From photonic devices to the interconnect fabric: enabling AI training clusters through integrated photonics and advanced assembly

Business as usual scaling of compute for AI will soon hit bandwidth and latency walls, whereby processing units will sit idle while waiting for new data to process. At least part of these bottlenecks can be addressed through deployment of optics deep into these compute systems.

This talk highlights imec’s integrated photonic device roadmap, and shows how it can be combined with advanced assembly technology to realize optical interconnect with ultra-high bandwidths and low energy consumption. The same technology platform can also be used to remove latency bottlenecks from today’s AI training clusters.