Is wPrime a unreliable benchmark?

I was looking over mobile benchmarks, and while super pi, 3d mark, and whetstone all seem consistent in ranking CPU orders, wPrime is all over the place. in wPrime, the i7 cpus 720qm stomps the 820 qm. 940xm times are almost double the 920xm. Why would the clearly superior processor have a much slower time in wPrime?

c2d p9700 is much slower according to wprime than c2d p8700. Just lots of seemingly random ordering of cpus if you go according to wprime.

They are faster in their proper order according to super pi, whetstone, 3dmark, etc. Why is this?

 

wPrime is highly dependent on the system, in particular resource-wise as any other processes using the CPU will slow down the test significantly whereas other tests don't use a full multithreaded CPU, they can afford some other processes sharing the CPU.

 

synthetic benchmarks of all kinds are 'unreliable' in that they are good only for comparing other machines running the same benchmark.

 

it's other machines running wprime, but wprime shows a bunch of such anomalies according to notebook check. wprime runs faster on an i3 330m than a i7 820qm?

 

I think that's more of a case of Notebookcheck being unreliable. It's a good resource for getting very general ideas of performance, but I don't put a ton of faith in it.

If there's one major issue with wPrime, it's that performance fluctuates greatly between versions. On my U30Jc, the difference between 1.55, 1.63, 2.00, and 2.03 varied by as much as 3 full seconds (which is substantial difference when you're talking about 21-25 second scores). It doesn't help that the version NBR standardized on a long time ago, 1.58, isn't available for download anywhere anymore.

When comparing inside one version, it's a great comparison of relative CPU performance. It does seem to favor AMD's chips somewhat, but then again SuperPI favored Intel's architectures pretty heavily, so some battles just can't be won.

 

But notebookchecks numbers are all consistent with logic when they are using super pi, whetstone, drystone, etc. I remember when wprime came out and a lot of people were reporting that their t7500 c2ds weren't running as fast as other other peoples t5500. Running 4 threads on a dual core processor would speed things up 5 seconds or more.

 

On a perfectly idle OS, 4 threads would run slower than 2 threads.

Each thread is given a set task, if there are background processes, one thread will slow down. The more threads you run, the less of a factor the background processes take, but the higher overhead of managing the threads, so on a perfectly optimized system, the ideal number of threads is the number of cores, but this goes up on unoptimized systems.

The simple fact is, being a multithreaded application that uses 100% of your CPU, it is very sensitive to platform optimization. Take two identical systems (other than CPU), and the faster CPU will perform better.

 

WPrime has issues. Not that it can't deliver. I used to do things you said and got 24.xx on a T9400. I have a i7 now wPrime is behind the times. I get a 17.xx. Thread setting trick does not work.

I set 1 core more than one core works. wPrime needs to look at this.

 

You need to set 8 on an i7 800 series because it has HT, 12 might be faster because the HT threads don't perform as well as the other threads.

 

wprime has always been consistent in my tests, I think maybe you were viewing testing that was not done correctly.

Also there is more than one version of wprime, the different versions do not give the same scores.