GTX 1070 have only 1920SP!

[jabroni619]

Not to most 980Ti owners it isn't, for example my 980Ti is 20% faster out of the box then stock so 1070 has a lot of ground to cover to be faster. GM200 was really under-clocked in reference cards which doesn't seem to be the case with GP200.

I highly doubt nVidia is trying to target existing 980Ti owners with the 1070. They aren't trying to give you a card that's 50-80% more powerful for <$400. That's why the 1080 exists. There's a ton of people who would love a 980Ti who are either unwilling or unable to spend the coin on one. 1070 gives that that level of performance at a much lower cost.

 

[Lepton87]

I highly doubt nVidia is trying to target existing 980Ti owners with the 1070. That's why the 1080 exists. There's a ton of people who would love a 980Ti who are either unwilling or unable to spend the coin on one. 1070 gives that that level of performance at a much lower cost.

With the prices the 980Ti is going for now I wouldn't be so sure 1070 is going to be much cheaper. I picked my for the price that 980 were selling for just a while ago and this is a card that used to be 700$ a full 100$ over reference and one of the more expensive aftermarket 980ti versions.

 

[RussianSensation]

Your math is a bit off here. Pascal doesn't have 11.5-14% lower IPC than Maxwell, it has 5-6% lower IPC (1.67-1.69 divided by 1.78 is equal to 0.94-0.95). Although TPU actually shows the average clock of 1080 as 1783 MHz, not 1733 MHz, so actual the IPC gap is roughly 10%. But then if one were to compare to the 980 Ti it goes back down to 5%. So all in all GP104 has 5-10% lower IPC.

That's not how math works. If the chart's fastest card is rated at 100%, and the 2nd card is rated at 78% and the 3rd is rated at 67%, it's not 1.67 / 1.78, but exactly 67%/78% as I stated. Furthermore, using your 1783mhz target actually supports my point even more because it means the IPC difference is even greater than 11.5-14%, not lower as you imply. In any case, your and my analysis arrives at the same conclusion that 1070 should be faster than a 980Ti reference.

Besides, if it performs like a 980Ti at a fraction of the price and power draw, the specs really shouldn't bother anyone.

Sure they should because all next generation mid-range cards from NV have matched or outperformed last gen's flagship card. The difference is the x70 card was never this neutered. Even the famously crippled GTX470 still outperformed the GTX285 by 12% at launch, and that lead extended over time.

The GTX470 also easily matched a stock 480.

Using your logic, as long as the next gen card is faster and costs less, it's an automatic win. Ok fine, so why doesn't NV split the generation into 4 parts and release a new card 10% faster every 6 months?

 

[jabroni619]

With the prices the 980Ti is going for now I wouldn't be so sure 1070 is going to be much cheaper. I picked my for the price that 980 were selling for just a while ago and this is a card that used to be 700$ a full 100$ over reference and one of the more expensive aftermarket 980ti versions.

980Ti's are still way more then a 1070 @ MSRP and haven't dropped in power consumption, then there's VR where the 1070 will clearly be the better option. Even with the price drops, someone looking for 980Ti level performance is better off with a 1070 in just about every metric you look at.

Using your logic, as long as the next gen card is faster and costs less, it's an automatic win. Ok fine, so why doesn't NV split the generation into 4 parts and release a new card 10% faster every 6 months?

If it's faster and costs less it is an automatic win. Why wouldn't it be? I'm not sure what relevance your last sentence has to the first.

 

[antihelten]

That's not how math works. If the chart's fastest card is rated at 100%, and the 2nd card is rated at 78% and the 3rd is rated at 67%, it's not 1.67 / 1.78, but exactly 67%/78% as I stated. Furthermore, using your 1783mhz target actually supports my point even more because it means the IPC difference is even greater than 11.5-14%, not lower as you imply. In any case, your and my analysis arrives at the same conclusion that 1070 should be faster than a 980Ti reference.

Yes, but the chart does not say 100%, 78% and 67%, it says 100% (980), 178% (theoretical 1080) and 167% (actual 1080).

This would be the same as 56% (980), 100% (theoretical 1080) and 94% (actual 1080). Or 60% (980), 107% (theoretical 1080) and 100% (actual 1080).

Either way you end up with the same 4-5% IPC difference. And yes using the 1783 MHz number gets close to your numbers, but that's a coincidence.

Even the famously crippled GTX470 still outperformed the GTX285 by 12% at launch, and that lead extended over time.

And the 470 was only 12.5% cheaper than the MSRP of the 285 ($350 vs. $400), whereas the 1070 will be 41.5% cheaper than the MSRP of the 980 Ti ($380 vs. $650).

So we got a 28% perf/$ improvement from the 470, and if the 1070 is 5% faster than the 980 Ti* we will get an 80% perf/$ improvement this time around.

So basically the 1070 provides a 40% bigger jump in perf/$ than the 470 did compared to the previous generations flagship.

I don't really see the issue here.

*Jen Hsun said the 1070 was faster than a Titan X at the presentation.

 

[RussianSensation]

If it's faster and costs less it is an automatic win. Why wouldn't it be? I'm not sure what relevance your last sentence has to the first.

Because that's not how it works. 960 was faster than 760 but it was the worst x60 series in the last 5 generations. Just because the new card is faster, doesn't mean it's a good buy.
http://www.computerbase.de/2015-03/geforce-gtx-460-560-660-760-960-vergleich/2/

Using your logic, Volta GTX2070 5% faster than 1080Ti for $549. Good deal. Then next generation GTX2080 goes up to $799 since it's 30% faster than a $799 GTX1080Ti, making it a good deal too. :thumbsup: Sounds like you haven't thought this through or you are very new to the PC gaming landscape. Out of curiosity, how old are you so just so I know if you actually studied the history of GPUs or even had PCs prior to 2011?

 

[MrTeal]

980Ti's are still way more then a 1070 @ MSRP and haven't dropped in power consumption, then there's VR where the 1070 will clearly be the better option. Even with the price drops, someone looking for 980Ti level performance is better off with a 1070 in just about every metric you look at.



If it's faster and costs less it is an automatic win. Why wouldn't it be? I'm not sure what relevance your last sentence has to the first.

Newegg had an EVGA ACX2.0 980Ti for $500 before a $25 MIR the other day. That's still above launch prices, but it's not much more than the $450 that the first FE cards will be going for. How long stock on new sub-$500 980Ti's will last is anyone's guess though. Reviews on the 1070 will be interesting to say the least.

 

[jabroni619]

Because that's not how it works. 960 was faster than 760 but it was the worst x60 series in the last 5 generations. Just because the new card is faster, doesn't mean it's a good buy.
http://www.computerbase.de/2015-03/geforce-gtx-460-560-660-760-960-vergleich/2/

Using your logic, Volta GTX2070 5% faster than 1080Ti for $549. Good deal. Then next generation GTX2080 goes up to $799 since it's 30% faster than a $799 GTX1080Ti, making it a good deal too. :thumbsup: Sounds like you haven't thought this through or you are very new to the PC gaming landscape. Out of curiosity, how old are you so just so I know if you actually studied the history of GPUs or even had PCs prior to 2011?

I'm in my 30's and have been gaming since the Wolfenstein 3D days. Since you seem to think you have it all figured out, enlighten me. If it's such a bad buy, name one card you can get with 980Ti performance for $380. I'm sure you have an extensive list given your last few posts.

 

[antihelten]

Using your logic, Volta GTX2070 5% faster than 1080Ti for $549. Good deal. Then next generation GTX2080 goes up to $799 since it's 30% faster than a $799 GTX1080Ti, making it a good deal too. :thumbsup:

The 1070 costs 60% as much as the 980 Ti, so the comparable situation would be if the Volta GTX 2070 also cost 60% as much as the 1080 Ti.

That would then be a GTX 2070 5% faster than a 1080 Ti for $350. That is indeed a pretty good deal.

 

[IllogicalGlory]

My estimation of the IPC loss:

(2560 * 1783)/(2816 * 1128) = 1.44x expected increase
actual increase is 1.37x

Loss = 100% - actual/expected = 100% - 1.37/1.44 = 5.1%

Pascal has about 5% less IPC than Maxwell based on that very rough calculation, which is cuda cores times average clockspeed per techpowerup.

 

[antihelten]

My estimation of the IPC loss:

(2560 * 1783)/(2816 * 1128) = 1.44x expected increase
actual increase is 1.37x

Loss = 100% - actual/expected = 100% - 1.37/1.44 = 5.1%

Pascal has about 5% less IPC than Maxwell based on that very rough calculation, which is cuda cores times average clockspeed per techpowerup.

Exactly the same as I got versus the 980 Ti (5% loss). But it's worth noting that the math looks slightly worse against the 980 (roughly a 10% loss). So it's probably more accurate to say a 5-10% loss for Pascal versus Maxwell.

 

[Arachnotronic]

I'm in my 30's and have been gaming since the Wolfenstein 3D days. Since you seem to think you have it all figured out, enlighten me. If it's such a bad buy, name one card you can get with 980Ti performance for $380. I'm sure you have an extensive list given your last few posts.

I'm coming up blank on trying to find a 980 Ti class card for $380. Some posters here just want to hate on NVIDIA because they feel "entitled" to more performance/$ than they are being offered. They also take serious issue with NVIDIA making good margins on products that people like to buy and would prefer if all vendors operated at low margins & posted losses.

 

[casiofx]

I'm coming up blank on trying to find a 980 Ti class card for $380. Some posters here just want to hate on NVIDIA because they feel "entitled" to more performance/$ than they are being offered. They also take serious issue with NVIDIA making good margins on products that people like to buy and would prefer if all vendors operated at low margins & posted losses.

As of 19 may 2016, there are none.
1080 and 1070 can't be sold to normal consumers yet, you can only get it in the future. Ahahaha get it?

The point is your point will probably valid when the 1070 hits the store shelfs, but apparently at this time of writing it is not, so your conclusion is not correct, you're predicting the future.

 

[Lepton87]

Pascal has about 5% less IPC than Maxwell based on that very rough calculation, which is cuda cores times average clockspeed per techpowerup.

You didn't take the memory bandwidth into the equation. It doesn't make sense for the pascal cores to be slower than the maxwell cores. Those are largely the same NV didn't make any significant changes to the architecture and they certainly didn't mess with the cores throughput. SMM design is the same. Pascal differs very little from maxwell the change is something similar to mawell 1.0 to maxwell 2.0 transition. They added a few features but they still couldn't get rid of their weak spots. Remember VLIW4? That was a pretty big change from VLIW5 but they still couldn't completely address their biggest shortcoming that was the tessellation performance. AMD doubled their tessellation throughput but it still trailed the competition just like pascal still is not great in DX12. The depth of changes doesn't seem to be as deep as the transition from VLIW5 to VLIW4 either. They could just call the architecture maxwell 3.0. Not making a radical architecture overhaul is actually the only reason I decided to replace my burned 980Ti with an another 980Ti, because if they did a significant architecture tweak then we could see the kepler fiasco all over again. With the things as they are I think NV can reuse Maxwell's driver maturity for Pascal. Obviously it also means we won't have any significant improvement in performance due to drivers as they are already mature but on the flip-side we shouldn't see maxwell 2.0 cards tanking in performance like their Kepler brethren at least not until the next significant architecture change - Volta.

 

[lopri]

GP104 is almost a direct shrink of GM204. Everything should behave likewise other than the clock speed and power consumption which is largely attributable to the node change.

 

[lopri]

And it is likely that NV is trying to salvage as many dies as possible so the margins of error needs to be conservative. Thus GTX 1070.

 

[jabroni619]

I'm coming up blank on trying to find a 980 Ti class card for $380. Some posters here just want to hate on NVIDIA because they feel "entitled" to more performance/$ than they are being offered. They also take serious issue with NVIDIA making good margins on products that people like to buy and would prefer if all vendors operated at low margins & posted losses.

Yeah, i'm at a loss as to what was going on with those replies. To go through all that analytical trouble and completely fail to realize, or intentionally ignore the 1070 is by far the leader in terms of priceerformance for high end GPU's is a little dubious. Granted, this may all change when AMD shows us it's cards, but to pass off the 1070 as a "bad buy" given what we currently know about it is silly.

 

[IEC]

Yeah, i'm at a loss as to what was going on with those replies. To go through all that analytical trouble and completely fail to realize, or intentionally ignore the 1070 is by far the leader in terms of priceerformance for high end GPU's is a little dubious. Granted, this may all change when AMD shows us it's cards, but to pass off the 1070 as a "bad buy" given what we currently know about it is silly.

There are zero reviews of the 1070, much less Polaris 10. It's a little early to say it's going to be a price performance leader when neither it nor its closest competition have launched.

 

[jabroni619]

There are zero reviews of the 1070, much less Polaris 10. It's a little early to say it's going to be a price performance leader when neither it nor its closest competition have launched.

Yes, its a little premature to say with 100% certainty, but it's hardly going out on a limb either. It would have to have a street price significantly higher then MSRP and perform significantly worse then what nVidia claims for it not to be the clear high end priceerformance king. I did mention that could all change when AMD reveals it's cards.

 

[Thinker_145]

There are zero reviews of the 1070, much less Polaris 10. It's a little early to say it's going to be a price performance leader when neither it nor its closest competition have launched.

If the GTX 1070 matches the 980Ti and it actually retails for $380 then its a freaking home run. I don't see how anyone can say otherwise.

I mean the 980Ti is what 40% faster than a 970 and used to cost twice the price.

Sent from my HTC One M9 using Tapatalk

 

[Wall Street]

If the GTX 1070 matches the 980Ti and it actually retails for $380 then its a freaking home run. I don't see how anyone can say otherwise.

I think that people expect a lot from a die shrink. We have had 2048 shader parts since 2012 and you can buy a new 2048 shader card in the $200 range now. Of course the GTX 1070 makes up for it in clocks.

Also, for the people who think that any card that performs better than the old king for a lower price is a good deal, this isn't necisarily the case. Exponential improvements in price to performance have been an annual hallmark of the semiconductor industry for decades. New cards which underperform the expected improvement from Moore's law can be disappointing.

 

[mohit9206]

Cynical Unicorn posted this interesting observation on Overclock.net forum
Rounded to the nearest integer:
The 1070 has 25% fewer CUDA cores than the 1080
The 970 has 20% fewer CUDA cores than the 980
The 780 has 14% fewer CUDA cores than the OG Titan
The 670 has 12% fewer CUDA cores than the 680
The 570 has 6% fewer CUDA cores than the 580
The 470 has 7% fewer CUDA cores than the 480

Notice how the gap keeps getting bigger?
What isn't getting much bigger is the price difference. Looking at launch prices:
The 1070 is $220 less than the 1080
The 970 is $220 less than the 980
The 670 is $100 less than the 680
The 570 is $150 less than the 580
The 470 is $150 less than the 480

The 1070 saves $9 per %CUDAs lost
The 970 saves $11 per %CUDAs lost
The 780 saves $25 per %CUDAs lost
The 670 saves $8 per %CUDAs lost
The 570 saves $25 per %CUDAs lost
The 470 saves $21 per %CUDAs lost

 

[flopper]

Cynical Unicorn posted this interesting observation on Overclock.net forum
Rounded to the nearest integer:
The 1070 has 25% fewer CUDA cores than the 1080
The 970 has 20% fewer CUDA cores than the 980
The 780 has 14% fewer CUDA cores than the OG Titan
The 670 has 12% fewer CUDA cores than the 680
The 570 has 6% fewer CUDA cores than the 580
The 470 has 7% fewer CUDA cores than the 480

Notice how the gap keeps getting bigger?
What isn't getting much bigger is the price difference. Looking at launch prices:
The 1070 is $220 less than the 1080
The 970 is $220 less than the 980
The 670 is $100 less than the 680
The 570 is $150 less than the 580
The 470 is $150 less than the 480

The 1070 saves $9 per %CUDAs lost
The 970 saves $11 per %CUDAs lost
The 780 saves $25 per %CUDAs lost
The 670 saves $8 per %CUDAs lost
The 570 saves $25 per %CUDAs lost
The 470 saves $21 per %CUDAs lost

Midrange 800/500 euro and then in october you have to shift cards again.
Bad value along the way

 

[Armsdealer]

Gentlemen,

Khalid moammar at wccf is claiming with a high degree of confidence that 1300-1400mhz is the core clock of uncut p10 reference.

If in fact that is true, and the uncut die has 36 or 40 compute units, it seems likely it will compete and likely be faster than the 1070. (7+ tflops vs 6.1)

Anybody know if the sisoft bench we've seen online is accurate in reporting number of cores?

 

[selni]

Cynical Unicorn posted this interesting observation on Overclock.net forum
Rounded to the nearest integer:
The 1070 has 25% fewer CUDA cores than the 1080
The 970 has 20% fewer CUDA cores than the 980
The 780 has 14% fewer CUDA cores than the OG Titan
The 670 has 12% fewer CUDA cores than the 680
The 570 has 6% fewer CUDA cores than the 580
The 470 has 7% fewer CUDA cores than the 480

Notice how the gap keeps getting bigger?
What isn't getting much bigger is the price difference. Looking at launch prices:
The 1070 is $220 less than the 1080
The 970 is $220 less than the 980
The 670 is $100 less than the 680
The 570 is $150 less than the 580
The 470 is $150 less than the 480

The 1070 saves $9 per %CUDAs lost
The 970 saves $11 per %CUDAs lost
The 780 saves $25 per %CUDAs lost
The 670 saves $8 per %CUDAs lost
The 570 saves $25 per %CUDAs lost
The 470 saves $21 per %CUDAs lost

That's not unfair wrt the 1070 but not the whole story historically (and there's some weird stuff in there - titan in a x80 vs x70 value comparison? if the argument is "well that was the real x80, what about the 9x0 series there?), here's a similar calc but accounting for clockspeed.

The 1070Ti here is complete fiction and is where a hypothetical 2240 SP 1070 with everything else unchanged would have been:

[FONT=Courier New] Generation        [/FONT][FONT=Courier New][FONT=Courier New]mhz*cores [/FONT]loss   mhz*cores/$ advantage
1080 vs 1070       27.16%           15.12%
1080 vs 1070Ti     15.02%[/FONT][FONT=Courier New]           34.30%
980 vs 970         21.29%           [/FONT][FONT=Courier New]31.34%
780 vs 770         19.63%[/FONT][FONT=Courier New]           30.73%
680 vs 670[/FONT][FONT=Courier New]         18.95%            [/FONT][FONT=Courier New]1.31%
580 vs 570[/FONT][FONT=Courier New]         11.11%[/FONT][FONT=Courier New]           27.10%
480 vs 470[/FONT][FONT=Courier New]         19.06%[/FONT][FONT=Courier New]           15.73%[/FONT]

So it's not unfair calling the 1070 more cut down or worse value than usual, but the historical trend doesn't seem to be there. It's also still not the whole story of course, memory performance etc matters.