Jump to content
HWBOT Community Forums

Recommended Posts

Posted

We tried to keep the actual test parts as short as possible while still getting reliable numbers. We know some run them over and over again (heck, we did in our QA testing) and after certain point you just want them to do their thing quickly and give numbers :)

Posted

While testing the new 3DMark I had two issus.

 

1st: When running Fraps in background the first game test of each benchmark runs without problems. When it’s finished and the loading screen of 2nd game test should appear an error message is displayed and the benchmark is ended with 0 points. Closing Fraps solves this issue.

 

2nd: There is an issue with 3-way-SLI in Fire Strike combined test. My daily rig has three GTX 670 4 gb installed and I can run gt1, gt2 and physics-test without problems. I get normal scores in it. But combined test is horribly low (and so is overall score). In best case frames jump up and down from 13 to 7 fps in the first two or three seconds of this test and from then 7 fps is all I get. But most of my runs show about 7 fps all the time. Overall score is about 7.3k then.

 

I tried different things to solve this:

 

- changed PhysX from cpu back to gpu: no change

- forced AFR2 in NVInspector: no change

- limited ram to 16 gb with maxmem: no change

- clocked down cpu a bit: no change

- renamed 3DMark.exe to 3DMark03.exe: no change (because of the other .exe-files I think this won’t help anyway)

- disabled SlI: now I get constant 13 fps in Fire Strike combined

 

Without SLI I get about 6.2k overall in Fire Strike. Only 1k points away from 3-way-SLI means there is something really wrong. I hope this will be solved with updated drivers (I already installed the latest beta) and/or a patch for 3DMark.

Posted (edited)

3- and 4-way SLI/Crossfire problems are entirely down to drivers. Even 2-way sli didn't work right until the very latest betas. Guess have to give some time for AMD and NVIDIA get those things sorted. I believe they have some vested interest too - whichever gets their quad-GPU support working first on 3DMark gets to dominate hall of fame lists everywhere until the other guy catches up :)

 

Also FRAPS incompatibility is probably something that needs to be fixed - not sure if this is FRAPS or 3DMark issue but my hunch would be that FRAPS does something when it hooks to DirectX stack that isn't entirely kosher.

Edited by FM_Jarnis
Posted

Im getting hooked on FireStrike already , by far the best looking 3DMarks of them all , well done futuremark. No problems here to report running flawlesly in 2 way sli and 2 way crossfire as well as APUs.

Posted

No tes is ako with other benches why would they change there stance now?

Posted

Tesselation for the new benchmark is the same problem as for 3DMark11: you need to verify only if Tesselation was used. So either force verification links or only allow screenshots with system connected to the internet. No more just taking a screenshot to compare, the a lot of manual procedure (run, screenshot, save file, upload, copy url, fill in details).

 

Should we open a different thread for this, maybe? I'm split 50/50 on the issue.

 

(the easiest way would of course be to use hwbot benchmark api for submissions :D)

Posted

At this point I would say leave it be

there is no turning back...its driver level.

Maybe soon we can get an nv card that will make it a mute point for awhile?

Posted
Tesselation for the new benchmark is the same problem as for 3DMark11: you need to verify only if Tesselation was used. So either force verification links or only allow screenshots with system connected to the internet. No more just taking a screenshot to compare, the a lot of manual procedure (run, screenshot, save file, upload, copy url, fill in details).

 

Should we open a different thread for this, maybe? I'm split 50/50 on the issue.

 

(the easiest way would of course be to use hwbot benchmark api for submissions :D)

 

Sounds like it's not too late to turn back then. :) A proper discussion thread seems like a good idea.

Posted

My experience with Titan and 3DMark so far: 3-Way-SLI with Geforce GTX Titan results in unusual behavior. With the exact same driver settings (used Geforce 314.09) the framerate drops from 14.8 Fps (2 x GPU) to around 5.5 Fps and the screen "twitches" already when loading the combined test. Also while Game test 1 scales from 2 to 3 GPUs (ca. +25 %), I noticed the framerate in Game test 2 is reduced by about 10 %. The GPUs were not overclocked and the same goes for the Core i7-3960X used fot the few runs.

 

2 GPUs vs. 3 GPUs (one dedicated for PhysX @ driver) improved the overall score by around 10 %. 3-Way-SLI results in lower scores because of the bugged combined test.

Posted

Not that many people I guess, but it's a great card - no doubt about that. :)

I forgot I was talking about Fire Strike Extreme, so it's likely the same problem with 3-way-SLI and the combine test already described in the thread. That means it's not an isolated Titan problem, but a general driver issue.

Posted

Combined test has always been a problem. The 6970 had the same Combined issue you're describing. As far as I can remember it was indeed because of a driver issue.

 

What happens if you disable SystemInfo and manually select the amount of cards to use for rendering? Just run combined.

Posted (edited)
My experience with Titan and 3DMark so far: 3-Way-SLI with Geforce GTX Titan results in unusual behavior. With the exact same driver settings (used Geforce 314.09) the framerate drops from 14.8 Fps (2 x GPU) to around 5.5 Fps and the screen "twitches" already when loading the combined test. Also while Game test 1 scales from 2 to 3 GPUs (ca. +25 %), I noticed the framerate in Game test 2 is reduced by about 10 %. The GPUs were not overclocked and the same goes for the Core i7-3960X used fot the few runs.

 

2 GPUs vs. 3 GPUs (one dedicated for PhysX @ driver) improved the overall score by around 10 %. 3-Way-SLI results in lower scores because of the bugged combined test.

 

We know there is a 3- and 4-way scaling problem right now and are working with AMD and NVIDIA to resolve this. We hope to have more news next week and this may require a patch to 3DMark.

 

At the moment our recommendation is not to use any scores of Fire Strike from 3/4 GPU systems as things are clearly not working right.

 

Drawbacks of working on the very cutting edge of DirectX and DirectCompute, doing stuff no game really is doing yet... the fluid simulation of the smoke particles on the bridge is somewhat ambitious as far as the code goes and multi-GPU adds complexity considerably :)

Edited by FM_Jarnis
Posted
Combined test has always been a problem. The 6970 had the same Combined issue you're describing. As far as I can remember it was indeed because of a driver issue.

 

What happens if you disable SystemInfo and manually select the amount of cards to use for rendering? Just run combined.

 

If you do manual GPU count select to something other than the actual GPU count in system that is going to participate in rendering, you just make the combined test completely pointless. It is a debug option only for cases where you get somehow bogus results with the autodetect (most commonly, 2-4 GPUs in system but Crossfire not enabled on purpose would require manual setting to 1 to get correct single card score).

 

6xxx series Radeons have a driver issue of some sort with 3DMark 11 combined test - more than 3 cards combined test score suddenly tanks. We've told AMD. I actually reminded AMD person about this just today.

 

This doesn't happen on 7xxx series, by the way (nor on any NVIDIA setups). It is specific to HD 6xxx cards and specifically with 4GPUs.

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...