Some of the best things about Ethernet are the range of vendors, the competitive pricing and the sheer ubiquity. If that happens with U/E then that will be nice.

Color me skeptical.
If I get it right, the plan is to make Software think it is using Ultra-Ethernet by the usual SHIM/ABstraction layer, and then the frames on the wire will be some “Native” transport, such as Omnipath? this means all switches along the way need to know OP?
The point is to have multiople vendors offering the HW, so it is as cheap as Ethernet, and as functional as OP or IB (and better), not to just be compatible at ehe SW level

Functionally, “Ethernet” today is a frame-format, and not much else. It does not cover Layers 0 and 1 itself (tough IEEE
defines L0/L1 interfaces, such as PHY’s and their functionality, mostly to fit Ethernet Frames being Sent/recieved), and a lot of what used to define “ethernet” is simply omitted/obsolete (CSMA/CD, anyone?)

Yes, thanks for the catch. And good to hear from you. It has been a while. . .

Cheers,

Bruce

Haha! And what about ROcE. There is no such thing as Converged Ethernet!

Ethernet is a well defined protocol stack covering layer 0 to layer 2 if I recall the ISO OSI model correctly. If you strip out parts of it leaving just enough to move upper layer protocols I would say you no longer have Ethernet. Just like if you take all the stuff that was built into the IB spec that lets one actually manage and reconfigure a network, and you add it to Ethernet, well that might as well be IB with a different frame format. Really, the reason to do that is to get something as functional as IB without being beholden to the Borg, er, ah, to nVidia.

At the hyperscalers and cloud builders, Ethernet is already not Ethernet. They have stripped down things with comparatively limited protocol support. If it runs BGP and looks like it is running TCP/IP, will it still be Ethernet?