The New 3DMark Benchmark - Testing Smartphones to Multi-GPU Gaming PCs

[YBS1]

There...Wanted 13K, got 13K

http://www.3dmark.com/fs/126367

 

[hyrule4927]

EDIT 3: At clocks of 1035/1250 i get a graphics score of 6940

Hmm, seems like my score is a little bit on the low side then? I am still running Catalyst 12.10 (every more recent version slows down Folding@Home), so that might be a factor. I'll bench again later to see if my scores are consistent.

 

[Termie]

So I just completed a round of 15 different runs on my HD7870 to determine overclock scaling for core and memory separately. I'll post a chart soon. Long story short, I discovered a serious anomaly in memory scaling between 1250 and 1350 where it actually went negative, only to pick up again at 1400. I'd encourage folks using 3dMark to fine tune their overclock to test core and memory separately. The problem had previously been masked by bumping both core and memory up equally with each test.

 

[3DVagabond]

So I just completed a round of 15 different runs on my HD7870 to determine overclock scaling for core and memory separately. I'll post a chart soon. Long story short, I discovered a serious anomaly in memory scaling between 1250 and 1350 where it actually went negative, only to pick up again at 1400. I'd encourage folks using 3dMark to fine tune their overclock to test core and memory separately. The problem had previously been masked by bumping both core and memory up equally with each test.

You should run the same 1250-1350 mem speed benches on say 3D Mark, and see if it does the same thing.

 

[Termie]

You should run the same 1250-1350 mem speed benches on say 3D Mark, and see if it does the same thing.

You mean something other than 3dMark, like Heaven or 3dMark11? I can go back and do that later today.

 

[Rikard]

I'm wondering if something might be inflating the scores - perhaps Virtu MVP? User Rhoxed scores 7399 with nearly the same clocks, and user Hyrule was slightly below that with 1035/1250 clocks.

Not saying you meant to do this - I'm just trying to figure out if 3dMark might be inconsistent or affected by other software. Your scaling is exactly what I expect though - about 3% for every 5% core, and 1% for every 3% memory. Hence the reason Rhoxed is about 15% faster than hyrule.

Everything in my setup is quite vanilla, though I am OCing the CPU by ASUS tools, so it does not pick up the actual frequency during system scanning:
http://www.3dmark.com/fs/102667
I really have no clue as to why the scores are what they are.

 

[Termie]

Here's that data I promised you all - I've gone through a great number of tests on my HD7870 to test the overclock but also scaling. I find the results pretty surprising:

Two big questions come to mind:
(1) Is error correction causing the memory scaling to turn negative in both 3dMark Fire Strike and 3dMark11?
(2) If so, why does positive scaling appear again as the memory clock gets higher? Is error correction taking place but cloaked by the increase in memory clock?

And two big conclusions:
(1) Memory overclocks do very little on the HD7870, and frankly, most professional review sites simply provide the highest "stable" overclock in their testing without determining whether it actually provides positive scaling.
(2) Taking my maximum tested core clock (1150, or 15% over reference) and maximum memory clock (1450, or 21% over reference), I get a total increase of 13.7% in performance.

 

[spinejam]

Gigabyte 7970 WF3 @ 1050 core 1600 memory:

http://www.3dmark.com/3dm/184934




Uploaded with ImageShack.us

 

[james1701]

http://www.3dmark.com/fs/269264

My latest, with my processor at a lower clock and my 7970ghz Crossfire.

But my best is this one.

http://www.3dmark.com/fs/203607

I do have the CPU clocked up more though.

 

[Durvelle27]

Maxed on my Samsung Galaxy S II

 

[Grooveriding]

How are people getting SLI to work in Firestrike ? SLI doesn't work for me and I've tried numerous drivers.

 

[Termie]

How are people getting SLI to work in Firestrike ? SLI doesn't work for me and I've tried numerous drivers.

Crossfire worked very poorly for me when I tried it yesterday with dual HD7870s. The Graphics score scaled at 38%, the Combined score scaled negatively. Look carefully at the scores above for the HD7970s - everyone's combined score is the same, regardless of whether it's one or two cards.

And no, the Combined Score isn't that CPU-bottlenecked. It scaled up nicely with a single card's overclock.

 

[Grooveriding]

Crossfire worked very poorly for me when I tried it yesterday with dual HD7870s. The Graphics score scaled at 38%, the Combined score scaled negatively. Look carefully at the scores above for the HD7970s - everyone's combined score is the same, regardless of whether it's one or two cards.

And no, the Combined Score isn't that CPU-bottlenecked. It scaled up nicely with a single card's overclock.

SLI literally does not work at all though. One card or two is the same score.

 

[alexruiz]

I am suspecting patches are on the way
Scores are all over the place.

 

[james1701]

There is a blurb in one of the TimetoLiveCustoms video where Tom Logan says 3dMark knows there is an SLI issue and is working on it.

I do get both gpu's going, but only parts of it I get 99% usage though.

 

[AdamK47]

Here are my results with 3x Titans.

http://www.3dmark.com/fs/248605

Yeah, I'd say SLI is a little broken.

 

[SPBHM]

any news regarding the Android/iOS/WinRT version?

 

[Grooveriding]

Looks like it is related to a recent Windows patch, KB2670838, that enables DX11.1 features. Apparently 3DMark uses some DX11.1 features if available and this is why SLI is not working.

http://community.futuremark.com/for...score-3DMark&p=1752972&viewfull=1#post1752972

Going to try removing that Windows patch.

 

[Grooveriding]

Removing that update works

 

[Mir96TA]

 

[Durvelle27]

 

[Mir96TA]

 

[Face2Face]

 

[YBS1]

Not quite 20,000 but I might get there with a little more effort.

The 680's came a long way in the seven months since my first post in this thread.

 

[AdamK47]

Results from my every day stable overclocks. No suicide runs here.

http://www.3dmark.com/3dm/1259234