Clockspeed

[perkam]

Hey all. Was wondering to what extent WCG and similar crunching progs are affected by clockspeed.

I'm not talking about the difference between a 2.66Ghz i7 920 and a 4Ghz i7 920, which of course will impact performance.

I'm talking about the difference in performance between a DP Nehalem setup @ 2.2Ghz and a DP Nehalem setup @ 3Ghz...or it could be a Harpertown or Clover. Point being, is the 3Ghz Xeon a hell of a lot more faster than the 2.2Ghz with the same number of cores/threads?

Perkam

[karl_eller]

Hey all. Was wondering to what extent WCG and similar crunching progs are affected by clockspeed.

I'm not talking about the difference between a 2.66Ghz i7 920 and a 4Ghz i7 920, which of course will impact performance.

I'm talking about the difference in performance between a DP Nehalem setup @ 2.2Ghz and a DP Nehalem setup @ 3Ghz...or it could be a Harpertown or Clover. Point being, is the 3Ghz Xeon a hell of a lot more faster than the 2.2Ghz with the same number of cores/threads?

Perkam
As far as I'm aware, it scales fairly linearly, so higher clock speeds = better. The problem, however, is that power consumption, vCore and heat output DON'T scale linearly with clock speed, so going from 2.66 to 4 Ghz on an i7 might net you a fairly healthy boost in PPD, you could be looking at double the heat output (just pulling a number out of my ass here, I don't know if it's actually double, but it is a lot more), as well as reducing CPU life and upping your power bills. So OC'ing is worth it, but for 24/7 WCG we usually stick with a fairly "safe" overclock.

The above doesn't really apply to buying a higher binned chip (eg buying a 3.0 GHz Xeon instead of a 2.2 GHz Xeon) since they'll have fairly similar power consumptions and heat output, but you'll find the 3.0 GHz Xeons are easily double the price of the 2.2's :p:

Eller

[jcool]

What Eller said... In theory, an increase in clockspeed nets to 100% in a boost of performance, thus the same CPU would be exactly 36,4% faster going from 2,2 to 3Ghz.
However, this only applies if the FSB/QPI/HTT/whatever bus speed isn't already saturated. Example, running a QX6700 at the same (low) FSB of 266Mhz, you will gain less and less PPD increase if you raise the multiplier further and further, simple because the slow FSB can't keep up with the CPU. This is even more evident in DP rigs, especially Dual Clovers but also Harpers. 445*9 is WAY faster than 400*10 for example.

With single socket 775 and high FSB, or AMD and i7, you should see pretty much linear scaling tho. Gainstown is pretty much twice as fast as a single i7, same goes for Istanbul I suppose.

[perkam]

Ty for the responses gentlemen. Looks like you've hit the nail on the head:

http://www.techreport.com/r.x/opteron-2435/cinebench.gif

Both the Shanghai and the Nehalem Cinebench results show linear scaling with clockspeed, albeit the base performance of each chip increases with an increase in cores or threads, as evidenced by the 36.7% increase in score going from 2.7Ghz Shanghai (2387) to the 2.6Ghz Istanbul (2435).

Perkam

[Snow Crash]

Think we can get Dave to chime in on exactly what he means when he say if set up correctly an i7 920 can get 30,000 PPD? I am close to that number at 4GHz but am always up for whatever tweaking might help.

{edit}I like 21x191 because I get closer to my 1600C8 ram speed but if 19x211 is better because the bus is saturated I can do that instead (just have to get 1866 ram to better support it cause I just can't seem to OC my ram enough). I can't run 20x200 at any reasonable vCore{/edit}