[benbenolson]

Back in the day when Xeon Phi was around, I'd run `make -j 256` to run on the ~240 available hyperthreads. Those things were build machine beasts, assuming there weren't too many dependencies. For example. the Linux kernel would build files approximately ~240 at a time, which greatly sped up the build process, but linking was extremely slow (single-threaded on one very slow Phi core).

Even more interestingly, the Knights Landing series had a PCIe coprocessor version, which ran a stripped-down Linux kernel, and you could SSH onto it. One of my friends got one for free at a conference, and I really wish I'd picked one up!

[Agingcoder]

Yes I had knl at some point . I tried it, tuned my code to work ok it, and since many optimizations carrried on to regular xeons … ended up buying Xeons.

[jacquesm]

Don't you get limited by memory bandwidth at that point? (assuming all disk contents are cached)