Using GTX 1070 to speculate GTX 1060 performance

[JustMe21]

Since the base clock speed and memory speed are the same, couldn't we use the results for the GTX 1070 to speculate on the 1060 performance? Granted the 1060 has a smaller memory bus, but the higher GPU boost rate offsets that somewhat.

So, I was looking at the Guru 3D review of the RX 480 and using the 1920 x 1080 scores since that's where the 1060 will do best.

On 3DMark Fire Strike, the 1070 gets a score of 16229. So, divide that by the 1920 CUDA cores and you get about 8.453 points per CUDA core. Now, multiply that by the 1280 CUDA cores the 1060 has and you get 10820.

On the Hitman test, the 1070 gets 86 FPS. Divide that by 1920 and you have .045 FPS per CUDA core. So, 53 FPS for the 1060.

For Battlefield Hardline, the 1070 gets 101 FPS or .0562 FPS per CUDA core. So, the 1060 gets 67.33 FPS.

In the end, the 1060 looks like a definite 970 killer, but not a 980 killer. Performance in DX11 is about on par with the RX 480, but, the 1060 will definitely have much better power efficiency.

 

[guachi]

The 1080 is about 60% faster than a 980.
The 1070 is about 60% faster than a 970.
So I suspect the 1060 will be about 60% faster than a 960.

Which puts it almost exactly as fast as a 480, though nVidia claims it's 15% faster. If true, the 1060 either got a little something extra from nVidia to beat the 480 or the 960 was a relatively poor card to begin with.

 

[BTRY B 529th FA BN]

I read it's somewhere around 71% faster.. I think it was on the Anandtech front page.

http://www.anandtech.com/show/10474/nvidia-announces-geforce-gtx-1060-july-19

 

[ultima_trev]

The 1080 is about 60% faster than a 980.
The 1070 is about 60% faster than a 970.
So I suspect the 1060 will be about 60% faster than a 960.

Which puts it almost exactly as fast as a 480, though nVidia claims it's 15% faster. If true, the 1060 either got a little something extra from nVidia to beat the 480 or the 960 was a relatively poor card to begin with.

GTX 960 was indeed a poor design, 128-bit memory interface and 32 ROPs versus 256-bit memory interface and 64 ROPs of GTX 980. Now we have GTX 1060 with a 192-bit memory interface and 48 ROPs, which is much closer as GTX 1070 and 1080 don't have a bigger memory bus or ROP cluster versus GTX 980.

 

[JustMe21]

Nvidia was definitely smart on their marketing. People look at the Geforce 1080 (2560 CUDA cores) and compare it to the Geforce 980 (2048 CUDA cores), when it really should have been called a Geforce 1080 Ti since the 980 Ti has 2816 CUDA cores. The 1070 with its 1920 CUDA cores is more in line with the 980, so it would have fit better to be called the 1080.

Of course, the 1060 with it's 1280 CUDA cores is in between a 970 with 1664 CUDA cores and a 960 with 1024 CUDA cores, so it seems like it should have been called a Geforce 1065.

If you think about it in those terms, the performance gains aren't as good, although the power efficiency is still a lot better than previous generations.

Well, looking back at the previous generation, the lower end cards did have more points per CUDA core, so with the leaked score of Fire Strike score of 11225, the 8.77 points per CUDA core does seem reasonable.

Other Fire Strike averages
1080 (2560 CUDA cores) - 19370 - 7.57 points per CUDA core
1070 (1920 CUDA cores) - 16229 - 8.45 points per CUDA core

980Ti (2816 CUDA cores) - 15656 - 5.56 point per CUDA core
980 (2048 CUDA cores) - 11168 - 5.45 points per CUDA core
970 (1664 CUDA cores) - 9568 - 5.75 points per CUDA core
960 (1024 CUDA cores) - 6485 - 6.33 points per CUDA core

It is interesting that since the previous generation were close in points per CUDA core, it seems like the 1080 isn't performing to its full potential. Could it be CPU limitation?

 

[Flapdrol1337]

GTX 960 was indeed a poor design, 128-bit memory interface and 32 ROPs versus 256-bit memory interface and 64 ROPs of GTX 980. Now we have GTX 1060 with a 192-bit memory interface and 48 ROPs, which is much closer as GTX 1070 and 1080 don't have a bigger memory bus or ROP cluster versus GTX 980.

Poor design? It was fine, it's half a 980, if you have half the cores you only need half the memory bandwidth, it gives pretty much exactly half the performance.

Now that the 1080 has gddr5x on a 256 bit bus you need a relatively wider bus if you want to use gddr5 on the 1060. If gddr5x was cheap and easily available they might have gone for 128bit gddr5x.

 

[Head1985]

I tried that.I downclock my 1070 to 1350Mhz and memory to match 192GB/s.Still it was 10% faster than GTX1060 will be(1070 must be at 1250-1270mhz to match GTX1060 Teraflops.I think GTX1060 will boost at 1800-1900Mhz stock)
Heaven 70fps at 1367Mhz so GTX1060 will be 10% slower at 63Fps.That is on GTX980 level

I also run Crysis3 benchmark and GTX1060 have 47Fps.That is on GTX970 aftermarket level.Slower than stock GTX980
Few more results in crysis3
GTX1070 at 2100-2140/9400
GTX970 at 1500/8000
GTX670 at 1280/7400

Btw stock GTX1070 at 1730-1770Mhz have in this benchmark 73FPS

 

[tviceman]

You can't just downclock a 1070 to simulate 1060 performance because the ratio of ROPs and cores is not 1:1 between the two products. The 1060 is physically 66% the cores and 75% the bandwidth, ROPs, and TMUs. It'll be slightly above 70% of a GTX 1070 at 1440p (maybe a little more at 1080p) due to the slightly higher boost clocks and more bandwidth.

 

[Head1985]

You can't just downclock a 1070 to simulate 1060 performance because the ratio of ROPs and cores is not 1:1 between the two products. The 1060 is physically 66% the cores and 75% the bandwidth, ROPs, and TMUs. It'll be slightly above 70% of a GTX 1070 at 1440p (maybe a little more at 1080p) due to the slightly higher boost clocks and more bandwidth.

Yes i can.
GTX1070 at 1270Mhz and with 192GB/s bandwidth have same TEraflops and bandwidth as GTX1060.
pixel fillrate:
GTX1070 at 1270mhz have 60GP/s
GTX1060 at 1900Mhz have also 60GP/s
I calculate 1070 with 48rops usable and GTX1060 with 32rops usable.Because we already know GTX1070 cant use 64rops(only 48) and GTX1060 with 1280SP cant use 48rops(only32)

Texel fillare:
GTX1070 120TMU at 1270mhz=152GT/s
GTX1060 80TMU at 1900Mhz=152GT/s

TERAFLOPS:
GTX1060 at 1900mhz=4.8TF
GTX1070 at 1270Mhz=4.8TF

SO we have same TERAFLOPS, same memory bandwidth,Same Pixel fillrate and same Texel fillrate.

 

[Head1985]

Btw witcher3 results 1440p max 4xhairworks and DOF off
GTX1060 simulated
Avg: 32.108 - Min: 25.76 - Max: 37.72
GTX1070 stock 1710-1770/8000
Avg: 44.851 - Min: 39 - Max: 52
GTX1070 2114/9400
Avg: 53.501 - Min: 46 - Max: 63

1070 is 40% faster and GTX1060 most likely have GTX980 performance in this game.In crysis3 it will be slower than GTX980.

EDIT crysis3 results 1440P max +4MSAA
GTX1060 simulated
Avg: 30.9 - Min: 27 - Max: 38
Stock 1070 1730-1770/8000
Avg: 45.959 - Min: 40 - Max: 56
1070 2120/9400
Avg: 55.162 - Min: 48 - Max: 67
In crysis3 GTX1070 is 48% faster than GTX1060

 

[tviceman]

Fake bench away to your heart's content. Real benches will come in a week or so.

 

[Head1985]

Fake bench away to your heart's content. Real benches will come in a week or so.

i didnt create this thread WTH you wnat from me?I only do my best to simulate GTX1060.

 

[AnandThenMan]

i didnt create this topic WTH you wnat from me?I only do my best to simulate GTX1060.

Ignore the trolling thanks for the taking the time to simulate the 1060 numbers.

 

[Azix]

might as well use more relevant cards. 980 and 480 based on 1070 performance and teraflops vs 980ti or 980

 

[Head1985]

Ok last game Assassins creed syndicate
2560x1440 all max +FXAA
GTX1060 simulated
Avg: 31.5 - Min: 29 - Max: 34
Stock GTX1070
Avg: 45.034 - Min: 41 - Max: 48
GTX1070 2100/9400
Avg: 54.737 - Min: 50 - Max: 58

GTX1070 is 42% faster.btw pretty similar results with techpower UP using GTX1070 and 980.GTX1070 is there 43% faster than GTX980.

 

[tg2708]

You can't just downclock a 1070 to simulate 1060 performance because the ratio of ROPs and cores is not 1:1 between the two products. The 1060 is physically 66% the cores and 75% the bandwidth, ROPs, and TMUs. It'll be slightly above 70% of a GTX 1070 at 1440p (maybe a little more at 1080p) due to the slightly higher boost clocks and more bandwidth.

I agree, their are other things in play that affects performance. Its a good effort but I'd leave its performance for true benchmarks.

 

[JustMe21]

Since we have to wait until July 19 for actual reviews, it's fun to speculate. It seems that hype on any new video card isn't quite what the reality turns out to be.

 

[Hitman928]

Head, thanks for taking the time to run those benches. It will be interesting to see how close your simulated performance numbers are to the real card. You don't happen to have Ashes do you?

 

[boozzer]

Nvidia was definitely smart on their marketing. People look at the Geforce 1080 (2560 CUDA cores) and compare it to the Geforce 980 (2048 CUDA cores), when it really should have been called a Geforce 1080 Ti since the 980 Ti has 2816 CUDA cores. The 1070 with its 1920 CUDA cores is more in line with the 980, so it would have fit better to be called the 1080.

Of course, the 1060 with it's 1280 CUDA cores is in between a 970 with 1664 CUDA cores and a 960 with 1024 CUDA cores, so it seems like it should have been called a Geforce 1065.

If you think about it in those terms, the performance gains aren't as good, although the power efficiency is still a lot better than previous generations.

Well, looking back at the previous generation, the lower end cards did have more points per CUDA core, so with the leaked score of Fire Strike score of 11225, the 8.77 points per CUDA core does seem reasonable.

Other Fire Strike averages
1080 (2560 CUDA cores) - 19370 - 7.57 points per CUDA core
1070 (1920 CUDA cores) - 16229 - 8.45 points per CUDA core

980Ti (2816 CUDA cores) - 15656 - 5.56 point per CUDA core
980 (2048 CUDA cores) - 11168 - 5.45 points per CUDA core
970 (1664 CUDA cores) - 9568 - 5.75 points per CUDA core
960 (1024 CUDA cores) - 6485 - 6.33 points per CUDA core

It is interesting that since the previous generation were close in points per CUDA core, it seems like the 1080 isn't performing to its full potential. Could it be CPU limitation?

woah, that is a really good point. :thumbsup:

 

[doggyfromplanetwoof]

Nvidia was definitely smart on their marketing. People look at the Geforce 1080 (2560 CUDA cores) and compare it to the Geforce 980 (2048 CUDA cores), when it really should have been called a Geforce 1080 Ti since the 980 Ti has 2816 CUDA cores. The 1070 with its 1920 CUDA cores is more in line with the 980, so it would have fit better to be called the 1080.

Of course, the 1060 with it's 1280 CUDA cores is in between a 970 with 1664 CUDA cores and a 960 with 1024 CUDA cores, so it seems like it should have been called a Geforce 1065.

If you think about it in those terms, the performance gains aren't as good, although the power efficiency is still a lot better than previous generations.

Well, looking back at the previous generation, the lower end cards did have more points per CUDA core, so with the leaked score of Fire Strike score of 11225, the 8.77 points per CUDA core does seem reasonable.

Other Fire Strike averages
1080 (2560 CUDA cores) - 19370 - 7.57 points per CUDA core
1070 (1920 CUDA cores) - 16229 - 8.45 points per CUDA core

980Ti (2816 CUDA cores) - 15656 - 5.56 point per CUDA core
980 (2048 CUDA cores) - 11168 - 5.45 points per CUDA core
970 (1664 CUDA cores) - 9568 - 5.75 points per CUDA core
960 (1024 CUDA cores) - 6485 - 6.33 points per CUDA core

It is interesting that since the previous generation were close in points per CUDA core, it seems like the 1080 isn't performing to its full potential. Could it be CPU limitation?

This guy is chief.

 

[Wall Street]

Nvidia was definitely smart on their marketing. People look at the Geforce 1080 (2560 CUDA cores) and compare it to the Geforce 980 (2048 CUDA cores), when it really should have been called a Geforce 1080 Ti since the 980 Ti has 2816 CUDA cores. The 1070 with its 1920 CUDA cores is more in line with the 980, so it would have fit better to be called the 1080.

I don't really agree with your analysis framework. The GTX 1080 isn't direct shrink of anything. Counting cores isn't really a helpful metric. The GTX 1080 isn't called a Ti because traditional the Ti part has been reserved for a ~500 mm2 chip with a 385-bit memory bus. One of the reasons that the GTX 1080 looks unimpressive given the number of cores and the clock speed it has is because the 384-bit bus at 7 gbps of the GTX 980Ti actually provides 5% more bandwidth than is available on the GTX 1080. The eventual GTX 1080 Ti will have either 385-bit GDDR5x or HBM 2 to make sure that it crushes the GTX 980Ti in ROP count and memory bandwidth. ROPs, texture units and memory bandwidth impact performance in a way that makes performance per shader less useful as a metric.

Of course the GTX 1080 has more cores than the GTX 980 and similar core count to the GTX 980 Ti. If they kept the core counts similar for cards with the same marketing tier between generations then the GTX 1060 should be called the GTX 1080 because the GTX 680 had 1536 cores and the 1280 core GTX 1060 has the most similar core count this generation.

 

[nkdesistyle]

Yes i can.
GTX1070 at 1270Mhz and with 192GB/s bandwidth have same TEraflops and bandwidth as GTX1060.
pixel fillrate:
GTX1070 at 1270mhz have 60GP/s
GTX1060 at 1900Mhz have also 60GP/s
I calculate 1070 with 48rops usable and GTX1060 with 32rops usable.Because we already know GTX1070 cant use 64rops(only 48) and GTX1060 with 1280SP cant use 48rops(only32)

Texel fillare:
GTX1070 120TMU at 1270mhz=152GT/s
GTX1060 80TMU at 1900Mhz=152GT/s

TERAFLOPS:
GTX1060 at 1900mhz=4.8TF
GTX1070 at 1270Mhz=4.8TF

SO we have same TERAFLOPS, same memory bandwidth,Same Pixel fillrate and same Texel fillrate.

you make sense with everything but I absolutely don't believe that 1070 doesn't use all 64 ROPS, it is not cut down the way 970 was. It's directly related to memory bus as well. Almost every review has said it uses full specs and there is no misinformation about it. If it had less ROPs than stated it would also be true it is using less memory just like the 970. So it is indeed full 64 ROPS and it can use all of them. Yes your numbers may be correct because ROPS won't make that much of a difference to be honest.

 

[JustMe21]

I don't really agree with your analysis framework. The GTX 1080 isn't direct shrink of anything. Counting cores isn't really a helpful metric. The GTX 1080 isn't called a Ti because traditional the Ti part has been reserved for a ~500 mm2 chip with a 385-bit memory bus. One of the reasons that the GTX 1080 looks unimpressive given the number of cores and the clock speed it has is because the 384-bit bus at 7 gbps of the GTX 980Ti actually provides 5% more bandwidth than is available on the GTX 1080. The eventual GTX 1080 Ti will have either 385-bit GDDR5x or HBM 2 to make sure that it crushes the GTX 980Ti in ROP count and memory bandwidth. ROPs, texture units and memory bandwidth impact performance in a way that makes performance per shader less useful as a metric.

Of course the GTX 1080 has more cores than the GTX 980 and similar core count to the GTX 980 Ti. If they kept the core counts similar for cards with the same marketing tier between generations then the GTX 1060 should be called the GTX 1080 because the GTX 680 had 1536 cores and the 1280 core GTX 1060 has the most similar core count this generation.

The reason I did an average based on CUDA cores is because that's what Nvidia likes to show on their specs for their site. It would be great to have an actual formula that would take into account the memory bandwidth, memory bus size, TMUs, stream processors, ROPs, MHz, and CUDA cores to show a more comparable performance.

 

[Head1985]

I tried that.I downclock my 1070 to 1350Mhz and memory to match 192GB/s.Still it was 10% faster than GTX1060 will be(1070 must be at 1250-1270mhz to match GTX1060 Teraflops.I think GTX1060 will boost at 1800-1900Mhz stock)
Heaven 70fps at 1367Mhz so GTX1060 will be 10% slower at 63Fps.That is on GTX980 level

Hey i was right.GTX1060 have 63.9Fps in heaven
http://videocardz.com/62122/nvidia-geforce-gtx-1060-official-performance-leaked

 

[Head1985]

btw simulated Tomb raider dx12 very high preset:
GTX1070 at 1400Mhz with 192GB/s have 36.5fps and to match GTX1060 Teraflops i need 1270Mhz.1400Mhz is 10% faster than 1270mhz so i need lower performance by 10%.

36.5-10% is 32.8Fps.

Simulated GTX1060 will have 32.8FPs.Pretty much on spot with VC results.