Has anyone tried linear algebra on a Threadripper yet?
I would expect the dual-die approach with the split memory controllers to make it a bit slower than Haswell-E (and to provide some benefit from messing about with NUMA controls and MPI), but you have an awful lot of cores to drive those memory controllers.
(I won't be buying a new machine until March, but I notice the Intel lots-of-core-Xeons-at-less-absurd-price range has now officially appeared)
Last fiddled with by fivemack on 2017-09-25 at 11:15
|