Radeon HD 6950 vs 6970 and 5870

[Apis]

I find the GPU benchmark results suspicious for the Radeon HD 6950:

6970: 3136, 102.9 %
6950: 3049, 100.0 %
5870: 2749, 90.2 %

The above results are not supported by tests on other sites, take a look at review at Anandtech for example.

The 6950 performs worse than the 5870 in some games and significantly worse (20%) when it comes to GPGPU performance. (which is what one expects from the specs since 6950 have fewer shader units and lower core clock speed.) It's only faster when it comes to very high resolutions and tessellation.

Also the difference between 6950 and 6970 should be more than 3%! (more like 10%)



The only explanation I can think of is that the 6950 cards have been unlocked (eg converted to 6970) and overclocked?

Or is there some other reason passmark favors the 6950?

EDIT:
Also the 3D benchmark says even the 6870 and 6850 is better than the 5870 which is simply impossible.

I read in another tread that the benchmark is somewhat CPU dependent. Could that be the explanation (the 6xxx series benchmarks are from newer CPU/motherboard combos)?

EDIT2:
Wikipedia has a comparison page for the different AMD GPUs.

---
This is now a problem for me because the store I bought my 5870 card has replaced it (without asking me first) with a 6950 (I had it serviced because of a broken fan). I primarily used my 5870 for openCl (GPGPU) computations so that is a significant downgrade for me.

The store only refers to this page and says it is much better, while I'm trying to explain to them that the results does not reflect the actual performance difference.

[David (PassMark)]

Don't really know what you are talking about.

Our results
========
6970: 3136, 103%
6950: 3049, 100%
5870: 2749, 90%


Anandtech Results (Crysis 1900x1200, 4xAA)
========
6970: 51.5, 106%
6950: 48.4, 100%
5870: 45.9, 95%

The results are in fact in pretty good alignment. It is silly to expect exact alignment to a tenth of a percent to all other benchmarks out there.

I am sure it would be trivial find another benchmark where the alignment was worse, but not all benchmarks and games will run equally well on all hardware. So it always make sense to look at a range of results. In an ideal world you would get benchmarks for you exact hardware, exact O/S, exact driver version, and exact software that you want to use.

If you main usage of your video card is OpenCL, then you should be looking at OpenGL benchmarks, not DirectX9 benchmarks.

Having said that, in the next major release of PerformanceTest, we'll be adding some DirectCompute tests. Which should be similar, but still probably not exactly the same, as OpenCL in performance.

[Apis]

Thanks for your answer!

Yes, if you look at the crysis results 6950 is 5% better, but it is actually 2-4% worse than the 5870 in some other games in the above test.

I'm not saying the 6950 is worse than 5870 when it comes to game performance, just that it's not that much of a difference.

And it doesn't explain why even the 6870 and 6850 is reported as better than the 5870...

Have I understood it correctly that the results are based on the average of many users on different cpu/motherboard/mem/gpumem (etc) combinations? That would favor newer cards a bit so that might explain the above difference?

DirectCompute tests sounds great.

----
EDIT:

Looking at the same crysis example:

Passmark
----------
5870: 2749, 100 %
6870: 2827, 103 %
6850: 2751, 100 %

Anandtech (Crysis 1900x1200, 4xAA)
----------------
5870: 45.9, 100 %
6870: 41.2, 90 %
6850: 36.1, 79 %

[David (PassMark)]

Yes, results are an average. Sometimes the average is over 1000 of machines for common cards, sometimes of less than 10 machines for rare cards.

In normal use you aren't going to notice a 5% difference in video card performance. Plus there is a margin for error in all these tests. Differences of a few percent aren't significant.

[Raoul Duke]

Suggestion: Do not have precision when it is not accurate. Change the display of results to indicate what the give-or-take range is. Etc. (This really is pretty much basic stuff for properly presenting data.)

Thanks for PassMark, I use it all the time. It helps me find deals. So I'm not trying to be a bitchy bitch about the whole enterprise, just trying to show this is a Teachable Moment.