energy efficiency Archives - EFFECT Photonics

Insights

May 29, 2024

Data Centers in the Age of AI

Article first published 14th June 2023, updated 29th May 2024. Artificial intelligence (AI) is changing…

Article first published 14^th June 2023, updated 29^th May 2024.

Artificial intelligence (AI) is changing the technology landscape in various industries, and data centers are no exception. AI algorithms are computationally heavy and will increase data centers’ power consumption and cooling requirements, deeply affecting data center infrastructure.

The Constraints of the Power Grid

Data centers famously consume a significant amount of energy, and power-hungry AI algorithms will lead to a further increase in data center power consumption. The world’s major data center providers are already gearing up for this increase. For example, a recent Reuters report explains how Meta computing clusters needed 24 to 32 times the networking capacity. This increase required redesigning the clusters and data centers to include new liquid cooling systems.

Despite the best efforts of the world’s tech giants to rethink their architectures, it’s clear that data centers and their new AI workloads are hitting electrical power grid li m itations. The capacity of the power grid is now increasingly regarded as the main chokepoint that prevents AI clusters from being more widely implemented in data centers.

Since changes in the power grid distribution would take decades to materialize, data center providers know they cannot continue to centralize their data center architectures. To adapt to the power grid constraints, providers are thinking about how to transfer data between decentralized data center locations instead.

For example, data centers can relocate to areas with available spare power, preferably from nearby renewable energy sources. Efficiency can increase further by sending data to branches with spare capacity. The Dutch government has already proposed this kind of decentralization as part of its spatial strategy for data centers.

Figure 1: High-speed fiber-optic connections allow data processing and storage to be moved to locations where excess (green) energy is available. Data can be moved elsewhere if power is needed for other purposes, such as charging electric vehicles.

Interconnecting Data Centers over Long Distances

Longer data center interconnects enable a more decentralized system of data centers with branches in different geographical areas connected through high-speed optical fiber links to cope with the strain of data center clusters on power grids.

These trends push the data center industry to look for interoperable solutions for longer interconnects over 80 to 120 km distances. The advances in electronic and photonic integration allowed coherent technology for metro DCIs to be miniaturized into QSFP-DD and OSFP form factors. This progress allowed the Optical Internetworking Forum (OIF) to create the 400ZR and ZR+ standards for 400G DWDM pluggable modules. With modules that are small enough to pack a router faceplate densely, the datacom sector could profit from a 400ZR solution for high-capacity data center interconnects of up to 80km.

After the success of 400ZR standardization, the data center industry and the OIF are starting to promote an 800ZR standard to enable the next generation of interconnects. In OFC 2024, we started seeing some demos from several vendors and the OIF on this new standards initiative.

Optical Interconnects Inside Data Centers

The increasing complexity of AI processing will impact not just the interconnections between data centers but also the architectures inside the data center. AI nodes inside data center racks are normally connected via electrical or RF signals, while the racks are connected via optical fiber interconnects. However, as AI systems do more parallelized processing, data center racks run into electrical memory and power consumption constraints. These electrical interconnects between AI nodes are increasingly becoming a bottleneck in the ability of data center architectures to scale and handle the demands of AI models sustainably.

AI data centre — Figure 2: How architectures might change inside the AI data center. AI nodes on a single rack are connected via electrical or RF connections to an electro-optic switch that connects to other racks with optical fiber. In the future (on the right side of the figure), these AI nodes might be interconnected via optics. Figure inspired by Rob Stone’s (Meta) presentation.

Optics will play an increasingly larger role in this area. As explained by Andrew Alduino and Rob Stone of Meta in a talk at the 2022 OCP Global Summit, interconnecting AI nodes via optics will be vital to decreasing the power per bit transmitted inside data center racks. This means changing the traditional architecture inside these racks. Instead of using an electro-optical switch that converts the electrical interconnects between AI nodes to optical interconnects with another, the switching inside the data center rack would be entirely optical.

Avoiding the power losses of electrical connections and electrical-optical conversions will improve the cost and power per bit of connections inside the data center. As more data center capacity is needed, these new optical interconnections might also need co-packaged coherent optics to scale effectively. This is the argument made recently by our very own Joost Verberk, EFFECT Photonics’ VP of Product Management, at the 2024 OCP Regional Summit in Lisbon.

Takeaways

As AI continues reshaping the technological infrastructure, data centers undergo significant transformations to meet the new demands. The shift towards AI-intensive operations has exacerbated the existing strain on power grids, pushing data center providers towards more decentralized solutions. This includes relocating to areas with spare power and transferring data with optical interconnects across geographically dispersed locations.

The adoption of advanced optical interconnections is happening also inside data centers, as data center racks might transition to all-optical switching to connect their AI nodes. These evolving strategies not only address the immediate challenges of AI workloads but also set the stage for more sustainable and scalable data center operations in the future.

PDF Download

Tags: Artificial intelligence (AI), Cooling requirements, Data center architecture, data centers, energy efficiency, Intelligent automation, Networking capacity, Optimal temperature and airflow, power consumption, Power usage effectiveness (PUE) ratio

Insights

January 24, 2024

Coherent Optics for AI

Artificial intelligence (AI) will have a significant role in making optical networks more scalable, affordable,…

Artificial intelligence (AI) will have a significant role in making optical networks more scalable, affordable, and sustainable. It can gather information from devices across the optical network to identify patterns and make decisions independently without human input. By synergizing with other technologies, such as network function virtualization (NFV), AI can become a centralized management and orchestration network layer. Such a setup can fully automate network provisioning, diagnostics, and management, as shown in the diagram below.

carrier-network — Figure 1: Example of a carrier network with different applications (mobile, home, office) automated via SDN control and AI management.

However, artificial intelligence and machine learning algorithms are data-hungry. To work optimally, they need information from all network layers and ever-faster data centers to process it quickly. Pluggable optical transceivers thus need to become smarter, relaying more information back to the AI central unit, and faster, enabling increased AI processing.

The Need for Faster Transceivers

Optical transceivers are crucial in developing better AI systems by facilitating the rapid, reliable data transmission these systems need to do their jobs. High-speed, high-bandwidth connections are essential to interconnect data centers and supercomputers that host AI systems and allow them to analyze a massive volume of data.

In addition, optical transceivers are essential for facilitating the development of artificial intelligence-based edge computing, which entails relocating compute resources to the network’s periphery. This is essential for facilitating the quick processing of data from Internet-of-Things (IoT) devices like sensors and cameras, which helps minimize latency and increase reaction times.

400 Gbps links are becoming the standard across data center interconnects, but providers are already considering the next steps. LightCounting forecasts significant growth in the shipments of dense-wavelength division multiplexing (DWDM) ports with data rates of 600G, 800G, and beyond in the next five years. We discuss these solutions in greater detail in our article about the roadmap to 800G and beyond.

Graph-pluggables-shipments-dwdm — Figure : Shipments of high-speed DWDM ports by data rate (historical data and forecast). Source: LightCounting Optical Communications Market Forecast, April 2022.

The Need for Telemetry Data

Mobile networks now and in the future will consist of a massive number of devices, software applications, and technologies. Self-managed, zero-touch automated networks will be required to handle all these new devices and use cases. Realizing this full network automation requires two vital components.

Artificial intelligence and machine learning algorithms for comprehensive network automation: For instance, AI in network management can drastically cut the energy usage of future telecom networks.
Sensor and control data flow across all network model layers, including the physical layer: As networks grow in size and complexity, the management and orchestration (MANO) software needs more degrees of freedom and dials to turn.

These goals require smart optical equipment and components that provide comprehensive telemetry data about their status and the fiber they are connected to. The AI-controlled centralized management and orchestration layer can then use this data for remote management and diagnostics. We discuss this topic further in our previous article on remote provisioning, diagnostics, and management.

For example, a smart optical transceiver that fits this centralized AI-management model should relay data to the AI controller about fiber conditions. Such monitoring is not just limited to finding major faults or cuts in the fiber but also smaller degradations or delays in the fiber that stem from age, increased stress in the link due to increased traffic, and nonlinear optical effects. A transceiver that could relay all this data allows the AI controller to make better decisions about how to route traffic through the network.

A Smart Transceiver to Rule All Network Links

After relaying data to the AI management system, a smart pluggable transceiver must also switch parameters to adapt to different use cases and instructions given by the controller.

Let’s look at an example of forward error correction (FEC). FEC makes the coherent link much more tolerant to noise than a direct detect system and enables much longer reach and higher capacity. In other words, FEC algorithms allow the DSP to enhance the link performance without changing the hardware. This enhancement is analogous to imaging cameras: image processing algorithms allow the lenses inside your phone camera to produce a higher-quality image.

network link with FEC — Figure 3: Simplified diagram of a network link with forward error correction (FEC). The FEC encoder adds a series of redundant bits (called overhead) to the transmitted data stream. The receiver can use this overhead to check for errors without asking the transmitter to resend the data.

A smart transceiver and DSP could switch among different FEC algorithms to adapt to network performance and use cases. Let’s look at the case of upgrading a long metro link of 650km running at 100 Gbps with open FEC. The operator needs to increase that link capacity to 400 Gbps, but open FEC could struggle to provide the necessary link performance. However, if the transceiver can be remotely reconfigured to use a proprietary FEC standard, the transceiver will be able to handle this upgraded link.

Reconfigurable transceivers can also be beneficial to auto-configure links to deal with specific network conditions, especially in brownfield links. Let’s return to the fiber monitoring subject we discussed in the previous section. A transceiver can change its modulation scheme or lower the power of its semiconductor optical amplifier (SOA) if telemetry data indicates a good quality fiber. Conversely, if the fiber quality is poor, the transceiver can transmit with a more limited modulation scheme or higher power to reduce bit errors. If the smart pluggable detects that the fiber length is relatively short, the laser transmitter power or the DSP power consumption could be scaled down to save energy.

Takeaways

Optical networks will need artificial intelligence and machine learning to scale more efficiently and affordably to handle the increased traffic and connected devices. Conversely, AI systems will also need faster pluggables than before to acquire data and make decisions more quickly. Pluggables that fit this new AI era must be fast, smart, and adapt to multiple use cases and conditions. They will need to scale up to speeds beyond 400G and relay monitoring data back to the AI management layer in the central office. The AI management layer can then program transceiver interfaces from this telemetry data to change parameters and optimize the network.

Article PDF

Tags: 400 Gbps links, artificial intelligence, coherent optics, data centers, Dense-wavelength division multiplexing (DWDM), diagnostics, Edge computing, EFFECT Photonics, energy efficiency, forward error correction (FEC), Internet of Things (IoT), Machine learning algorithms, Network automation, Network function virtualization (NFV), network optimization, optical networks, optical transceivers, Reconfigurable transceivers, remote management, SDN control, telemetry data

Insights

May 31, 2023

Building a Sustainable Future with Fully Integrated PICs

Article first published 27 September 2021, updated 31st May 2023. The demand for data and…

Article first published 27 September 2021, updated 31st May 2023.

The demand for data and other digital services is rising exponentially. From 2010 to 2020, the number of Internet users worldwide doubled, and global internet traffic increased 12-fold. By 2022, internet traffic had doubled yet again. While 5G standards are more energy-efficient per bit than 4G, the total power consumption will be much higher than 4G. Huawei expects that the maximum power consumption of one of their 5G base stations will be 68% higher than their 4G stations. These issues do not just affect the environment but also the bottom lines of communications companies.

Keeping up with the increasing data demand of future networks sustainably will require operators to deploy more optical technologies, such as photonic integrated circuits (PICs), in their access and fronthaul networks.

Integration Impacts Energy Efficiency and Optical Losses

Lately, we have seen many efforts to increase further the integration on a component level across the electronics industry. For example, moving towards greater integration of components in a single chip has yielded significant efficiency benefits in electronics processors. Apple’s recent M1 and M2 processors integrate all electronic functions in a single system-on-chip (SoC) and consume significantly less power than the processors with discrete components used in their previous generations of computers.

𝗠𝗮𝗰 𝗠𝗶𝗻𝗶 𝗠𝗼𝗱𝗲𝗹	𝗣𝗼𝘄𝗲𝗿 𝗖𝗼𝗻𝘀𝘂𝗺𝗽𝘁𝗶𝗼𝗻
𝗠𝗮𝗰 𝗠𝗶𝗻𝗶 𝗠𝗼𝗱𝗲𝗹	𝗜𝗱𝗹𝗲	𝗠𝗮𝘅
2023, M2	7	5
2020, M1	7	39
2018, Core i7	20	122
2014, Core i5	6	85
2010, Core 2 Duo	10	85
2006, Core Solo or Duo	23	110
2005, PowerPC G4	32	85
Table 1: Comparing the power consumption of a Mac Mini with an M1 and M2 SoC chips to previous generations of Mac Minis. [Source: Apple’s website]

Photonics is also achieving greater efficiency gains by following a similar approach to integration. The more active and passive optical components (lasers, modulators, detectors, etc.) manufacturers can integrate on a single chip, the more energy they can save since they avoid coupling losses between discrete components and allow for interactive optimization.

Let’s start by discussing three different levels of device integration for an optical device like a transceiver:

Discrete build – The transceiver components are manufactured through separate processes. The components are then assembled into a single package using different types of interconnections.
Partial integration – Some components are manufactured and integrated on the same chip, but others are manufactured or sourced separately. For example, the transceiver laser can be manufactured separately on a different material and then interconnected to a chip with the other transceiver components.
Full integration – All the components are manufactured on a single chip from a single material simultaneously.

While discrete builds and partial integration have advantages in managing the production yield of the individual components, full integration leads to fewer optical losses and more efficient packaging and testing processes, making them a much better fit in terms of sustainability.

The interconnects required to couple discrete components result in electrical and optical losses that must be compensated with higher transmitter power and more energy consumption. The more interconnects between different components, the higher the losses become. Discrete builds will have the most interconnect points and highest losses. Partial integration reduces the number of interconnect points and losses compared to discrete builds. If these components are made from different optical materials, the interconnections will suffer additional losses.

On the other hand, full integration uses a single chip of the same base material. It does not require lossy interconnections between chips, minimizing optical losses and significantly reducing the energy consumption and footprint of the transceiver device.

Figure 1: Example of integrating a laser to other photonic components through different integration approaches.

More Integration Saves Scarce Resources

When it comes to energy consumption and sustainability, we shouldn’t just think about the energy the PIC consumes but also the energy and carbon footprint of fabricating the chip and assembling the transceiver. To give an example from the electronics sector, a Harvard and Facebook study estimated that for Apple, manufacturing accounts for 74% of their carbon emissions, with integrated circuit manufacturing comprising roughly 33% of Apple’s carbon output. That’s higher than the emissions from product use.

Figure 2: Apple’s carbon-emission breakdown. Manufacturing accounts for 74% of total emissions, and hardware use accounts for 19%. Carbon output from manufacturing integrated circuits (i.e., SoCs, DRAM, and NAND flash memory) is higher than that of hardware use. [Source: ]

Early Testing Avoids Wastage

Testing is another aspect of the manufacturing process that impacts sustainability. The earlier faults can be found in the testing process, the greater the impact on the use of materials and the energy used to process defective chips. Ideally, testing should happen not only on the final, packaged transceiver but in the earlier stages of PIC fabrication, such as measuring after wafer processing or cutting the wafer into smaller dies.

Figure 3: Package level (top row) vs. die level (bottom row) testing. Manufacturers can find faults earlier by testing at the die level, which avoids wasting packaging materials.

Discrete and partial integration approaches do more of their optical testing on the finalized package, after connecting all the different components together. Should just one of the components not pass the testing process, the complete packaged transceiver would need to be discarded, potentially leading to a massive waste of materials as nothing can be ”fixed” or reused at this stage of the manufacturing process.

Full integration enables earlier optical testing on the semiconductor wafer and dies. By testing the dies and wafers directly before packaging, manufacturers need only discard the bad dies rather than the whole package, which saves valuable energy and materials.

Full Integration Drives Sustainability

While communication networks have become more energy-efficient, further technological improvements must continue decreasing the cost of energy per bit and keeping up with the exponential increase in Internet traffic. At the same time, a greater focus is being placed on the importance of sustainability and responsible manufacturing. All the photonic integration approaches we have touched on will play a role in reducing the energy consumption of future networks. However, out of all of them, only full integration is in a position to make a significant contribution to the goals of sustainability and environmentally friendly manufacturing. A fully integrated system-on-chip minimizes optical losses, transceiver energy consumption, power usage, and materials wastage while at the same time ensuring increased energy efficiency of the manufacturing, packaging, and testing process.

Download Article – PDF

Tags: ChipIntegration, Data demand, DataDemand, EFFECT Photonics, Energy consumption reduction, energy efficiency, EnergySavings, Environmental impact, Fully Integrated PICs, Green Future, GreenFuture, Integrated Photonics, Integration benefits, Manufacturing sustainability, Optical technologies, OpticalComponents, photonic integration, PIC, PICs, ResponsibleManufacturing, sustainability telecommunication, Sustainable, Sustainable future, SustainableNetworks, Transceiver optimization

Insights

December 14, 2022

Fit for Platform DSPs

Over the last two decades, power ratings for pluggable modules have increased as we moved…

Over the last two decades, power ratings for pluggable modules have increased as we moved from direct detection to more power-hungry coherent transmission: from 2W for SFP modules to 3.5 W for QSFP modules and now to 14W for QSSFP-DD and 21.1W for OSFP form factors. Rockley Photonics researchers estimate that a future electronic switch filled with 800G modules would draw around 1 kW of power just for the optical modules.

Around 50% of a coherent transceiver’s power consumption goes into the digital signal processing (DSP) chip that also performs the functions of clock data recovery (CDR), optical-electrical gear-boxing, and lane switching. Scaling to higher bandwidths leads to even more losses and energy consumption from the DSP chip and its radiofrequency (RF) interconnects with the optical engine.

Figure 1: Normalized power breakdown of the internal blocks of a 400ZR pluggable module. Although this breakdown may vary a bit among different vendors, the authors find it to be generally true for various DSP-based modules, both IM-DD and coherent. Source: R. Nagarajan et al., IEEE JLT (2021)

One way to reduce transceiver power consumption requires designing DSPs that take advantage of the material platform of their optical engine. In this article, we will elaborate on what that means for the Indium Phosphide platform.

A Jack of All Trades but a Master of None

Transceiver developers often source their DSP, laser, and optical engine from different suppliers, so all these chips are designed separately from each other. This setup reduces the time to market and simplifies the research and design processes but comes with trade-offs in performance and power consumption.

In such cases, the DSP is like a Swiss army knife: a jack of all trades designed for different kinds of optical engines but a master of none. For example, current DSPs are designed to be agnostic to the material platform of the photonic integrated circuit (PIC) they are connected to, which can be Indium Phosphide (InP) or Silicon. Thus, they do not exploit the intrinsic advantages of these material platforms. Co-designing the DSP chip alongside the PIC can lead to a much better fit between these components.

Co-Designing with Indium Phosphide PICs for Power Efficiency

To illustrate the impact of co-designing PIC and DSP, let’s look at an example. A PIC and a standard platform-agnostic DSP typically operate with signals of differing intensities, so they need some RF analog electronic components to “talk” to each other. This signal power conversion overhead constitutes roughly 2-3 Watts or about 10-15% of transceiver power consumption.

Figure 2: Simplified diagram of the building blocks of a coherent transceiver in which the DSP drives the optical engine directly from its digital-to-analog converter (DAC). This setup can minimize the power conversion overhead of optical transceivers.

However, the modulator of an InP PIC can run at a lower voltage than a silicon modulator. If this InP PIC and the DSP are designed and optimized together instead of using a standard DSP, the PIC could be designed to run at a voltage compatible with the DSP’s signal output. This way, the optimized DSP could drive the PIC directly without needing the RF analog driver, doing away with most of the power conversion overhead we discussed previously.

Figure 3: Comparison between the drive power and power consumption of three different PIC + DSP pairings: standard DSP with a silicon PIC, standard DSP with an InP PIC, and an optimized DSP with an InP.

Additionally, the optimized DSP could also be programmed to do some additional signal conditioning that minimizes the nonlinear optical effects of the InP material, which can reduce noise and improve performance.

Taking Advantage of Active Components in the InP Platform

Russell Fuerst, EFFECT Photonics’ Vice-President of Digital Signal Processing, gave us an interesting insight about designing for the InP platform in a previous interview:

When we started doing coherent DSP designs for optical communication over a decade ago, we pulled many solutions from the RF wireless and satellite communications space into our initial designs. Still, we couldn’t bring all those solutions to the optical markets.

However, when you get more of the InP active components involved, some of those solutions can finally be brought over and utilized. They were not used before in our designs for silicon photonics because silicon is not an active medium and lacked the performance to exploit these advanced techniques.

For example, the fact that the DSP could control laser and modulator components on the InP can lead to some interesting manipulations of light signals. A DSP that can control these components directly could generate proprietary waveforms or use non-standard constellation and modulation schemes that can boost the performance of a coherent transceiver and increase the capacity of the link.

Takeaways

The biggest problem for DSP designers is still improving performance while reducing power use. This problem can be solved by finding ways to integrate the DSP more deeply with the InP platform, such as letting the DSP control the laser and modulator directly to develop new waveform shaping and modulation schemes. Because the InP platforms have active components, DSP designers can also import more solutions from the RF wireless space.

Tags: analog electronics, building blocks, coherent, dispersion compensation, DSP, energy efficiency, Intra DCI, Photonics, PON, power consumption, reach, simplified

Core Industries

Emerging Industries

Our Technology

Pluggable Modules – Coherent

Sub-Assemblies

ASICs

View All

Insights

Press Releases

In the News

Events

Careers

Knowledge Center

Company Info

Contact Us

Core Industries

Emerging Industries

Our Technology

Pluggable Modules – Coherent

Sub-Assemblies

ASICs

View All

Insights

Press Releases

In the News

Events

Careers

Knowledge Center

Company Info

Contact Us