Paper Title: Optical Chiplets Meet the Demands of AI Compute Clusters

Paper Abstract: AI applications (such as ChatGPT®) will surely demand enormous computing power. AI clusters will require low-latency networking at terabit speeds. One way to scale AI systems is with optical connections rather than electrical ones. However, today’s optical interfaces are large, expensive, and power-hungry. Optical chiplets to the rescue! Such chiplets are small and use much less power than pluggable optics and can just be one more standard part in the package. And they also avoid the need for retimers or DSP circuits that drive long copper interconnects. They will deliver energy efficient, low latency, and dense optical connectivity needed for AI compute clusters.

Paper Author: Jeff Hutchins, Director CTO Office, Ranovus
Hojjat Salemi, Chief Business Development Officer, Ranovus

Author Bio: Jeff Hutchins is a director in the CTO Office at Ranovus, a developer of optical interconnect and co-packaged optics. He has been active in standards development and is currently a vice chair of the Energy Efficient Interfaces effort at the OIF. He is a board member of the OIF and has also co-authored white papers and delivered presentations at many conferences including OFC. He earned an MSEE in electron devices from Stanford University and holds 5 patents.

Author 2 Bio: