Comparing CPU core productivities

In previous posts you mention the performance difference between 32 nm and 45 nm. You can't use the top-Xeons to compare those differences, they are 32 nm, just like your i5 430. Comparing the runtimes to the Opteron is more valid. But only if you want to compare AMD and Intel. That difference is at least as important as the 32 nm and 45 nm differences.

Edit: I'm talking a about the top 5 here. :)

The basic error here is ignoring the question of whether hyperthreading is on or off.

As the Lynnfield die possesses 4 physical cores (as does the Nehalem), Lynnfield i5 hosts for which BOINC reports 4 processors are NOT running HT, while all i7 hosts reporting 8 processors are (there is not yet an i7 die shipping with eight physical cores).

The Arrandale and Clarkdale i5 dice, on the other hand, have only two physical cores, so if you found a Core i5 Arrandale or Clarkdale reporting four processors it would be running HT.

If you actually wanted to compare the per-core productivity, you would just divide the HT hosts times by two.

As to process generation, the i5 is a split decision--the Lynnfield products run on 45 nm, while the Clarkdale and Annandale are on 32nm. Some of the actual dice are used for both i7 and i5 products. This is true of Arrandale and Lynnfield.

As you have chosen to hide your own hosts, I can't just look at yours, but you can: Just check the "Number of processors" entry on the host detail page.

One other thing: a host running less than a full load of Einstein will rather commonly report shorter run times. Nehalem has a lot of memory bandwidth, but the processes do conflict at both the cache and memory access level.

If you actually believe i5's have a large fundamental architectural advantage over i7's you are mistaken.

If anyone round here understands the inner workings of Intel processors, it's archae86. If you ask nicely, he might tell you something of his CV.

Ok to add some comparible (ish) results to this discussion

i have an i7 930 and an i5 655k(multiplier unlocked) both doing Einstein

i7 http://einsteinathome.org/host/3625774/tasks

i5 http://einsteinathome.org/host/3025270/tasks

i7 is oc to 3.6 and running HT - (mem @ 180) - other two cores feeding gpu's(8no)

i5 is oc to 4.0 and running HT - (mem @ 133) - other two cores feeding gpu's (6no)

i7 is doing 6 cpu wu @ once

i5 is doing 2 cpu wu @ once

ignore the last few results from the i5 as had overclock turned down to 3.2gig

turbo mode is off on both cpu's

i5 being at a slightly higher speed is quicker than the i7 by about a ~1000 sec - consistent over several days crunching - get same times if i use less cores say 2 on i7 times are still the same

at stock for both i7 at 2.8gig and i5 at 3.2gig - i5 is about an hour or more quicker per wu - this is from memory as results are long gone - i7 does a cpu wu in about ~22k seconds at stock and 16.5k seconds when over clocked

from limited (non scientific) testing cpu wu are more dependant on cpu clock speed than memory speed - i.e. FLOPS bound

if i put the i5 at approx 3.8 then i would get almost the same times as i7

basically at the same clock speed the i7 is slightly faster but not by much per single thread

(have not tried turning off HT on either yet)

Orgil,

archae86, me and probably quite a few others understand perfectly well what you want to say. There's no need to repeat any of this. However, we've been pointing out a fundamental flaw in your interpretation of your data, which may totally invalidate the point you're trying to make. We need to get past that point.

In your initial post you said "both cpu's processing 2 wu's in 2 threads of a core". I tried to make you clarify what this means, as it's not unambiguous. So far you are assuming that you are doing the same work as the CPUs you are comparing to, just with 2 cores instead of 4 cores.

May I suggest you're wrong? What I think you're doing is running "1 WU on each CPU core". HT may be active for your CPU, but you're not using it. The OS is smart enough to schedule tasks on the fastest ressources. That means you're finishing 2 WUs in 22 - 23 ks. If you used HT you could run 4 WUs in parallel and achieve about 43% higher throughput for GC tasks (as archae86 measured in his setup).

The last results you've shown here show nicely how your WUs always finish in pairs, not quads.

Archae86s Westmere needed 14899 s for a GC WU at 3.4 GHz. At 2.27 GHz I'd expect about 14899 * 3.4 / 2.27 = 22300 s on average for these WUs and this CPU architecture. Notice how this is just about what you're getting? That's because the architectures perform about the same at Einstein.

If you compare to runtimes of 24 ks for i7 CPUs then these are using HT. That's a huge difference: these Quads are producing 8 WUs in almost the same time as you're doing 2 WUs with your dual core. Per core productivity is almost a factor of 2 better thanks to higher clock speeds (no offense - lower clock speeds are healthy for laptops) and the use of HT.

Running 4 GC WUs you should be able to deliver 4 of them every ~31 ks.

MrS