Once again, @ErrataRob derails my brilliant plan to build an MSP430 workstation and server platform.https://twitter.com/ErrataRob/status/980907462973902849 …
Personally my only use for cpu perf is big compile jobs, which are (modulo awful recursive-configure scripts) 100% //izable and don't share data. But yeah it varies with what you're doing.
-
-
I'm in a similar boat (also want SMT solvers, but their execution characteristics seem roughly similar to compilers, but more memory-bound).
-
seems like the keys are memory bandwidth and huge last-level cache. so far an ARM farm doesn't do anything for me, but I haven't got hold of a Thunder X2 box yet.
-
My suspicion is that for memory-bandwidth constrained tasks that CPUs designed to go to sleep faster (and not consume power) will be more power efficient.
-
Maybe with some sort of fancy hardware-driven round-robin sleep for all cores contending for memory?
-
One of the classic "many-smaller-cores" problems as been threading. Cavium has 8 SMT threads per core, so that when one thread waits on memory, other threads can continue processing. Centriq has no threading, so cores stop completely.
-
While Cavium's approach is superior in theory, many have long doubted it and suspected the approach Centriq is now using might be better.
-
Partly it's because many-slower cores fail when they are too slow, and partly because when memory bandwidth is constrained, you want the cores to sleep rather than trying other work.
End of conversation
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.