Geforce GTX 980 Ti vs GTX 1070 vs GTX 1080 Overclocked Performance

[Sweepr]

Overclocker's Club did a close comparison of three of the fastest VGAs right now (all reference launch models), with and without overclock in 10 different games.

Summary of Performance - 1440P Overclocked:

GTX 1070 is 3% faster than GTX 980 Ti
GTX 1080 is 23% faster than GTX 980 Ti

GTX 1070 is 27% faster than R9 Fury
GTX 1080 is 47% faster than R9 Fury

GTX 1070 is 31% faster than GTX 980
GTX 1080 is 51% faster than GTX 980

Comparison Video Cards:

XFX R9 390X DD
MSI GTX 970 Gaming 4G
PNY GTX 980 XLR8
XFX R9 Fury
PowerColor R9 390
NVIDIA GTX 980Ti
NVIDIA GTX TitanX
NVIDIA GTX 1080 Founders Edition

The cooling results again show that NVIDIA has installed a cooling solution that easily handles the thermals on the GTX 1070 Founders Edition. The three heat pipe-based cooler does well enough that the delivered temperatures when the fan speeds are maxed out is only two degrees warmer than MSI's righteous Twin Frozr V cooler, albeit a bit noisier in the sound department. This is a significant achievement with a blower-style cooler. The other upside is that the blower-style cooling solution used on NVIDIA's top tier Pascal architecture cards directs the thermal load outside the chassis, reducing any heat build up in the chassis. You can point to this as a win, but with the way chassis are built for clear directed airflow, the larger coolers have a place in the ecosystem.

Noise levels are quite impressive on the GTX 1070 Founders Edition card using the old ears. Much like the GTX 1080, the GTX 1070 does ramp up fairly quickly when under load with a noise commitment that is much like that of the GTX Titan X and GTX 980 Ti. Cranking up the fan speeds when overclocked does bring a higher noise commitment to the equation, but you get the benefit of significantly improved cooling. Ultimately, 57 °C is not a bad number as lower temperatures equal higher core clock speeds on the GTX 1070.

One of the key targets for NVIDIA over the past few generations has been to reduce the power consumption of its architecture, especially after taking a lot of heat (figurativley) on the Fermi launch for both the power and thermal results. With each successive generation, the power consumption curves keep getting better. Now serving up Pascal to the user base, NVIDIA has delivered solid gains in power efficiency. Looking at a direct comparison between the GTX 970 and GTX 1070, the power consumption reported with the GTX 1070 is the hands down winner. At stock speeds, the margin is just 18 watts under load. When overclocked, the GTX 1070 shows a massive 44 watt improvement over the GTX 970. Pretty impressive, to say the least.

...The best possible core clock speed/memory speed I was able to wring from this GPU was a solid 2050MHz to 2063MHz under load at the maximum reported temperatures. If I could keep the core temperature another 10 °C cooler, the card will run at 2126MHz all day long. Overclocking would not be used if it did not have an added benefit. With stock speeds on the GTX 1070 Founders Edition offering GTX Titan X performance for a fraction of the cost, you can do even better than Titan X performance in many games. Overclocking the GTX Titan X does even up the field, but it's amazing that for a potential AIB card price of $379, you cannot go wrong.

Launching June 10, the GTX 1070 Founders Edition will be available for $449. Much like the GTX 1080, the GTX 1070 has got style, runs cool and quiet, and is tough to beat at the current pricing.

www.overclockersclub.com/reviews/nvidia_geforcegtx_1070_overclocking

Just for my own curiosity, I took the overclocked results at 1080p and 1440p and compared the new cards to the 980Ti.

The green icons represent a greater than 5% improvement, yellow dash is 5% or less change, and the red arrow represents a decrease.

 

[ShintaiDK]

Just keeps showing how good GP104 is. No competition for the rest of 2016 for it

 

[sm625]

23% faster for 41% higher core clocks and 39% higher memory clocks? Something is clearly bottlenecking the 1080.

 

[Sweepr]

I don't get the rumour about 'slower IPC' based on these results. OCed Gefore GTX 980 Ti should be 4.1% faster than OCed Geforce GTX 1070 here (if this was Maxwell), but it's actually a bit slower. Beating that beast with a 1920 SPs VGA using 256-bit GDDR5 is fairly impressive IMHO. Maybe Geforce GTX 1080 could scale a bit better though.

 

[antihelten]

Just keeps showing how good GP104 is. No competition for the rest of 2016 for it

At stock the 1080 is 32% faster than a 980 Ti, when overclocked this shrinks down to 23%. A relative loss of about 7%.

So no this doesn't really show how good GP104 is, on the contrary it shows that it's a worse overclocker than GM200 (which we kinda already knew).

 

[ShintaiDK]

At stock the 1080 is 32% faster than a 980 Ti, when overclocked this shrinks down to 23%. A relative loss of about 7%.

So no this doesn't really show how good GP104 is, on the contrary it shows that it's a worse overclocker than GM200 (which we kinda already knew).

I know you dont like the GP104, but its a beast and unmatched in performance and perf/watt by any other GPU. And it will continue so from the looks of it for a long, long time.

 

[Timmah!]

23% faster for 41% higher core clocks and 39% higher memory clocks? Something is clearly bottlenecking the 1080.

It has less CUDA cores than 980Ti, you gotta take those into account. 1080 is not a better GPU at every single parameter than 980Ti - thats why the perf difference between them is so meh.

Anyway, all this talk about Pascal having lesser IPC than Maxwell...cause the paper stats dont project themselves into measurable ingame performance - well comparably to GM200 anyway...could it be the lack of ROPs is the main reason for this? If yes, would the CUDA compute apps suffer from it too? Are ROPs used in the computing stuff?

 

[flash-gordon]

GTX 1080 is 51% faster than GTX 980

Really? 2 nodes jump and just 51% faster for almost 50% price increase.

 

[ShintaiDK]

Really? 2 nodes jump and just 51% faster for almost 50% price increase.

Its not 2 nodes. More like 1+finfet.

GM200=8B transistors 601mm2.
GP104=7.2B transistors 314mm2.
GP100=16B transistors 610mm2.

 

[Head1985]

Btw they used reference throtthling 980TI

Look at this:
He have 10k 3D mark score on 1480Mhz 980TI(aftermarket) vs their throttling 980TI claiming at 1450Mhz, but it is throtthling and they have only 8700 points
http://www.overclock.net/t/1601896/...80ti-vs-gtx-1070-vs-gtx-1080/60#post_25224292
So Aftermarket non throttling 980TI will be 11-15% faster than max OC 1070 and much more close to GTX1080.Just add 11% to all GTX980TI results and you have idea.

 

[antihelten]

I know you dont like the GP104, but its a beast and unmatched in performance and perf/watt by any other GPU. And it will continue so from the looks of it for a long, long time.

Actually I do like the GP104, I think it's a very impressive GPU. I don't particularly like the price of it nor the FE cooler (or perhaps more accurately, Nvidias PR claims about the cooler), but that's it really.

The GP104 is absolutely a beast, but my point was that this particular test doesn't actually help portray that, since it focuses on one of the only weak points of GP104, it's poor overclockability (relative to GM200, which admittedly was an exceedingly good overclocker).

How long it will remain unmatched is up in the air at this point. I'm not sure if I believe the Vega rumours of launch in October, but if true it could potentially be a rather short stay at the top (there's also the GP102 of course).

Btw they used reference throtthling 980TI

Look at this:
He have 10k 3D mark score on 1480Mhz 980TI(aftermarket) vs their throttling 980TI claiming at 1450Mhz, but it is throtthling and they have only 8700 points
http://www.overclock.net/t/1601896/...80ti-vs-gtx-1070-vs-gtx-1080/60#post_25224292
So Aftermarket non throttling 980TI will be 11-15% faster than max OC 1070 and much more close to GTX1080.Just add 11% to all GTX980TI results and you have idea.

To be fair they are also using the reference throttling 1080/1070 (founders edition). So isn't this just apples to apples?

 

[Face2Face]

I gotta say, it's an impressive showing for performance per watt. Interested to see how well Polaris does in this regard as well.

 

[parvadomus]

Not impressive. People forgets that this performance is basically achieved by a node shrink and nothing more.

 

[kraatus77]

I don't get the rumour about 'slower IPC' based on these results. OCed Gefore GTX 980 Ti should be 4.1% faster than OCed Geforce GTX 1070 here (if this was Maxwell), but it's actually a bit slower. Beating that beast with a 1920 SPs VGA using 256-bit GDDR5 is fairly impressive IMHO. Maybe Geforce GTX 1080 could scale a bit better though.

compare their tflops difference vs performance difference, pascal is less perf/tflops compared to maxwell. that means lower ipc. it's around 3-5% slower. and basically same as hawaii's ipc. ( according to tpu chart)

 

[Det0x]

Btw they used reference throtthling 980TI

Look at this:
He have 10k 3D mark score on 1480Mhz 980TI(aftermarket) vs their throttling 980TI claiming at 1450Mhz, but it is throtthling and they have only 8700 points
http://www.overclock.net/t/1601896/...80ti-vs-gtx-1070-vs-gtx-1080/60#post_25224292
So Aftermarket non throttling 980TI will be 11-15% faster than max OC 1070 and much more close to GTX1080.Just add 11% to all GTX980TI results and you have idea.

Bingo.. Yet some people refuse to acknowledge this :thumbsdown:

 

[Sweepr]

Btw they used reference throtthling 980TI

If you want to compare the best custom Geforce GTX 980 Ti out there do it against a similar Geforce GTX 1080/1070 model, not Founder's Edition. Also weren't many people here saying FE also throttles? So I don't see the problem.

Used/abused Geforce GTX 980 Ti is simply not a better card than a new Geforce GTX 1070, unless you find if it for a very low price. Similar performance but older architecture, much more power hungry and none of Pascal features.

 

[zentan]

compare their tflops difference vs performance difference, pascal is less perf/tflops compared to maxwell. that means lower ipc. it's around 3-5% slower. and basically same as hawaii's ipc. ( according to tpu chart)

Can't be sure of accuracy in such measurements. The ROP configuration is different 96 vs 64 and a bit different memory structure overall.
If such comparisons can be made with such minute accuracy while having very different configurations then you will soon have some fellows saying how GCN SPs in Fiji chips were a regression in performance from those in Hawaii since they definitely don't scale that much in all scenarios over Hawaii although it's widely known that Fiji is limited elsewhere.

 

[thilanliyan]

1080 is a beast. AMD better bring something to the high end before the year is up!!
Too bad the 1080 costs CAD$900+ here in Canada.

 

[coercitiv]

If you want to compare the best custom Geforce GTX 980 Ti out there do it against a similar Geforce GTX 1080/1070 model, not Founder's Edition. Also weren't many people here saying FE also throttles? So I don't see the problem.

1080 FE is a premium card (+$100), using premium components including vapor chamber cooling, hence should not be compared with reference 980Ti.

See how easily this kind of argument can be turned around?

PS: wouldn't it be more appropriate to to keep all cards from throttling, since this is an oc test?

 

[Sweepr]

1080 FE is a premium card (+$100), using premium components including vapor chamber cooling, hence should not be compared with reference 980Ti.

See how easily this kind of argument can be turned around?

Personally I wouldn't buy it, seeing how custom models like ASUS's Strix perform in comparison.

 

[parvadomus]

1080 is a beast. AMD better bring something to the high end before the year is up!!
Too bad the 1080 costs CAD$900+ here in Canada.

AMD can easily match it with a scaled Polaris to 1.5x, 96rops, 384-bits probably same power consumption and die size of GP104.
But I think they are trying to secure as much market share as possible first, replacing their inefficient hawaii and tonga chips.

 

[Headfoot]

I like the comment about "noise commitment." How very meaningless. Of course coolers running at 100% fan speed cool well. The reference 290 fan cools well at 100% fan speed. It also sounds like a jet engine. As has been abundantly clear since the mid 90's, the way to qualify coolers and fans are the noise/perf, perf/$ ratios.

 

[kraatus77]

Can't be sure of accuracy in such measurements. The ROP configuration is different 96 vs 64 and a bit different memory structure overall.
If such comparisons can be made with such minute accuracy while having very different configurations then you will soon have some fellows saying how GCN SPs in Fiji chips were a regression in performance from those in Hawaii since they definitely don't scale that much in all scenarios over Hawaii although it's widely known that Fiji is limited elsewhere.

than compare it to 980, same rops there (same regression in ipc). as long as architecture is same (fiji/hawaii) you can compare them if everything is scaling. like tx was 50% better in everything beside clocks vs 980. that's not the case with fiji. so we cannot compare them. because they didn't change shaders's arch. and didn't scale everything irt shaders.

you will hardly find 2 different gpus with everything same. so going by that logic we cannot really compare anything. in that case we compare gpus which has more similarities. so in this case compare 1070/980 and you will still find regression in ipc.

best comparison is 390/1080 , both have same cc,rop, and even bandwidth. and going by tpu chart , 1080 is 78% faster than 390 by having 80% more frequency. so that puts ipc pretty even.


also before pascal , people always said gcn's tflops doesn't really translate to gaming performance compared to nvidia. but nobody talked about 980ti's rop advantage. suddenly it's something to consider now ?

why people always change goalposts ? the fact is pascal has lower ipc than maxwell. it's doing well because of frequency advantage. run similar gpus at same clocks and you will see maxwell gpu will beat pascal.

i'm not saying it's a bad thing, but certainly not really as impressive as it was compared to kepler vs maxwell.

 

[beginner99]

Really? 2 nodes jump and just 51% faster for almost 50% price increase.

Exactly. That is the real fail here. the problem isn't performance, it's price. It's pure rip-off and NV fanboys are all falling into the trap (or at least pretend to, you know they might just be PR-Guys trolling forums).

 

[zentan]

than compare it to 980, same rops there (same regression in ipc). as long as architecture is same (fiji/hawaii) you can compare them if everything is scaling. like tx was 50% better in everything beside clocks vs 980. that's not the case with fiji. so we cannot compare them. because they didn't change shaders's arch. and didn't scale everything irt shaders.

you will hardly find 2 different gpus with everything same. so going by that logic we cannot really compare anything. in that case we compare gpus which has more similarities. so in this case compare 1070/980 and you will still find regression in ipc.

best comparison is 390/1080 , both have same cc,rop, and even bandwidth. and going by tpu chart , 1080 is 78% faster than 390 by having 80% more frequency. so that puts ipc pretty even.


also before pascal , people always said gcn's tflops doesn't really translate to gaming performance compared to nvidia. but nobody talked about 980ti's rop advantage. suddenly it's something to consider now ?

why people always change goalposts ? the fact is pascal has lower ipc than maxwell. it's doing well because of frequency advantage. run similar gpus at same clocks and you will see maxwell gpu will beat pascal.

i'm not saying it's a bad thing, but certainly not really as impressive as it was compared to kepler vs maxwell.

Neither am i saying it's bad but the 3-5% of measurement accuracy of it and given the way the boost clocks behave(very random since it's intro with kepler and more so maxwell onwards). 3-5% may be well right or may fall into region of error. Just that. It's just not ROPs as well,there are some improvements in other areas which might not be taken advantage at the moment but when if they come into act this 3-5% may shift other direction.
Anyway it's not a big issue.