As AI continues to evolve, the need for new computing models and advanced data transmission methods has never been more crucial. PCIe over Optical is emerging as a vital technology to support the increasing demands for high-performance, low-latency connectivity, which is essential for the growth of AI.
The data centre industry is currently undergoing significant changes, driven by the rapid expansion of AI models. For example, ChatGPT has seen a massive increase in parameters from 1.5 billion in version 2.0 to 175 billion in version 4.0, with future generations expected to utilise around 100 trillion parameters, matching the number of synapses in the human brain. This surge in AI complexity and the resulting explosion in data consumption are placing immense pressure on data centre networks. Global data creation is projected to surpass 180 zettabytes by next year, a significant increase from previous years.
To manage these growing data volumes, robust connectivity solutions capable of high-speed data transfer with minimal latency are essential. Disaggregated computing, where memory and storage are shared in centralised pools, is becoming increasingly important in this context. Low-latency connectivity solutions are critical to the success of this model, especially as distances increase and data rates rise. Linear pluggable optics (LPOs) are seen as a potential solution, providing longer reach with minimal latency and power consumption. Unlike traditional optical transceivers, LPOs eliminate the need for clock data recovery (CDR) and digital signal processing (DSP) components, resulting in reduced power consumption and latency. However, this approach requires advanced SerDes technology to ensure efficient data transmission.
Interoperability is a key challenge in implementing LPOs. The Optical Internetworking Forum (OIF) has been working on the CEI-112G-Linear specification since 2020, and recent demonstrations have shown that multi-vendor LPO configurations are possible. These efforts are paving the way for PCIe to increase bandwidth while reducing power, latency, and costs. The IEEE 802.3 specification has already established optical budgets for CEI-112G-Linear, ensuring that links can be closed without the use of DSPs. At the OFC 2024 event, Alphawave and several other companies demonstrated the interoperability of LPOs, marking a significant step forward.
Alphawave also conducted an un-retimed optics study using its PipeCORE™ PCIe 6.0 subsystem IP, achieving impressive performance results with BERs of less than 1×10-9. Further optical demonstrations have been conducted by Alphawave and its ecosystem partners, generating excitement for the future applications of PCIe and CXL over Optical.
Despite the progress, challenges remain, particularly in managing optical component nonlinearities, part-to-part variation, and large-scale deployment. However, the potential benefits of PCIe over Optical, including higher bandwidth, lower power consumption, and reduced latency, make it a promising solution for the future of data transfer.
As the data landscape continues to evolve, PCIe over Optical is poised to play a critical role in enabling efficient, high-speed data transfer. With PCIe 7.0 and faster rates unlikely to work across standard copper connections, technologies like retimers and LPOs are becoming increasingly essential. The ecosystem supporting these technologies is growing, with ongoing advancements and real-world demonstrations proving their capabilities. PCIe and CXL over Optical are set to revolutionise data centres in the near future.
The ongoing development of PCIe over Optical is not just addressing current challenges but is also setting the stage for the next generation of data centres. With AI demands increasing, this technology is well-positioned to provide the necessary connectivity solutions to support the future of data processing and transmission.
Alphawave IP Group plc (LON:AWE) is a semiconductor IP company focused on providing DSP based, multi-standard connectivity Silicon IP solutions targeting both data processing in the Datacenter and data generation by IoT end devices.