Tuesday, 26 March, 19:30 – 21:30

Room 6F

How Much Optics Does AI Need?

AI systems have attracted enormous interest over the past couple of years and commensurate investments into the AI infrastructure for data processing and movement. This rump session will debate how deep optical technologies should penetrate AI clusters. Topics include:

Processing: Substantial efforts have gone into building all-optical compute engines; will those be the future for AI, or is data processing best left to CMOS?

Switching: Will all-optical switching complement or even replace electronic switching for AI cluster networking? If so, at what switching granularity – circuits, bursts, or packets?

Data Movement: Undoubtedly, the proven strength of optics is in data movement. What form will interconnects take in AI systems? Will Ethernet continue to be the driving design paradigm, or will the AI era give rise to new protocols?

Computer I/O: How will computer IO evolve in the AI era? With per-lane speeds saturating, will this change the architecture? Will active optical substrates/interposers connect chiplets? Or will chipletized co-packaged optics on electrical substrates be the better choice?

Packaging: When will near-package optical engines be deployed in AI systems, or will standard pluggable optics prevail? Are single-wavelength or WDM approaches the preferred architectural choice for AI?

Organizers