The open future of networking hardware for AI

> Over the past year we have developed a Disaggregated Scheduled Fabric (DSF) for our next-generation AI clusters to help us develop open, vendor-agnostic systems with interchangeable building blocks from vendors across the industry. DSF-based fabrics allow us to build large, non-blocking fabrics to support high-bandwidth AI clusters.

Is that somewhat akin to Infinibands credit-based flows, where someone is making a decision about who is allocated traffic.

It'd be really cool to see this takeoff. The OCP NIC is seemingly the only piece of OCP that is really truly well adopted. OCP servers and power supplies are pretty rare I think. Some of the related more software side things seem to be ok - OpenBMC & some of the related specs.

But like, networking has very few options right now. Ultra Ethernet Consortium clearly wants to improve but I'm not sure what pokers they have in the fire, what their focus really is. Nvidia owns Marvell, which was the last Infiniband standing, and they have Nvlink. A company bought Intel's pretty excellent Infiniband derived Omni-Path, Cornelis Networks, but not sure where if anywhere they have presence.

Really miss the days when Intel server chips came with variants that has OmniPath. Interconnect on chip should definitely be beyond the pizza-box, but we seemi fly regressed to the semi-costly space-using NIC model. Boo.