitions, 6 directions. These tests were done while transmitting and receiving 150Gbps with no errors.

LR: That’s impressive. What distance and latency are we talking about?
GP: Distance in AVs are typically less than 100 m, and latency is less than 1 microsecond.

LR: Do you see Reflex Photonics involved in Autonomous Vehicles (AVs) someday?
GP: Yes, AVs will require fiber-optics for security, bandwidth, latency, and SWaP. As the leading provider of rugged high bandwidth optical transceivers, Reflex is well positioned to deliver the most reliable optical interconnect for AVs. For large AV industries like commercial automotive one big challenge will be reducing the price of optical transceivers while keeping all the ruggedization testing in place. This will happen over a number of years, and so we will invest accordingly to track market prices.

LRWhen do you think AVs will start to get traction?
GP: When the technology is considered safe adoption will happen. This will require years of education and trials. One area of concern is cybersecurity—nobody wants a hacker taking over their vehicle at 60 miles per hour. An effective strategy will be needed to isolate critical control functions from infotainment. This separation is done in commercial aircraft and similar standards will be imposed on AVs. Fiber-optics provide the first level of defense since they are immune to electromagnetic interference and therefore harder to disrupt. As well, learning machines will be smart enough to initiate automatic protection from dangerous threats. Protection techniques commonly used by military aircraft could be deployed.

LR: How do you think the Autonomous Vehicle is going to play out, in reality?
GP: The benefits of autonomous vehicles have long been known, but safety has always been a barrier. The recent advances in AI and low-cost sensors has generated great hope for convenient, safe and cost-effective people transport. Like everyone, I see a gradual shift to AV starting with assisted driving available now to special lanes for AV followed by AV completely dominating the roads. I see China embracing this technology to solve local pollution issues while seizing the opportunity to lead the automotive industry.

Lynnette Reese is Editor-in-Chief, Embedded Intel Solutions and Embedded Systems Engineering, and has been working in various roles as an electrical engineer for over two decades.




Novet, Jordan. “Apple Packed an AI Chip into the IPhone X.” CNBC, CNBC, 12 Sept. 2017, www.cnbc.com/2017/09/12/apple-unveils-a11-bionic-neural-engine-ai-chip-in-iphone-x.html.

[ii] Chin, Josh. “Chinese Police Add Facial-Recognition Glasses to Surveillance Arsenal.” The Wall Street Journal, Dow Jones & Company, 7 Feb. 2018, www.wsj.com/articles/chinese-police-go-robocop-with-facial-recognition-glasses-1518004353.

Share and Enjoy:
  • Digg
  • Sphinn
  • del.icio.us
  • Facebook
  • Mixx
  • Google