Crossfired 5850's Issue

[Mussels]

I do not see a single thread maxing out a core though. I wish it was so simple...but it's not. I think AMD would really have no excuse if that was truly the case.


And that last bit..OC...really, should NOT be nessecary. I was fine with my qx9650@ 3.6ghz, and was getting 100% load on both of my 4890's, but again, AMD's driver has DRASTICALLY changed since those times. It's that change that has lead to the super-small gains to be had with stock cpus.

I have good idea as to what's going on, but I'm to willing to go into it until I'm 100% for sure. At least I now have some direction though...I'll be posting ALOT more about this exact issue in the days to come.

it jumps between cores, if you're looking for a static one in task manager.

as was mentioned earlier, it could be something related like memory bandwidth, CPU-NB link, etc.

The games themselves could be to blame here as well, for poor multithreading.

 

[cadaveca]

Actually, I watch threads in performance manager, not taskmanager. That gives a slightly different perspective, as then you can view handles for each thread, and the cpu time consumed...looking at taskmanager's graph really doesn't tell much...there's no way to indentify what's consuming the cpu time...is it hte app, the driver, or something else? Again, task manager is useless for this sort of thing.


you do bring up a good point about it being the game engine itself, and that's why I'll be testing other apps, as I did with testing Company of Heroes and World in Conflict today. But I see the same in BFBC2(I mentioned before that CFX scaling was broken here), as well as every other app, when the cpu is stock. There's alot more depth to that...but all in due course.

I plan to basically do all the leg work here for AMD, and maybe they can take the info and find proper fixes. I have ALOT of games to go through though ,so it's going to take some time. I'm gonna hvae to indentify the exact bottleneck for each game.

 

[Mussels]

Actually, I watch threads in performance manager, not taskmanager. That gives a slightly different perspective, as then you can view handles for each thread, and the cpu time consumed...looking at taskmanager's graph really doesn't tell much.


you do bring up a good ponit about it being the game engine itself, and that's why I'll be testing other apps, as I did with testing Company of Heroes and World in Conflict today. But I see the same in BFBC2(I mentioned before that CFX scaling was broken here), as well as every other app, when the cpu is stock. There's alot more depth to that...but all in due course.

I plan to basically do all the leg work here for AMD, and maybe they can take the info and find proper fixes. I have ALOT of games to go through though ,so it's going to take some time. I'm gonna hvae to indentify the exact bottleneck for each game.

one thing to test: try cranking AA up, and see how that goes. just because you're CPU limited, doesnt mean you cant do GPU only tasks to get some more candy at the same FPS

 

[cadaveca]

one thing to test: try cranking AA up, and see how that goes. just because you're CPU limited, doesnt mean you cant do GPU only tasks to get some more candy at the same FPS

Yes, and I was actually gonna edit about that...it's the driver not feeding the cards enough data. AA has basically become free...with stock cpu the cards only load @ 50% or less, clear indicator that yes, detail settings and such can be upped...but from 4xAA, there's not much that is actually going to have a noticible impact on the visuals...I'm already cranking all the details to the max.

Now, I know a very easy way to test whether it's the engine or not...the same limits should be there whether singlecard or not. If the app behaves fine with a singlecard, then I feel safe to say that the additional driver load managing the cards is at fault, and that's why earlier I mentioned how "stable" the singlecard results are in comparison to the CFX results...sure sign it's the additional driver load, and not the game, causing the problem.


Adding a second card doesn't make the game have to do more work, at all. That bit is basically unimportant to the app...unless the app specifcially has switches within the engine to change data flow for multiple cards(as some engines do), but even those that do, don't exactly do what you'd expect...


Like I said, very interesting topic...there's alot of ground to cover here, that will keep me busy for some time. At least it's something to do with this stupid cards...playing games, quite obviously by the example in this thread, isn't always a great experience for CFX users.


EDIT: What really strikes me as funny, Mussels, is that you mentioned in that other thread that maybe I should check my overclock for stability...but here, it seems the issue comes from NOT having an OC, and has NOTHING to do with stability...the exact OPPOSITE problem.

 

[Mussels]

i've had an (unverified) theory for a while, that adding a second GPU adds to the CPU load (as you called it, driver load) thus causing CPU bottlenecks to happen quicker than in single GPU setups (and potentially lowering performance below single card levels, when insufficient CPU power is present)


your best bet really is to OC your CPU, and if that cant be done, figure out some other method to get best benefit (underclocking/volting the GPU's for power saving, using more AA/AF, etc)

 

[cadaveca]

i've had an (unverified) theory for a while, that adding a second GPU adds to the CPU load (as you called it, driver load) thus causing CPU bottlenecks to happen quicker than in single GPU setups (and potentially lowering performance below single card levels, when insufficient CPU power is present)


your best bet really is to OC your CPU, and if that cant be done, figure out some other method to get best benefit (underclocking/volting the GPU's for power saving, using more AA/AF, etc)

Yeah, I agreee 100%(and have posted that exact same thing many times). Funny thing is though...we really shouldn't need to overclock, at all. The fact that these companies cannot release products that work well without an overclock says ALOT about the industry...yet another way that OC'ing, in my perspective, is ruining the industry.

AMD needs to put out some disclaimers..."PROPER CFX Performance can only be had with and OVERCLOCKED cpu, which will void your warranty."



...and clearly, reviewers doing vga reviews with OC'd systems now makes ALOT of sense...it covers the faults in the tech. I wondered why some sites that previously waved the "I TEST AT STOCK" flag all of a sudden all started reviewing @ 3.8ghz...seems I got my answer.

 

[Mussels]

Yeah, I agreee 100%(and have posted that exact same thing many times). Funny thing is though...we really shouldn't need to overclock, at all. The fact that these companies cannot release products that work well without an overclock says ALOT about the industry...yet another way that OC'ing, in my perspective, is ruining the industry.

AMD needs to put out some disclaimers..."PROPER CFX Performance can only be had with and OVERCLOCKED cpu, which will void your warranty."



...and clearly, reviewers doing vga reviews with OC'd systems now makes ALOT of sense...it covers the faults in the tech. I wondered why some sites that previously waved the "I TEST AT STOCK" flag all of a sudden all started reviewing @ 3.8ghz...seems I got my answer.

its more that GPU power is far above CPU power, currently. if you had a 4GHz i7 i bet the story would be a little different.


This is also why i tell people not to get 5970's, let alone the madmen with 2 of them... without a 4.5GHz+ i7 you aint gunna see any gains.

 

[mastrdrver]

Back and yep no difference. 1 fps on average 2 fps on the minimum. The interesting thing though is that they are consistent from run to run when doing 3 loops.

Fwiw, Dave Baumann has said several times on B3D forums that, at least ATI gpus, are cpu limited in games and not gpu limited. So there may be more credence to your claim cadaveca than you think.

 

[cadaveca]

So there may be more credence to your claim cadaveca than you think.

Oh, i know. I'm on B3D as well, just never post. I don't have anything to really offer that crowd.

And I'm not "claiming" anything...the numbers are hard to refute. The difference, now, for me, is that I've learned to keep my mouth shut until I have hard evidence.

I mean really, if I had just said "cpu limitation", and left the thread, most of you would not be looking at it as you are now. I went step by step for a reason...

Scientific method...state the theory, and then either prove it, or disprove it. Ill post the data...you guys can judge whether I'm right or not.

 

[alexsubri]

@Armored Calvary

I'm throwing this out there....what does your 12v say? I had a bad motherboard before any my 12v said 0 even though I had power...Turns out one of my transmitter broke off...I was getting similar results with my 2nd gpu running really low in crossfire

 

[ArmoredCavalry]

@Armored Calvary

I'm throwing this out there....what does your 12v say? I had a bad motherboard before any my 12v said 0 even though I had power...Turns out one of my transmitter broke off...I was getting similar results with my 2nd gpu running really low in crossfire

Everything power-wise seems fine. I think after some tinkering with overlocking the cpu, I have to agree with Cadaveca's conclusion, that the cpu is limiting the scaling of crossfire.

For now, I'm just gonna return the second card, and wait for the 6K series. Performance of one 5850 is pretty darn good as is. It was a fun experiment though, and I'll have to watch this thread to see if there are any other developments. Thanks for the help anyhow.

 

[mastrdrver]

Scientific method...state the theory, and then either prove it, or disprove it. Ill post the data...you guys can judge whether I'm right or not.

The hard part, which seems to be abundantly prevalent these days, is follow the evidence to its logical conclusion and then accepting it.

If ATI is claiming they pulled off those improvements in the [H] review because of the driver's over use of the cpu, that's beyond pitiful. Knowing that now with Dave Bauman's comments about being cpu limited, gets you to wondering if they will ever get it right. :shadedshu

 

[cadaveca]

If ATI is claiming they pulled off those improvements in the [H] review because of the driver's over use of the cpu, that's beyond pitiful. Knowing that now with Dave Bauman's comments about being cpu limited, gets you to wondering if they will ever get it right. :shadedshu

What I was attempting to say within the confines of 140 characters, is that we did indeed add new functionality. Period. The map that now runs faster was running slow with 10.8, 10.7, 10.6, 10.5 and yes even 10.4.... in other words it was always running that way. We found a CPU optimization that basically used four cores effectively (in a four core CPU example).

I did NOT say nothing was broken. What I said is that we fixed one issue by the above stated new functionality. If there was a map that was working great up until Cat 10.5a and then stopped running as fast with Cat 10.6, 10.7 and 10.8 please let me know which map it is and a quick FPS comparison of 10.5a and 10.x

I am here to help... as tough as it may seem to believe

http://hardforum.com/showpost.php?p=1036107963&postcount=78


It's scary for me to read such statements, 100% for sure. I'm not sure whether they are really that ignorant, or if it's because they know about the issue, but lack the skill to fix it...and then deny the problems until they have a fix...

It's really hard for me to remain confident in AMD's ability to address issues, as a consumer, given that examples like this crop up time after time.

The whole "I don't have the problem, so there must be none" attitude I get from them really doesn't help that either.

I do understand, however, that without having the issue, or being able to replicate it, addressing it is very hard.

Really, what I want, at this point, is for AMD to start publically publishing thier review guides. They are covered under NDA at the moment, but this has GOT to change. We'll never get good performance until we can see exactly how AMD is running thier systems for development. Ar they AMD-cpu based, or Intel? Overclocked, or stock? These things are VERY important.

I'd really like reviewers to stop using overclocked systems for reviews too, but I know that that is not going to happen any time soon. And that's very upsetting to me...issues like this one in this thread are never going to be covered in a review until that happens.


What really concerns me with this particular fix is that my own testing shows that this supposed "multi-core fix", really, is still not using cpu resources effectively...at least, not with AMD's own cpus. Maybe once I've compiled all the data, and can show exactly where the bottleneck for Crossfire is, AMD's error here might become quite obvious...at least I hope so...so they can fix it. CPu usage is NOT effective, clearly, given the issue in this thread, using a profile that supposedly features this fix.