Jump to content
HWBOT Community Forums

3DMark11 and AMD VGA multi-gpu


Massman

Recommended Posts

Just a heads up for everyone who's playing AMD VGA multi-GPU and doesn't have good results.

 

1) Play with the tesselation level in the Catalyst driver

2) Disable System Info scanning under help tab.

3) Re-run, re-run, re-run.

 

The first option is something everyone probably already knows about. Just for reference, you might want to read this thread: http://hwbot.org/forum/showthread.php?t=17844.

 

The second one is something that hasn't been discussed much before; only saw a thread about it on the Futuremark forums (click). Basically, with AMD crossfire multi-GPU the combined test will get a serious hit if you enable System Scan. On this SB-E setup, the score went down from ~11K to ~7K, which is a pretty big difference. By disabling system info, you won't be able to get your scores verified in the ORB Hall of Fame ... but adjusting Tesselation prevented you from doing that anyways, so who cares :D.

 

The third remark is something I noticed after disabling System Info. It seems that GT1 is especially 'buggy' on this 4x 6970 setup fluctuating between 95FPS and 103FPS. You should be able to quickly see if it's a good or a bad run, though ... so shouldn't be too difficult.

 

Underneath, two screenshots displaying the effects 2) and 3).

 

p18307.png

 

p19996.png

 

You can see that I gained almost 3k points in Combined by disabling system info, but I lost ~6FPS in GT1 (even with increased frequencies). If you re-run enough, you will get a normal GT1 again:

 

p20267.png

 

Another example of good vs bad GT1 run: 20418 with 95.83FPS, 20634 with 103.68FPS.

 

p20418.png

 

20634.png

 

Just a heads up ...

Link to comment
Share on other sites

I hate this benchmark. My pain is GT4. In most cases I get really poor results with 2xHD5970 in this particular test. With my old Core i7 960 re-runing enough times eventually brings me a normal score but now with my Xeon W3670 I have no success. Combined score is another pain but playing with System Info check box and re-runing enough times bring decent score. I will try playing with the tesselation level.

Edited by OLDcomer
Link to comment
Share on other sites

Test setup:

 

- i7 3930k @ 4500

- MSI 6970 Lighting x4

- Rampage 4 extreme

- 4x 2GB GTX8

- all air

 

Okay, so wtf is going on ... seems like you get screwed over when running AMD, LOL!

 

amd1.jpg

 

Yellow indicates the score you get when running SI enabled. For each configuration, SI detects the amount of GPUs correctly.

 

Fyi, it doesn't matter how much GPUs I select in 3DMark11 with crossfire enabled, I always get ~40% load on all cards in combined. So, it's not like you get the bad scores because some GPUs aren't under load.

 

load.jpg

Link to comment
Share on other sites

  • 2 weeks later...

Be advised that any run with SystemInfo disabled produces incorrect results from Combined Test with multi-GPU setups. The benchmark needs to know the number of GPUs for that test.

 

The issues you are seeing with 3+ Radeon GPUs in this test are due to AMD driver issues. We've told them, they have so far disregarded the issue. Sorry, out of our hands.

 

The result from SI-disabled, while the score seems great, it is meaningless. The benchmark doesn't do the same workload that it would do with SI enabled (or with SI enabled SLI run).

Edited by FM_Jarnis
Link to comment
Share on other sites

Disabled Tessellation invalidates the score on any configuration. You are effectively telling the driver to skip work that the benchmark requests it to do. You can do so, but the score is meaningless. Naturally up to you under what "rules" you benchmark. I'm talking about FM hall of fame and the result being flagged in 3dmark.com result browser.

 

Just disabling SI affects the score only on multi-GPU configurations (and only with Combined Test).

Link to comment
Share on other sites

If disabling SI makes the result meaningless because of policy and enabling SI makes the result meaningless because of driver issue ...

 

... how can I obtain a meaningful 4xCF 3DMark11 result?

 

Ask AMD? Currently FM has no solution for this problem and the issue is within the video drivers. We have told AMD this (several times) and can only wait for them to correct the issue.

 

All Single GPU configurations work

2 GPU Crossfire works

SLI with 2, 3 and 4 GPUs works

Link to comment
Share on other sites

Physics? Or Combined? That number-of-GPUs-switch / SI enabled should not affect any other test except Combined Test. In combined test if the switch is set to something other than the actual GPU count (or auto, as per what SI detects), yes the score changes because the test workload is different. GPU count determines how many copies of the GPU physics simulation is spawned. Normally there should be a copy of the simulation running on each individual GPU (to eliminate cross-GPU data traffic on physics simulation state). The whole "GPU physics using DirectCompute" is fairly complicated issue on multi-GPU setups.

 

In any case, AMD just released new drivers and we're going to investigate in-house if anything has changed with either of them (11.12 or the 12.1 preview drivers) and will bug AMD again if there are still issues. I've also suggested internally that FM should make a full ("lots of big technical words") disclosure on this specific issue - we'll see if it can be made happen. Could explain exactly what is going on.

Link to comment
Share on other sites

@FM_Jarnis:

And what about GT4 and my problem? Check the GPU usage in the first and second image which I provided above. It is very annoying.

 

Sorry, no idea. It would appear that for some reason the GPU isn't fully stressed on GT4 on some runs. As every run asks the driver to do exact same things, I can only summarize that there is something funky going on with the drivers.

 

We have not seen this kind of effect in-house and considering the thousands of runs the benchmark went through during pre-launch verification testing, this kind of massive score disparity randomly on some runs would have stood out like a big flashing beacon. So I'm fairly certain that this is not a 3DMark 11 bug, but an issue with the rest of the system doing what 3DMark 11 asks it to do during the benchmark run.

Link to comment
Share on other sites

  • 1 month later...
The current decision is to not approve the existing 7900 drivers during this contest, so it now looks like AMD Radeon HD 7900 series cards will not be able to enter this contest.

 

~ http://futuremark.yougamers.com/forum/showthread.php?162091-ASRock-Overclock-King-Global-OC-competition&p=1656330&viewfull=1#post1656330

 

Heh, seems like the Radeon cards are not supposed to be approved :P

 

Perhaps in April.

Link to comment
Share on other sites

Heh, seems like the Radeon cards are not supposed to be approved :P

 

Perhaps in April.

 

No, this is purely due to no combined AMD driver available. We do not approve beta drivers. 12.1 will be approved this week. It doesn't support 7xxx series, so...

 

Edit: Also note that GTX 560Ti experienced exact same issue.

Edited by FM_Jarnis
Link to comment
Share on other sites

  • 5 months later...

Join the conversation

You can post now and register later. If you have an account, sign in now to post with your account.

Guest
Reply to this topic...

×   Pasted as rich text.   Paste as plain text instead

  Only 75 emoji are allowed.

×   Your link has been automatically embedded.   Display as a link instead

×   Your previous content has been restored.   Clear editor

×   You cannot paste images directly. Upload or insert images from URL.

×
×
  • Create New...