Sponsor Talk
Nick Harris
Founder and CEO of Lightmatter
Title: BiDi DWDM: The Interconnect Foundation for High-Performance AI Training and Inference
Abstract:
Photonic interconnects are solving frontier model training constraints, but inference now presents an even more consequential challenge. Inference accounts for 80 to 90 percent of AI compute cycles and demands high interactivity, while balancing varied batch sizes and context windows up to one million tokens. Interconnect bandwidth, radix, and hop count set the performance ceiling for agentic AI workloads. On the supply side, tightening optical fiber supply raises infrastructure costs and constrains buildout. This talk analyzes these requirements and presents bidirectional dense wavelength-division multiplexing as the foundational physical layer for multi-rack scale-up, scale-out, and scale-across architectures, using Passage interconnect and Guide lasers as the hardware blueprint.
Compute performance has scaled a millionfold over the past thirty years; interconnect bandwidth roughly a thousandfold — a 10,000x gap that widens as accelerator density compounds. Closing it demands a different optimization principle: prioritizing per-fiber capacity over modulation complexity, a conclusion the Optical Compute Interconnect (OCI) Multi-Source Agreement has now reached, and one Lightmatter established years ago by designing from first principles. Bidirectional single-mode fiber links double logical radix and halve fiber and connector requirements. Dense wavelength-division multiplexing (DWDM) maximizes bandwidth using NRZ modulation to minimize FEC overhead. Lightmatter silicon reflecting three OCI generations is being validated today, powering a roadmap spanning near-packaged optics, co-packaged optics, and photonic interposer form factors.
Biography: