Authors: Dan Olds (Olds Research), Phil Pokorny (Penguin Solutions), Andy Mills (Penguin Solutions), Preet Virk (Marvell), Luis Ancajas (Micron Technology)
Abstract: As AI and HPC systems scale to unprecedented levels, traditional electrical interconnects are becoming a limiting factor. Silicon photonics offers a promising path to address bandwidth, energy efficiency and interconnect scalability constraints, but questions remain around deployment readiness, integration complexity and ecosystem maturity. Led by industry analyst Dan Olds, this session brings together experts from Penguin Solutions, Marvell and Micron to explore photonic interconnects from multiple perspectives. Panelists will discuss the technology’s value today, the technical and operational challenges ahead, where photonics can make the greatest impact, and what the broader ecosystem must do to accelerate real-world adoption.
Long Description: As the rapid growth of AI and HPC workloads continues to place unprecedented demands on system architectures, bandwidth, latency, energy consumption, and interconnect scalability have emerged as critical constraints to overall application performance. Traditional electrical interconnects have reached practical limits, prompting interest in silicon photonics as the enabler of next-generation HPC and AI infrastructure.
However, despite the growing interest in photonic technologies, questions remain regarding their practical deployment, operational tradeoffs, and readiness for production-scale systems.
This session will provide a forum for the HPC community to examine photonics beyond research demonstrations and technology roadmaps, diving into realistic deployment scenarios, measurable benefits, and the technical challenges that must still be addressed.
The goals of this session are to:
• Examine the architectural challenges driving industry interest in photonic interconnects for AI and HPC systems.
• Discuss where photonic technologies are delivering measurable value today, including improvements in bandwidth, latency, energy efficiency, and scalable memory architectures.
• Explore the engineering, operational, and ecosystem challenges that continue to limit broader adoption.
• Foster collaboration among system architects, vendors, cloud providers, researchers, and HPC practitioners on the future evolution of photonic infrastructure.
The backdrop for the discussion is the architectural pressures driving adoption—specifically the growing memory wall—and will examine how photonic interconnects can enable new approaches to memory disaggregation and high-performance communication while reducing dependence on electrical signaling.
The session will explore beyond theoretical advantages, examining practical deployment considerations and covering topics including integration with existing system architectures, interoperability with complementary technologies such as CXL, packaging and thermal challenges, operational complexity, reliability, software implications, and the performance tradeoffs organizations should consider when evaluating photonic solutions.
The topic is highly relevant to SC26 attendees, as their organizations are actively planning future AI and HPC infrastructure investments in the face of growing constraints in memory architecture, energy consumption, and interconnect scalability. Researchers, system designers, infrastructure providers, hyperscalers, and app developers all have a stake in understanding the state of photonics in AI and HPC systems.
The panel will be led by longtime HPC thought-leader and frequent SC speaker Phil Pokorny, CTO of Penguin Solutions and moderated by independent data center analyst Dan Olds. It will also include: Andy Mills, a memory expert with 30 years of industry experience; and industry expert from Marvell Technology.