Ep. 76: GlobalFoundries VP on Optical for Data Center Scale-Up

Before AI, data centric internet was straightforward.

Anything local, you would use copper or electrical connectivity and anything long distance, you would use optical.

But with AI coming in, all of that has changed.

With AI, where you have increased demand on the interconnect speeds, there are larger models coming up every day, newer and larger models that need huge amounts of GPUs that need to be interconnected, as well as connected to fast memory with extremely high data speed links.

You know, as this evolved, data center interconnect became a bottleneck.

And the need for higher and higher speeds, you know, rapidly increased.

It went from 200 Gbps to 400 to 800 and now moving on to 1.6 terabits per second.

According to some estimates, interconnect speeds are kind of doubling every 1.5 to 3 years.

And there's no sign of that stopping.

So this is making the whole interconnect technology move from electrical to optical.

So in today's AI data centers, scale-up connectivity, which is within the rack, is mostly electrical.

But rapidly moving to optical, a scale out, which is connectivity between racks within the same cluster is mostly optical and scale lower, which is connecting different data centers, which has always been optical and continues to be optical.

And we'll focus today's discussion specifically on scale up, which is again within the rack.

And we look at the market needs, challenges, current solutions and future trends.