And it was such an outstanding success! Intel's customers were able to get so much use out of it - it'll always be remembered as a high point in supercomputing.
Seriously - it was a resounding failure, even with Intel spending big bucks to convince this class of customers it was the future. There was little uptake and vast dissatisfaction with achieved results. IMHO I expect Xe to follow the same (very shallow) trajectory in HPC as in gamer graphics.
There is one remaining large machine (DUG "Bubba") using Xeon Phi and that's only an artifact of software tolerance combined with a _really_ great deal received on Intel's remaining "Knights Whatever" parts inventory.
But whatever real innovation underlay Xeon Phi/KNx was performed two decades ago. Current Intel is only capable of SKUsmanship - minutely partitioning function for projected max margins.
Edit: Also too, remember that KNx was really an attempt to salvage Intel's failed Larrabee GPU effort.
> Also too, remember that KNx was really an attempt to salvage Intel's failed Larrabee GPU effort.
I do, and I don't think you know that the first "Knights" project was simply a rename of Larrabee. Same exact product. It was called Knights ... Landing I think? Then it became Xeon Phi.
Knights Ferry was the 1st. Unclear about the actual diffs between it and Larrabee silicon. I do know the 1st boards still had a video out port but it was disconnected.
Seriously - it was a resounding failure, even with Intel spending big bucks to convince this class of customers it was the future. There was little uptake and vast dissatisfaction with achieved results. IMHO I expect Xe to follow the same (very shallow) trajectory in HPC as in gamer graphics.
There is one remaining large machine (DUG "Bubba") using Xeon Phi and that's only an artifact of software tolerance combined with a _really_ great deal received on Intel's remaining "Knights Whatever" parts inventory.