²ÝÁñÉçÇø

Data centers are currently grappling with the escalating need to boost bandwidth capacities. This is primarily driven by the emergence of new technologies, particularly the growing demand for AI/ML applications. As these technologies evolve, bandwidth requirements are expected to grow exponentially. Specifically, as Large Language Models (LLM) become more precise and expansive, they demand increasingly higher processing speeds. This surge in demand for rapid processing of LLM models has highlighted inefficiencies within data centers. This technical bulletin will delve into the realm of PCIe over Optics, a promising solution for escalating bandwidth demands in data centers. We¡¯ll explore resource limitations, latency challenges, and energy consumption.  

PCIe the interface of choice within rack servers, linking resources together through copper cables or a backplane. With over seen over six generations of deployment, and the ratification of the PCIe Gen 7.0 specification coming up, PCIe will continue to be a key player for high-speed interconnects. Figure 1 illustrates the full stack of data communication over the PCIe link, alongside its associated components. 

Resource limitation 

Current data centers are experiencing efficiency challenges due to memory bandwidth and memory utilization. The restriction of only accessing local memory not only limits the speed of data processing but also leads to underutilization of data center memories. This occurs despite the evolution of processors to include more and faster cores. 

Latency  

Latency currently poses a significant bottleneck for most AI/ML applications. The transfer of high data rates and complex modulation schemes over copper cables, backplanes, requires the use of advanced equalization techniques and algorithms such as forward error correction (FEC), but these further contribute to system latency.  

Energy consumption  

Power is the scarcest resource in data centers, and current technologies demand the utilization power-hungry chips. An estimated 25% of a data center¡¯s total power is used solely for point-to-point data transfers. As the need for data transfer grows, particularly with the advent of AI/ML applications, this energy consumption is predicted to rise dramatically. 

Scaling challenges  

Demand for data transfer and data processing is going higher and higher with emerging requirements and technologies, this would directly lead to higher memory and faster memory access. Datacenter growth requires network architecture to scale accordingly and designing networks that could be scaled without too much financial burden becomes very important. The ability to scale resources up or down based on demand is crucial for AI workloads, which can be highly variable.

Optical links offer higher bandwidth density compared to electrical links. Initially, PCIe interfaces were developed to be utilized over copper, DAC, and PCB interconnects. However, as data rates increase and electrical losses escalate, this approach is becoming less appealing. 

Optical links have the advantage of covering longer distances. Resource limitations, particularly memory constraints, are becoming increasingly challenging to address using the current architecture of PCIe over copper, which only permits access to local memory. Optical technology, however, can overcome this limitation by enabling different processing units to access further memory units in different server units or racks. This is beneficial for resource pooling or sharing over CXL switch and other similar applications. 

When it comes to energy efficiency and cost-effectiveness over longer distances, optical links excel. They are far less lossy compared to electrical links, which means they demand fewer re-timers and signal conditioning units over the same distances. Additionally, the use of low-cost, high-yield optical components could further reduce costs per distance. Copper interconnects, on the other hand, occupy a lot of space in data centers and are not suitable for dense data centers. In contrast, optical fibers are more flexible and take up less space, making them a better option for increasing density in data centers. 

Finally, linear or direct drive optical links can also help to reduce latency and power consumption. Different optical architectures can be employed for PCIe over optics, leading to improved latency. For instance, linear direct drive optics avoids an extra timer in the link, resulting in reduced latency. 

Figure 2 shows a PCIe over optics use case scenarios for data center intra rack and rack to rack configurations based on requirements from OCP (Open Compute Project). These applications range from compute, storage, accelerator, and memory connectivity scenarios for NVMe & CXL enabled disaggregated data centers. 

The retimed topology is a key approach where a maximum of two retimers are permissible within an end-to-end link. Some important aspects to consider within this topology include the strategic placement and the precise quantity of retimers deployed. 

Conversely, the non-retimed or linear topology, introduces a more complex set of challenges. This is primarily because a linear link disrupts the continuity of the path, making it more difficult to reconcile with the existing PCIe standards and compliance stipulations. Effective regulation of channel losses is paramount in this topology. Moreover, it may necessitate substantial alterations to the protocol layer, and potentially to the PHY layer as well. A comprehensive feasibility study with all types of optical engines is also a critical aspect of this topology. 

PCIe over optics is essential for establishing interconnectivity among rack units, thereby enabling them to operate as a cluster. The role of PCIe is central as it acts as a controller ¡ª the digital logic that interfaces with a particular software. One of the major hurdles is to ensure the transition to optical PCIe does not disrupt the control process of the software stack. 

An even greater challenge is managing the physical layer and the interoperability of the electro-optical interface. ²ÝÁñÉçÇø, in collaboration with OpenLight, plays a critical role here by providing an electrical IP solution that can function alongside a photonic one. Once a universal standard is established, any vendor of photonic die will be capable of integrating PCIe. ²ÝÁñÉçÇø and OpenLight showcased during OFC 2024 the world¡¯s first PCIe 7.0 data rate demonstration over optics, using a linear drive approach, in addition, we also featured a PCIe 6.x over optics demo. This demonstration showcased end to end link BER performance  orders of magnitude better than the FEC threshold, showcasing feasibility of PCIe 7.0 over optics running at 128Gbps PAM4. This performance was achieved using discrete electrical and optical components to build the PCIe over optics link. As demonstrated during OFC24, the same ²ÝÁñÉçÇø SerDes that drives electrical PCIe links with an excellent PPA and latency was not limited by this non-ideal and worse case use case scenario, showcasing flexibility and robustness of ²ÝÁñÉçÇø SerDes. 

In an age defined by AI/ML and its consequent bandwidth-demands, it is clear that PCIe over optics represents the future of signaling. Its development and adoption depends on the enablement of a supportive ecosystem, which ²ÝÁñÉçÇø is actively pursuing. , with ongoing interoperability demonstrations and excellent field results with PCIe 7.0 data rate and PCIe 6.x over optics, help reduce integration and risk and make first-pass silicon success possible.