Advanced Photonic Technologies for Future Systems: Optical Computing or Optics for Computing?

Author: Bert Jan Offrein, IBM Research Europe – Zurich, Rüschlikon, Switzerland; University of Twente, Enschede, The Netherlands Collection: 2026-OCP-EMEA-Summit-Photonics

Summary

Bert Jan Offrein of IBM Research Europe frames the core question facing the photonics industry: is optics best deployed as a communications medium within computing systems ("optics for computing") or as the computational substrate itself ("optical computing")? He grounds the discussion in a striking historical data point — the IBM Roadrunner supercomputer, the first petaflop system built in 2008, would now need to run for 10,000 days to train a single modern AI neural network — illustrating how AI's compute demands have outpaced historical infrastructure by many orders of magnitude. The central compute bottleneck is matrix multiplication, which accounts for 99.9%+ of all operations required for AI inference and training, and scales as O(N²) with network size while neuron activations scale only as O(N).

Offrein surveys three technology domains where IBM has been active: (1) optical interconnects for computing systems, tracing IBM's roadmap from the 2008 Roadrunner (first use of pluggable optics) through the 2011 Power 775 (co-packaged optics with 56 VCSEL-based transceivers per switch hub) to future integrated waveguide approaches; (2) adiabatic coupling via polymer waveguides as a manufacturable solution to the chip-to-fiber size mismatch problem (SOI waveguides at 0.5×0.2 µm versus 9 µm fiber core); and (3) in-memory optical computing using cascaded Mach-Zehnder interferometers on SiN waveguides to perform convolution operations at 30 gigasamples per second. The full power budget analysis of the SiN lattice filter at 32 GSps yields 2.49 TOps/W — comparable to GPU performance, not clearly superior.

The key conclusion is that optical in-memory computing does not yet offer a power efficiency advantage over digital hardware, though it uniquely enables ultra-fast signal processing at 30+ GSps. Optics for computing (communication) remains the justified near-term investment, with dedicated optical computing applications as a plausible but longer-term trajectory.