# Optical Interconnects: enabling large-scale AI clusters

## Broadcom is developing foundational optical technologies that lower AI infrastructure power and costs, while enhancing system reliability

Broadcom leads the industry in three core optical interconnect technologies essential to building large-scale AI networks.

AI server clusters are large and are only going to get larger. Optical links provide the connections needed to allow these clusters to scale at distance. Optical technology can support that growth in terms of both scale and cost.

#### The Core Optical Interconnects for AI

There are three core optical interconnects:

- [Vertical-Cavity Surface-Emitting Lasers](https://www.broadcom.com/products/fiber-optic-modules-components/components-broadband/short-reach-gaas-die/lasers), or VCSELs, are a workhorse for optical AI interconnect technology across the industry. Its low power and low cost make it ideal for data communications and sensing applications. The only limitation is that it operates best at shorter link distances.
- Electro-absorption Modulated Lasers, or EMLs, are ideal for AI systems that scale to greater distances and hundreds of thousands, or even millions of units. This technology provides better performance at very high bandwidth and typically are first to reach volume deployments at next generation data rates.
- [Co-Packaged Optics](https://www.broadcom.com/info/optics/cpo), or CPO, is an advanced heterogeneous integration of high-speed silicon photonics onto application-specific integrated circuits for addressing next-generation bandwidth and power challenges. We see this new technology providing the power and cost leadership for future generations of AI systems and enabling infrastructures supporting large-scale AI networks.

#### Leading the Way

Broadcom VCSEL technology has led the industry in time-to-market, volume, performance and reliability at 25, 50 and 100 gigabit per second per lane (G/lane). We are also very encouraged by our progress on 200G/lane VCSELs, an achievement that many thought impossible technically as it was assumed the technology could not sustain the data rates needed for the future of AI. We have also extended our EML technology from 100G/lane to 200G/lane and will be shipping large scale production volumes of these new 200G/lane EMLs very soon.

#### Foundational Technologies

Just a few months ago, we had our first commercial shipments of our [Bailly 51.2-Tbps CPO Ethernet switch system](https://investors.broadcom.com/news-releases/news-release-details/broadcom-delivers-industrys-first-512-tbps-co-packaged-optics). This is a combination of our 51.2-Tbps Tomahawk 5 switch with complete optical links directly integrated into the package. What that means is that all 512 lanes are directly, optically attached to the Tomahawk 5 switch itself. This is a foundational technology for AI infrastructure that could scale for all kinds of applications where optics are required.

Why is CPO needed? With AI systems, the bandwidth and the amount of components continues to scale. As it does, the cost of the optics continues to rise. Our solution is integration into silicon photonics, allowing us to put more components onto an individual chip. In the history of semiconductors, that has been the proven way to reduce costs.

Another benefit is that the optics can be placed on a common package with the core chip. That placement eliminates the complex electrical lanes between the ASICs and the optics. Typical 800G pluggable transceivers consume about 16 watts per link. The Bailly CPO system would reduce that optical link power to 5 watts, a 70 percent savings over typical deployments today. Perhaps even more impressive, it scales and delivers better efficiency when the optical link migrates to 1.6T, typically pluggable transceivers will consume 25 watts vs. 8 watts with CPO. That’s a big deal in terms of both power savings and the optics.

A third major benefit is enhanced reliability. Pluggable transceivers have approximately a 2 percent failure rate. By integrating more components into the chip, reliability is enhanced with CPO. Rather than integrate lasers directly on silicon, what we have done is maintain the laser system as a pluggable and easily replaceable component in the system. Everything else is built on core silicon technology with its long history of reliability.

Broadcom is focused on providing silicon photonics in a high-density fashion. Doing so allows us to move the optics off the subsystems that are these pluggable transceivers and move them directly onto ASIC substrates. We are doing exactly that for the first time with our [Tomahawk 5 switch](https://www.broadcom.com/products/ethernet-connectivity/switching/strataxgs/bcm78900-series): 512 lanes of optical connectivity across the entire device.

This is just the beginning. We see this technology being extended to many areas. It is foundational for enabling AI infrastructure and driving the entire technology industry forward. We see CPO engines being applied to various types of ASIC substrates for scale-up and scale-out AI interconnects. Customers essentially get the performance of pluggable transceivers but with lower power, costs, and a tremendous reliability advantage. Reliability that will enable AI infrastructure and allow our customers to sleep at night.

To learn more and watch my recent presentation, [please visit our page here](https://www.broadcom.com/company/events/ai-day). You can also listen to my [podcast interview here](https://www.bloomberg.com/news/audio/2024-07-02/broadcom-s-cpo-will-connect-massive-ai-clusters-tech-disruptors) with Bloomberg Intelligence.

### Topics

[optical interconnects](https://www.broadcom.com/blog-search?filters[pages][blog_topics][values][]=optical%20interconnects) [optics](https://www.broadcom.com/blog-search?filters[pages][blog_topics][values][]=optics) [large-scale AI clusters](https://www.broadcom.com/blog-search?filters[pages][blog_topics][values][]=large-scale%20AI%20clusters) [ai infrastructure](https://www.broadcom.com/blog-search?filters[pages][blog_topics][values][]=ai%20infrastructure)
