a

Advancing the Next-Generation of AI: The Evolution of Optical I/O with the SuperNova™ Light Source

by | Apr 30, 2024

Recent rapid innovations in artificial intelligence (AI) have highlighted the need for speed and efficiency in data. Ayar Labs is at the forefront of this transformation, championing a paradigm shift with optical I/O technology. Our latest evolution, the second-generation SuperNova™ remote light source, represents a significant building block for high-bandwidth connectivity that promises to reshape the marketplace.

Let’s explore how laser packaging techniques and strategies impact the performance and scalability of optical I/O solutions and lay the foundation for a deeper understanding of SuperNova’s role in advancing AI infrastructure.

The Role of Laser Packaging in Silicon Photonics

Silicon photonics has revolutionized data transfer within and between logic and memory chips by using light instead of electrical signals. This has become increasingly important to deployments targeting generative AI. The flow of data across these complex systems is increasingly constrained by traditional copper-based solutions that limit the scale and data rates at which compute devices and memory capacity can be clustered together while driving up power consumption, system complexity, and costs.

Pluggable optics have been the most common method of converting electrical signals into optical signals (and vice versa) to address these shortcomings. However, co-packaged optics (CPO) have achieved maturity and are increasingly being deployed in interconnect fabrics, CPO moves the I/O module away from the faceplate by integrating the module components into a single package alongside the compute or switch chip.

Figure 1. Size comparison between a pluggable optic and Ayar Labs’ TeraPHY™ optical I/O chiplet.

This proximity reduces the dependence on copper connections, which often have shortcomings due to signal loss and limited transmission distances at higher data rates. Co-packaging brings the functions of the pluggable transceiver right next to the application-specific integrated circuit (ASIC).

In a blog post last year titled “In-Package Optical I/O versus Co-Packed Optics – Let’s Get Techincal,” our CTO, Vladimir Stojanovic, outlined the differences and applications between in-package optical I/O and CPO. This post identifies optical I/O as a more integrated, energy-efficient solution designed for distributed compute systems, such as AI and data center designs, that require high bandwidth density, low energy, and low latency. CPO, on the other hand, is described as a slightly improved alternative to pluggable optics, mainly intended for large Ethernet network switches, and does not provide the same level of integration or efficiency benefits as optical I/O.

CPO solutions can use either integrated or remote laser sources. Let’s take a look at the two options and explore their benefits and drawbacks.

Integrated Light Sources

Integrated light sources can either be individually manufactured and co-packaged with the photonic integrated circuits (PICs) or monolithically fabricated with the PICs. In either case, the lasers that produce the light are typically in contact with the PIC and in close proximity to the ASIC die that operates at high temperatures. However, lasers—particularly at the high output powers necessary for advanced data rates—are the most likely component to fail within a connectivity solution when subjected to high temperatures, potentially taking down the entire link and potentially rendering the package with expensive compute units like GPUs worthless

Remote Light Sources

On the other hand, remote light sources, also known as disaggregated or external lasers, are independently packaged and physically separate from the electronic and photonic components, including drivers, modulators, photodiodes, etc. As a result, remote lasers are not subject to the same intense temperatures of integrated light sources. This allows for designs to enable them to operate under less demanding thermal conditions, extending their operational lifespan and dramatically reducing failure rates. 

Remote lasers can also easily be serviced or replaced without compromising the other components in a system. A replacement laser can be introduced and the ASIC package can continue to function.

The industry has recognized the importance of external, replaceable lasers and established the External Laser Small Form-Factor Pluggable (ELSFP) specification. This common form factor leverages both the serviceability and replaceability benefits of pluggable modules, and CPO’s cost, latency, and channel loss advantages, while also uniting the ecosystem of suppliers and customers around a single form factor.

Figure 2. Ayar Labs’ optical I/O solution, including two remote SuperNova™ light sources, is integrated and working in the same package with Intel Corporation’s Agilex FPGA technology.

Factors to Consider When Choosing a Light Source

The laser light source is often the most sensitive component of an optical connectivity solution from a cost and quality perspective. Designers and architects should prioritize the diversity of suppliers and standardized wavelength grids, such as the O-band LR4 grid that has been used for over two decades. This approach ensures a low design and supply risk and also establishes an attractive high-volume cost structure that is essential for the successful deployment of optical I/O. 

The formation of the Continuous-Wave Wavelength Division Multiplexing Multi-Source Agreement (CW-WDM MSA) has brought together a broad spectrum of industry stakeholders, including laser suppliers, transceiver manufacturers, CPO and optical I/O connectivity suppliers, and others, to promote interoperability across solutions and reduce dependency on any single supplier or technology. Such standardization efforts are key when enabling solutions for AI, HPC, and other high-value applications.

Choosing the right light source design involves balancing considerations around thermal management, scalability, flexibility, cost, and the specific application’s environmental constraints.

External Light Source of Choice: SuperNova

For many of today’s most demanding applications, like AI and HPC, a remote light source is the best fit for high-bandwidth, low-latency, high-reliability, and low-power consumption optical connectivity solutions. Ayar Labs’ SuperNova is a culmination of advancements in remote laser source solutions, bringing a significant paradigm shift to optical interconnect technology.

Our second-generation SuperNova light source is the industry’s first CW-WDM MSA-compliant 16-wavelength light source, which can drive 256 optical carriers across 16 optical fibers. This significantly increases the bandwidth capacity per fiber relative to conventional VCSEL-based active optical cables (AOCs) or other single-wavelength per fiber solutions like DR4/DR8, a level of bandwidth that is essential for AI workloads.

Figure 3. Ayar Labs’ 16-wavelength SuperNova™ light source

The Role of the SuperNova in AI Applications

The advancement of AI technologies and the exponential growth of large language models (LLMs) demand new data transfer solutions to keep pace with the exponential growth in model sizes and tokens. Remote light sources, like the 16-wavelength SuperNova, are critical enabling technologies for optical I/O solutions like the Ayar Labs’ TeraPHY™ chiplet. By increasing the number of wavelengths, we can increase the I/O bandwidth and radix to and from different silicon components in an AI system—compute elements, like CPUs and GPUs, memory elements, such as high bandwidth memory (HBM) or dynamic random access memory (DRAM), or connectivity solutions like switches.

Traditional systems rely heavily on switches for multiple connections, which introduce a large latency and limit the size of high-bandwidth domains. SuperNova’s multi-wavelength, multi-port capabilities make it possible to offer direct, low-latency connections between a larger number of devices. This approach simplifies the system architecture, enhances the speed and scalability of data exchange across the network, and enables significant innovation in system architectures. 

One of the biggest challenges in AI architectures is the “memory wall,”—the rapidly increasing memory-to-compute ratio leading to operational efficiency being bottlenecked by the low-latency bandwidth available to move data between processors and memory units. The SuperNova light source, when combined with TeraPHY, relieves this bottleneck by enabling large disaggregated memory clusters connected to GPUs via high bandwidth, low latency optical I/O.

By powering a single low-cost, low-latency, and high-bandwidth data transfer solution, SuperNova and TeraPHY enable more efficient connectivity between a variety of compute elements. This allows system designers to create vast, scalable AI networks while also empowering them to train larger, more complex AI models that demand increased computational resources.

Technology in Action: Revolutionizing Telecom with Remote Light Sources

One recent example of the benefits of a remote laser source is our joint concept with Corning and Ericsson. Traditional mobile networks need to rapidly transform at the edge to handle the exponential growth in traffic and high-value, performance-hungry AI-driven services. This requires smarter radios with significantly enhanced processing power and I/O throughput than is possible today.

Ayar Labs is working with Ericsson to enable smarter, more powerful, and AI-ready radios to handle the explosion of edge traffic. This joint solution integrates TeraPHY optical I/O chiplets with radios to provide massive I/O density to each ASIC, which are interconnected via a revolutionary Corning glass waveguide module with built-in optical connectors. This innovative architecture replaces onboard electrical connections with a compact network of optical links, enabling a bandwidth density that cannot be achieved with bulky, low-throughput pluggable optics.

Figure 4. A conceptual model illustrating the integration of Ayar Labs’ TeraPHY™ optical I/O chiplets with Corning’s waveguide module and a remote SuperNova light source.

This solution requires the SuperNova light source to be disaggregated from the radio and placed in a remote, controlled location away from the harsh environmental conditions experienced by telecom deployments. This makes the solution far simpler, faster, and safer to service while providing network operators with reduced costs and network downtime. Greater compute density in the radio enables data and user requests, such as AI inferencing, to be processed more effectively at the edge, greatly reducing latency for customers and data volumes on the backhaul network. The end result is a future-ready, AI-enabled mobile network that delivers more performance and capabilities than ever before while also reducing operating costs and complexity.

It is an exciting time at Ayar Labs. Our technology is poised to redefine the capabilities of AI infrastructure through the power of optical I/O by tackling the existing limitations of data transfer and addressing the future needs of computing. With SuperNova, we are helping set the pace for AI’s continued rapid growth.

To learn more about the future of optical connectivity and how the SuperNova remote light source is revolutionizing AI infrastructure, visit: https://ayarlabs.com/supernova/.

Join our mailing list

Recent News

More News →

Resources

More Resources →

Follow Us

Related Blog Posts

Optics on SC23

Optics on SC23

Discover the key insights and groundbreaking advancements from SC23 with Ayar Labs’ top observations. Dive into the future of HPC and AI.

See More

Pin It on Pinterest

Share This