With the boom in generative AI technology, we are stepping into a new era where everything is stored, everything is smart, and everything is connected. In this era, the flood of data is coming in like a tidal wave, posing unprecedented challenges to data center infrastructure. To meet the growing computational demands of AI models, Large Language Models (LLMs) need to simultaneously process massive multimodal datasets, including text, images, audio, and video, which has driven the demand for AI processing resources to skyrocket and require efficient interconnections across the data center.

In the face of this challenge, traditional data interconnections have struggled to meet the scaling needs of modern AI clusters. With the diversification and customization of AI platform architectures, as well as the significantly accelerated pace of annual upgrades, data centers must find more efficient and reliable connectivity solutions. Against this backdrop, "PCIe over optics" technology has emerged, offering new possibilities for expanding and upgrading AI infrastructure.

The PCIe (Peripheral Component Interconnect Express) interface, which is natively available on AI gas pedals and GPUs, has always played an important role in connecting various components within the data center. However, as AI clusters continue to grow in size, from a few racks and dozens of GPUs to large pods spanning more than a few racks and hundreds of GPUs, interconnect lengths are quickly becoming a key constraint to performance. While active cables are able to cover distances up to 7 meters at PCIe 5.0 data rates, at higher data rates such as PCIe 6.x and PCIe 7.x, traditional copper solutions can no longer meet the demand for cross-rack connectivity.

AI Interconnection Challenge Response

Since 2017, Astera Labs has been committed to unlocking the potential of AI and cloud infrastructures with its Intelligent Connectivity Platform, which builds scalable, customizable architectures based on PCIe®, CXL® and Ethernet semiconductor solutions and the COSMOS software suite.

The platform enables reliable connectivity over long distances, accelerating GPU cluster deployments today with optics that enable PCIe-to-line connectivity; shortening AI platform deployment time with software-defined architecture and interoperability testing; and improving infrastructure uptime and utilization with features such as deep diagnostics.

Its broad product portfolio includes the Aries®PCIe®/CXL® Smart DSP retimer with its high bandwidth and wide range of applications, the Aries PCIe/CXL Smart Cable Module (SCM), which provides 7-meter active cables for rack connectivity, the Taurus Ethernet Smart Cable Module (SCM), which supports high-speed Ethernet connectivity, and the Leo CXL® Smart Memory Controller, which optimizes memory management. Astera Labs excels in technology innovation, laying the foundation for large-scale deployment of advanced solutions.

Starting a new chapter in seamless AI connectivity

As AI infrastructures scale, traditional passive direct-attached cables (DACs) are struggling to keep up with demand. Signal loss issues are exacerbated at high data rates, limiting the effectiveness of passive solutions. In response, Aries PCIe/CXL® SCM™active cables (AECs) were created with 7-meter coverage to address the connection distance limitations of DACs, providing more possibilities for AI gas pedal cluster expansion, with cabling that can be easily extended outside the rack. However, as data rates increase to higher levels such as PCIe 6.x and 7.x, passive and active cables will be limited to a single rack. At this point, PCIe over optics solutions, especially active optical cables (AOCs), will play a key role in inter-rack connectivity, providing long-distance, high-stability connectivity to ensure the efficient operation of AI clusters.

Exploring the Future of PCIe Optical Connectivity

 Fiber optic links have become the cornerstone of high-speed Ethernet connectivity, supporting long-distance data transfers within hyperscale data centers. Incorporating this advantage into PCIe connectivity by developing new PCIe over optics solutions (with AOC) significantly extends the range of PCIe connectivity to rack clusters and optimizes cable management compared to copper cables.

The use of PCIe/CXL® in optics is driven by the need for lower latency compared to Ethernet, especially in high load applications such as cache coherent memory transactions and parallel processing between GPUs. In addition, these applications require dedicated software to fully manage the link to ensure full protocol compatibility and high system reliability.

In short, the emergence of "PCIe over optics" technology provides a new solution for data transmission within the data center. It not only improves the efficiency and stability of data transmission, but also reduces operation and maintenance costs, providing strong support for the accelerated deployment of AI platforms and the rapid development of next-generation generative AI applications. With the continuous progress of technology and the expansion of application scenarios, "PCIe over optics" is expected to become one of the mainstream technologies in the field of data storage and transmission in the future.