Request: 980ti SLI Benchmarks

[Dirk Diggler]

TLDR:
Early indications are that bandwidth does make a difference in applications that have good SLI scaling, and can be as much as 10% which equates to a decent overclock to your GPUs

We've been looking at numbers in the 980ti owners thread as one of the guys is considering SLI. I ran some benchmarks for him, but we're now discussing the difference 6 cores, 40 lanes and resultant PCI-E bandwidth makes. I've also seen this discussed elsewhere when people have been trying to decide which hardware to buy. Some say it makes a difference, some say it doesn't.

Could we settle this in this thread?

It would require someone with 980ti SLI driven by a 4 core i7 and a mobo supporting only x8/x8 bandwidth. Bench the following games @ 2560*1440 with your cards running 1418/3500.
- Added : As long as it's a pair of 980ti then it doesn't matter what the CPU/Bandwidth - it's all relevant.

ROTTR
Far Cry Primal
Dragon Age Inquisition
Metro Last Light Redux
Firestrike
Firestrike Extreme
Firestrike Ultra
Heaven

Is there any way I could force my cards to x8 bandwidth? - PCI-E Gen 2 in BIOS

(Remember, GPU's must be 980ti @ 1418/Stock)
Resolution should be 2560*1440 where it's optional.

ROTTR Settings:

Spoiler

Heaven Settings:

Spoiler

Valley Settings:

Spoiler

Far Cry Primal Settings:

Spoiler

Dragon Age Inquisition Settings:

Spoiler

Metro Last Light Redux Settings:

Spoiler

Current 980ti SLI (1418/stock) Leaderboard :

ROTTR

1. String (5930k @ 4.4GHz x16/x16) - 100.25
2. MrKambo (4790K @ 4.75GHz x8/x8) - 91.36
3. String (5930k @ 4.4GHz x8/x8) - 90.48

Far Cry Primal

1. String (5930k @ 4.4GHz x16/x16) - 68
2. String (5930k @ 4.4GHz x8/x8) - 68

Dragon Age Inquisition

1. String (5930k @ 4.4GHz x16/x16) - 112.7
2. String (5930k @ 4.4GHz x8/x8) - 105.9

Metro Last Light Redux

1. String (5930k @ 4.4GHz x16/x16) - 82.18
2. String (5930k @ 4.4GHz x8/x8) - 78.77

Firestrike

1. String (5930k @ 4.4GHz x16/x16) - 24248
2. MrKambo (4790K @ 4.75GHz x8/x8) - 22931
3. String (5930k @ 4.4GHz x8/x8) - 22632

Firestrike Extreme

1. String (5930k @ 4.4GHz x16/x16) - 15106
2. MrKambo (4790K @ 4.75GHz x8/x8) - 14861
3. String (5930k @ 4.4GHz x8/x8) - 14082
4. Awahwah (4690k @ 4.5 x8/x8) - 13487

Firestrike Ultra

1. Gabbatek (5820k @ 4.2GHz x8/x8) - 8469
2. MrKambo (4790K @ 4.75GHz x8/x8) - 8376
3. String (5930k @ 4.4GHz x16/x16) - 8347
4. String (5930k @ 4.4GHz x8/x8) - 8065

Heaven

1. String (5930k @ 4.4GHz x16/x16) - 2730
2. String (5930k @ 4.4GHz x8/x8) - 2763

Valley

1. String (5930k @ 4.4GHz x16/x16) - 4164
2. String (5930k @ 4.4GHz x8/x8) - 3861

 

[Dirk Diggler]

** Reserved for future single card leaderboard if required for CPU comparisons **

 

[Dirk Diggler]

If you can limit PCIE to gen2 in BIOS would that give a similar effect ?

I'll have a look.

 

[Dirk Diggler]

If you can limit PCIE to gen2 in BIOS would that give a similar effect ?

I tried Gen 1 and Gen 2 but still getting x16.

This is my result for ROTTR

4790K @ 4.75GHz
980Ti SLI @ 1418/3500

My minimums suck real bad for some reason, might have to investigate that!

Spoiler

Interesting. Over 10% FPS difference and as you say a big difference in minimums. The minimums is something I think a lot of benchmarks and reviews don't take into consideration.

Could you run the three Firestrike benches at those settings (Normal, Extreme and Ultra)?

 

[Dirk Diggler]

Perfect. I need to head out for a couple of hours but I'll get those benches done at the 1418 clock speed later.

Cheers.

 

[Dirk Diggler]

I still have my 980ti's lying around, if im not busy tommorow afternoon i could help you out

Cheers, the more data the better. I'm just doing my Firestrike benches now...

 

[Dirk Diggler]

Firestrike benchmarks :

Spoiler

Heaven Benchmark :

Spoiler

I'll update the OP with what we have so far.

 

[Dirk Diggler]

Leaderboard added, could do with some more data.

 

[Dirk Diggler]

I can't play benching till the weekend now, but ill spent the afternoon on Saturday going through all my bench runs.

No worries, probably best sticking to actual benchmarks with a score (see OP).

 

[Dirk Diggler]

I've got a few games with built in benchmarks, i feel thats more of a accurate representation.

But ill stick to scored benchmarks if that what you need.

No, games with benchmarks built in is what I meant. Sorry, my bad.

 

[Dirk Diggler]

Gen 2 x16 offers identical performance to Gen 3 x8

I just redid the ROTTR bench on Gen 2 and it's dropped 10%.

 

[Dirk Diggler]

And, Firestrike scores for x8/x8. There's a clear difference from x16. It appears to make a bigger difference than 6 cores over 4.

 

[Dirk Diggler]

Far Cry Primal x8/x8






Far Cry Primal x16/x16






This appears to go against the trend, until you realise SLI scaling is poor therefore the bandwidth is moot.

 

[Dirk Diggler]

Dragon Age Inquisition (x16/x16)

 

[Dirk Diggler]

Metro Last Light Redux (x16/x16)

Spoiler

 

[Dirk Diggler]

Valley (x16/x16)

 

[Dirk Diggler]

x8/x8 benchmarks. I'm done for now as there's a good mix of games and synthetics.

Early indications are that bandwidth does make a difference, as much as 10% which equates to a decent overclock.

Spoiler

DA:I






Metro LL Redux






Valley






Heaven

 

[Dirk Diggler]

I'm just curious, why are you putting this effort in it when there are dozens reviews online?

I've seen it disputed so often, I wanted to see for myself with equivalent clocked cards (which is crucial).

Here's an example, the first result in Google...

http://www.tomshardware.co.uk/answers/id-2054999/whats-sli-pci-x16-x16.html

You are correct that two x16 lanes will give the best performance. However, all multi GPU will only ever have the need for dual x8 lanes, for now. The performance difference between two x16 lanes and two x8 lanes is very minimal, at best. Something like 1600 CL9 RAM vs 1866 CL9 RAM.

A motherboard running at x8/x8 is considered optimal without breaking the bank for a minimal performance increase.

Now that I have seen the difference can actually be equivalent to a decent overclock on the GPU, I'm starting to wonder at the point of a 5820k for gaming? What is the point of a 28 Lane extreme chip for gaming? If there's no extra bandwidth to be had, then why not just go for a 6700k?

I'd like to see some 5960x and 5820k benchmarks as well as some i5's. Probably wishful thinking as it's restricted to specific cards, but if people pop their head in and offer a result or two then all the better.

 

[Dirk Diggler]

Just a quick 3d Mark Firestrike Extream run since my second Ti arrived. Both card @ 8x PCIE 3.

Spoiler

Score seems a bit low, mate. Did you have the cards set at 1418 core clock?

 

[Dirk Diggler]

No, no OC going on and I think my machine could do with a re-build.

No worries. Both cards need to be running at 1418 or the data isn't any use.

Thanks anyway, Bud.