XFX 6950 2GB Memory Bandwidth Issue?

[3DVagabond]

You think AMD would have tested with Intel's latest processors. :\ If there was a bug they would have found it. Stranger things have happened though.

Could it be driver related? If so, someone should repeat the test with release drivers and see if it was present when the cards were released.

 

[Ben1979]

You think AMD would have tested with Intel's latest processors. :\ If there was a bug they would have found it. Stranger things have happened though.

Could it be driver related? If so, someone should repeat the test with release drivers and see if it was present when the cards were released.

See: http://forums.amd.com/devforum/mess...=130923&STARTPAGE=1&FTVAR_FORUMVIEWTMP=Linear

 

[techie81]

Interesting read. I wonder if they are working on this at all?

 

[Ben1979]

with a friend with x58-i7 920 setup we did some tests.. which at least to my eyes seem definitely to be a problem.. even in "real" life benches/games.
The i7 at same clock should trash the Q9650.. but except the very cpu limited 3dmark 06 SM2 test(@1280x1024 resolution..), and a slight advantage in Vantage P(1280 resolution again...) it loses everywhere..

so assuming a +5% advantage the i7 should have in the benchmarks it lost, i'm eager waiting the next super duper Catalyst release "Improves 3D performance up to 7%"

, silently fixing the bug..(or releasing the "brake", depending how you see this "bug"

)

Posted by papatsonis. On http://forums.guru3d.com/showthread.php?p=3886512&posted=1#post3886512.

 

[cusideabelincoln]

The i7 at same clock should trash the Q9650.. but except the very cpu limited 3dmark 06 SM2 test(@1280x1024 resolution..), and a slight advantage in Vantage P(1280 resolution again...) it loses everywhere..

Most of those tests are GPU dependant. The ones that aren't clearly show the i7 spanking the C2Q.

The Q9650 at 4.2 GHz is pretty fast - fast enough to adequately push a 6950, especially in those tests which are mostly GPU bound. Again the CPU bound tests show the i7 ahead by a significant margin.

The small (almost within margin of error) performance penalty seen here could very well be something that Nehalem has been known for all along. See:

http://www.anandtech.com/show/2658/4

So good results, but I don't think they are reflecting the problem you had originally. I think they are showcasing something else. The answer as to why the PCI-E speed test is wonky with your 6950 is still alluding us.

 

[papatsonis]

i think the nehalem cache penalty should kick in, in the exact opposite situation.. in low gpu loads (like 3dmark 06) where it kills q9650.. as we move to gpu limited workloads, should be negated... but it isn't the case. How you explain the fact that while in Vantage 1280x1024 resolution wins the 9650, in 1920x1080 loses..? Additionally the AVP was choosed because of this review where it seems to be texture heavy game, thus more likely to see a bottleneck in pci-e throughput

 

[cusideabelincoln]

i think the nehalem cache penalty should kick in, in the exact opposite situation.. in low gpu loads (like 3dmark 06) where it kills q9650.. as we move to gpu limited workloads, should be negated...

The brute force of the i7 is going to overcome the cache penalty when the GPU load is too low, which is probably what is happening in 3DMark06 and why the i7 is faster.

How you explain the fact that while in Vantage 1280x1024 resolution wins the 9650, in 1920x1080 loses..? Additionally the AVP was choosed because of this review where it seems to be texture heavy game, thus more likely to see a bottleneck in pci-e throughput

An overlooked test so far is the multitexture fillrate of 3DMark06. This one may be impacted by this PCI-E bottleneck.

Perhaps you guys should do a few more tests. Open world games would probably work well.