NEWPhotonics: Shaping the Future of Data Centers Through All-Optical Chips for Digital Signal Processing
As the demand for cloud computing power skyrocketed in recent years, so did the energy consumption of data centers, leading to huge additional energy expenses for cooling and a significant carbon footprint.
One particularly tricky bottleneck is that data is transmitted optically through fiber cables, but, currently, it requires processing being done digitally on silicon DSP chips – and the conversion from digital to optical signals (and vice versa) through a pluggable optical transceiver involves a lot of energy consumption and heat dissipation.
Founded by Yaniv Ben-Haim, Dr. Yossi Zilberfarb, and Prof. Yosef Ben-Ezra in 2020, the startup NEWPhotonics addresses this painful challenge by building all-optical silicon chips that handle the digital signal processing in an optical domain in the transceivers. Thereby, NEWPhotonics is cutting their energy consumption by more than 50% while improving their bandwidth, reliability, and latency and enabling additional capabilities not present in current transceivers, e.g., for AI, security, quantum computing, and blockchain applications.
NEWPhotonics raised two funding rounds from NextLeap Ventures, VAHOCA, AWZ Ventures, and TRIREC and received grants from the European Innovation Council (EIC) and the Israel Innovation Authority (IIA). In addition, three Data Centers have identified NEWPhotonics as a strategic investment for them, as they experience the necessity to cool off their power-hungry infrastructure constantly.
Learn more about the future of all-optical processing engine chips for digital signal processing from our interview with the founder and CEO, Yaniv Ben-Haim:
Why Did You Start NEWPhotonics?
Nine years ago, we ran a company that designed and manufactured optical and photonic solutions for the in-building cellular coverage market with a spin-off with vast experience in photonics chip design. The company was merged in 2016, and after a three-year journey of leading a successful company that developed autonomous car technologies, we got back to photonics. We founded NEWPhotonics in early 2020 with the goal in mind of helping data centers consume less energy, drastically reduce their power consumption and minimize their carbon emission.
Energy consumption is an acute problem for data centers: It’s slowing down their upscaling, requires lots of additional energy for cooling, and leads to substantial carbon emissions. It’s such an important measure that the size of data centers is typically characterized by the energy it consumes, not the square meters it occupies. Relying on our previous experience in the silicon and networking industry, we developed an understanding of why certain elements in a data center consume so much energy – and set out to find a solution.
How Does Optical Digital Signal Processing Work?
Data centers waste most of their energy on interfaces moving data in or out from the chip level to the entire networking layers. Fiber optics transfer data optically between components. Yet, it needs to be converted to electronic signals for processing within a chip – and each conversion costs up to 60% of the electrical-optical energy. The transceiver which is processing this conversion also turns parallel data – being transferred through multiple channels in electrical to fiber optics (and vice versa) – into muxed/de-muxed data, which a digital chip can process through a so-called SerDes (serializer/deserializer) or GearBox. And in addition, the digital chip that includes the SerDes as part of the digital processing chip (DSP) is responsible for filtering the signals, format conversion, clock data recovery (CDR), and the gearbox function.
Nowadays, SerDes are being implemented using silicon chips. The industry is trying to improve the DSP by going to shorter transistor sizes, i.e., 7 nm or 5 nm chips, to save power consumption and increase their capacities. Yet, this gives less than 5% performance improvements. Also, it may take about four to six years to double the bandwidth from currently 112 GBit/s to 224 GBit/s, which is likely the maximum that cooper-based electronic SerDes can achieve.
We found a way to implement the DSP functionalities optically using silicon photonics, i.e., before the optical signal gets converted into an electronic one. This allows us to not only implement more DSP functions than a silicon chip could but also to be 50% more energy-efficient, have 40% less latency, and support bandwidths in the range of 400 GBit/s to 3.2 Terabit/s – that’s eight times better the current state-of-the-art!
By replacing billions of silicon transistors with only a few optical elements, our photonics chips also offer better reliability – so there’s no need to replace them every three to five years. And we can manufacture our chips on a 45 nm node or even use mature node processes such as 90 nm or even 180 nm. I.e., we’re not competing with state-of-the-art chip manufacturing and can use mature silicon foundries where there’s less competition. While we replace silicon CMOS chips with silicon photonic chips, we could also work with other materials – but silicon photonics is the most advanced platform for mass production.
We developed several patents on how to implement DSP in silicon photonics, and we developed essential technology for monitoring and stabilizing photonic integrated circuits (PiC). As light moves through a waveguide, it’s susceptible to changes, e.g., in temperature or process variations. So to align the phases and functions, one has to stabilize the circuit. Typically, the industry uses heaters – little integrated cooper wires wrapped nearby around the waveguide – to alleviate temperature differences along the circuit and keep the light signal in the required phase, which is, nowadays, measured by beam splitters and photodetectors. Yet, heaters create a lot of overhead and consume additional energy, so it becomes nearly impossible to build large and complex photonic circuits needed for processing, quantum & AI. That’s why in our technology, we’re not using heaters but have developed our unique method for noninvasive monitoring and stabilizing the light’s phase across the circuit in a digital way. We also developed our own algorithms to simulate and then calibrate our photonic chip in the simulation phase before the tape-out and also post-tape-out for real-time calibration.
How Did You Evaluate Your Startup Idea?
We started by understanding the problem, where our previous industry experience helped us greatly. Then, we began to develop ideas on how to solve it, and naturally, we chose photonics, i.e., moving the DSP functions into the optical domain, as the chosen solution.
Next, after deciding on the right technology and architecture, we started doing research and simulations to develop the technology and patents – we had a lot of time to think about it during the Covid quarantines in 2020. Back then, we evaluated our simulations based on initial physics theories. Today, after receiving the first results from the lab of measuring discrete components and after releasing seven successful tape-outs, we can successfully compare our theoretical simulation results with an actual evaluation of the components. In order to develop our technology further, early support and seed funding from our investors was crucial.
Importantly, last October 2022, we were defined by the EU Innovation Council (EIC) as a disruptive “Blue Ocean Innovation,” which is essential for the future of data centers – ensuing a grant and an investment by the EIC. We believe that such recognition will greatly support us on the road toward commercialization.
What Advice Would You Give Fellow Deep Tech Founders?
Start by understanding the problem first. Then find a solution to solve it. Don’t fall in love with technology, as you have to solve problems to build a business.
Also, make sure that there is a market for the technology you are building. That’s why we decided to focus on pluggable transceivers and optical switches, as it solves a hair-on-fire problem in an existing market rather than some hypothetical future market, e.g., as is the case for photonic AI processors or co-packed optics.
Focus on solving a problem in an existing market with disruptive technology because then you don’t need to educate the market about your solution. Disrupt an existing market: It’s 20% about the right technology and 80% about the right way to bring it to the market. And make sure, right from the very first moment, that your technology can be mass-produced. Otherwise, you’ll face serious issues when attempting to scale it.
Israel’s 3 new technology trends for 2023 – The Israel Innovation Authority announced that it would launch three innovation consortiums, and NEWPhotonics is among the companies participating.
European Innovation Council: New wave of deep tech start-ups set to receive Accelerator grants and equity investments – More context on the EIC program, with NEWPhotonics being among the selected companies of the EIC Accelerator 15 batch.