single threaded performance.

[robinlawrie]

trying to squeeze the last bit of juice from my 3930k on a rampage gene.

this isnt for beating benchmarks, its for my day-to-day 3d workstation. i need fast rendering, reliablility, and the best possible single threaded perfromance too, as an embarassing amount of 3dsmax is single threaded.

wanted to try the per-core multipliers, so i can get a healthy speedup for single threaded stuff.

thought id test what temps it currently runs at with IBT on only 1 thread. (its oc'ed to 4.4 throughout with an "all-cores" 44 multi at the mo) so i could decide how far to push the one threaded multiplier.


when loaded up with 1 thread, according to realtemp and cpu-z, the cpu multiplier is way down at 12x, and the cpu is clocking in at 1200 mhz.


this obviously sucks for single threaded apps. seems speedstep is deciding the cpu isnt loaded and is downclocking it.


is this something that setting per-core- multipliers will fix?


next up, if i want to use the vcore offset, does it scale voltage with load, or frequency? obviously if i want a single core to go up to say 4.8, ill need more voltage than im happy with to run all cores at 4.4 (trying to keep things cool, currently at 1.28v)

so im hoping it scales with multiplier/frequency, so i can keep my 1.28v at 4.4 (all cores) and have the voltage jacking up only when 1 core is overclocking further.

any advice greatly appreciated.

 

[pgi947]

Kill speedstep and ht may help.

 

[robinlawrie]

hmm.. well HT gives me an extra 30% rendering speedup ( more than my oc'ing)

and speedstep stops my pc from cranking the watts when its idle, and without it i dont think i can even do the per-core overclocking im asking about.


is this normal behaviour for speedstep? bit silly if so, as a lot of apps are single threaded and would end up running at 25% clock speed..

 

[pgi947]

Speedstep should fire the clock speed back up when load is detected, I think some of the c states are linked to speedstep so may be worth a look at those settings.

I've never really played with the per core multiplier.

 

[robinlawrie]

well, turns out it was because i had one multiplier set for all cores.

i dont know if this is a monitoring error, or if it is actually limiting the clocks on low -threaded apps... if so its a major issue for all-core multiplier oc'ing, on the rampage at least.


once i switched to a per-core multiplier it worked as it was supposed to.

i now have a machine that sits happily at 4.4, but when using single threaded stuff, cranks up to 4.8. temps never get above 65 degrees on a corsair h80.

might see if i can push it further. im surprised more people dont use this.. since apart from rendering and benchmarking, most things are not using all six (or even 4) cores..

 

[Nuclear Fusion]

If you use offset, it will scale with frequency/load (not sure which one, but doesn't matter since frequency scales with load anyway). I use it, to keep temps down under low loads. I assume you're familiar with the way a positive/negative offset works?

 

[robinlawrie]

ive been playing with offset, but it doesnt help in this case, since its tied to load, and with per-core multipliers, you have a situation where you have a higher clock (and hence need more voltage) on a lightly loaded system (only 1-2 cores maxed out)

im finding i cant actually get a single core stable much higher than all the cores together, since the cpu sees less than 100% load, and drops the voltage.

if it was linked to the frequency of the highest clocked core, id be very happy.. but its not.

im having to go back to a fixed voltage, meaning nice toasty idle temps, and still limited single core speedup, since with a high enough voltage for a high single core clock, when i run all the cores, it gets too hot for my liking.

really annoying since with enough voltage, i can get a single core up to 5 ghz, and still keep the cpu temp below 60 degrees. this would be fantastic for my work.