Polaris 10 benchmarks...

[SelenaGomez]

Videocardz released 3dmark benchs of Polaris 10. They show that its slightly better score than a 980 and even crossfired it is slower than a 1080. Why would AMD release a card like that 1.5 years after they released a 980 and right after the release of the 1080?? What exactly are they thinking?

 

[.vodka]

Too early to tell from those scores. We don't know clocks, we don't know anything yet. CF scaling is clearly broken, especially on a synthetic benchmark like 3dmark. What we do know is that Polaris targets the mainstream market (x70/x80 tier) where the vast majority of money is made. Vega targets the high end (x90/Fury). Polaris doesn't compete with GP104. Yet it could (or not) get dangerously close to the 1070 as the 480/x. That leaves the 490/x tier for Vega later in the year.

Different market segments, one offering doesn't compete with the other.

In other words, NDA lifts on June 29 following AMD's schedule for Polaris that they have outlined early in the year. Everything is going according to the known plan.Talking about videocardz, there will be an update on Polaris in less than a day, so we'll know more today.

 

[SelenaGomez]

I have never bought an AMD card and I have nothign against AMD either. I hope the Polaris is faster than a 1070 and less money than a 1070. We need more competition to bring the prices down.

 

[RussianSensation]

I have never bought an AMD card and I have nothign against AMD either. I hope the Polaris is faster than a 1070 and less money than a 1070. We need more competition to bring the prices down.

That's impossible. 232mm2 chip with rumoured 1266mhz GPU clocks was never designed to beat a 314mm2 chip with 1683mhz GPU boost clocks, and also you want it to cost less? Think of x70/x80 cards as GTX670/680 while Vega 10 is HD7950/7970. That makes Polaris 10 HD7850/7870 level cards that competed with GTX660 (aka GTX1060/1060Ti GP106). Polaris 10 will also cost substantially less than the $379-449 prices of 1070. Another way to look at it is GP104 is a 1440p 60Hz card while Polaris 10 is a 1080p 60Hz card. Both make perfect sense for their respective market segments.

 

[trane]

Looks like you have not been paying attention. AMD have said time and time again that they are targeting the mainstream and performance segments first. Translation - $300 and down, where >80% of the market lies (in units). Their goal is to regain market share they lost last generation due to the 20nm debacle. So they will ship the mainstream chips first, with high-end later. Nvidia is going the opposite route releasing their high-end first, mainstream later. The end result is Nvidia is unchallenged in the high-end space, while AMD will be unchallenged in the mid-range space.

Don't worry, AMD haven't given up on high-end, there's Vega coming for that. Similarly, Nvidia haven't given up on the <$300 market, they have GP106 and GP107 coming (and possibly a further cut down GP104).

 

[Bacon1]

That's impossible. 232mm2 chip with rumoured 1266mhz GPU clocks was never designed to beat a 314mm2 chip with 1683mhz GPU boost clocks, and also you want it to cost less? Think of x70/x80 cards as GTX670/680 while Vega 10 is HD7950/7970. That makes Polaris 10 HD7850/7870 level cards that competed with GTX660 (aka GTX1060/1060Ti GP106). Polaris 10 will also cost substantially less than the $379-449 prices of 1070. Another way to look at it is GP104 is a 1440p 60Hz card while Polaris 10 is a 1080p 60Hz card. Both make perfect sense for their respective market segments.

If AMD doesn't release a card thats 50% faster than Titan X for $150 I'm never buying from them again. I'd rather pay Nvidia $10,000 for a 25% faster than Titan X card!

:thumbsdown:

AMD is going to be bringing high end GPUs to mainstream, this has been their goal forever and what they've talked about. They want everyone to be able to have a VR ready computer, not only the select 7.5m gamers with 290/970+ performance card @ $300+ entry price.

http://wccftech.com/amd-polaris-mainstream-vr-market/

https://youtu.be/p010lp5uLQA?t=933

 

[SelenaGomez]

Looks like you have not been paying attention. AMD have said time and time again that they are targeting the mainstream and performance segments first. Translation - $300 and down, where >80% of the market lies (in units). Their goal is to regain market share they lost last generation due to the 20nm debacle. So they will ship the mainstream chips first, with high-end later. Nvidia is going the opposite route releasing their high-end first, mainstream later. The end result is Nvidia is unchallenged in the high-end space, while AMD will be unchallenged in the mid-range space.

Don't worry, AMD haven't given up on high-end, there's Vega coming for that. Similarly, Nvidia haven't given up on the <$300 market, they have GP106 and GP107 coming (and possibly a further cut down GP104).

They always say this... Yet NVidia dominates them in every way. IF AMD truly targeted ther mainstream >80% they would be far bigger and far more profitable than they are. Their Vega better be faster than the 1080ti or Titan since they will probabl be released together.

 

[SelenaGomez]

If AMD doesn't release a card thats 50% faster than Titan X for $150 I'm never buying from them again. I'd rather pay Nvidia $10,000 for a 25% faster than Titan X card!

:thumbsdown:

AMD is going to be bringing high end GPUs to mainstream, this has been their goal forever and what they've talked about. They want everyone to be able to have a VR ready computer, not only the select 7.5m gamers with 290/970+ performance card @ $300+ entry price.

http://wccftech.com/amd-polaris-mainstream-vr-market/

https://youtu.be/p010lp5uLQA?t=933

Why be so sarcastic and why exaggerate so much? Nvidia peopel arent paying $10000 for card and why cant AMD release a fast card for cheaper? They are the vastly smaller company and seem to be playing catch up quite often. Let be real Nvidia dominates AMD for a reason. AMD has messed up too many times and have had the same strategy for a decades now. I want AMD to pick it up. I am not a fanboy at all. I want whatever is the best bang for the buck. But AMD pisses me off

 

[Bacon1]

Why be so sarcastic and why exaggerate so much? Nvidia peopel arent paying $10000 for card and why cant AMD release a fast card for cheaper? They are the vastly smaller company and seem to be playing catch up quite often. Let be real Nvidia dominates AMD for a reason. AMD has messed up too many times and have had the same strategy for a decades now. I want AMD to pick it up. I am not a fanboy at all. I want whatever is the best bang for the buck. But AMD pisses me off

Because AMD has constantly had better price/perf cards compared to Nvidia since Hawaii first came out. Hasn't done them much good as far as market share goes and people constantly want cheaper and faster or no buy. How the hell is the company supposed to grow if they have to constantly provide cheaper parts that are faster? Nvidia loves their massive profits per sale compared to AMD.

 

[SelenaGomez]

Because AMD has constantly had better price/perf cards compared to Nvidia since Hawaii first came out. Hasn't done them much good as far as market share goes and people constantly want cheaper and faster or no buy. How the hell is the company supposed to grow if they have to constantly provide cheaper parts that are faster? Nvidia loves their massive profits per sale compared to AMD.

So why is Nvidia far bigger if AMD has constantly had better price/performance cards?

 

[showb1z]

Why be so sarcastic and why exaggerate so much? Nvidia peopel arent paying $10000 for card and why cant AMD release a fast card for cheaper? They are the vastly smaller company and seem to be playing catch up quite often. Let be real Nvidia dominates AMD for a reason. AMD has messed up too many times and have had the same strategy for a decades now. I want AMD to pick it up. I am not a fanboy at all. I want whatever is the best bang for the buck. But AMD pisses me off

That's exactly what they've been doing for years? Other than the top-end name one segment where AMD doesn't deliver better price/performance.

So why is Nvidia far bigger if AMD has constantly had better price/performance cards?

Mostly better marketing. And some bad execution at key moments, like releasing Hawaii with terrible reference coolers which dragged down the 200 series for its entire lifespan.

 

[slavovid]

That's impossible. 232mm2 chip with rumoured 1266mhz GPU clocks was never designed to beat a 314mm2 chip with 1683mhz GPU boost clocks, and also you want it to cost less? Think of x70/x80 cards as GTX670/680 while Vega 10 is HD7950/7970. That makes Polaris 10 HD7850/7870 level cards that competed with GTX660 (aka GTX1060/1060Ti GP106). Polaris 10 will also cost substantially less than the $379-449 prices of 1070. Another way to look at it is GP104 is a 1440p 60Hz card while Polaris 10 is a 1080p 60Hz card. Both make perfect sense for their respective market segments.

However isn't 1070 consisting of just 75% of the 314mm2 die in terms of performance & Cuda Cores. Well it keeps the size but is a cut down version.
That maybe makes it easier to cool down etc. But still Polaris 10 should easily be head to head to 1070 or even above it considering the slightly better density between the 16nm FinFET of the 314mm2 chip vs the 14nm FinFET
As we know from the apple chips the difference in the 2 nodes provides about 9% size difference . And when you add the fact that that comparison is between the 16nm FinFET Nvidia is using vs the 14nm FinFET LPE and not the 14nm FinFET LPP that AMD is using for Polaris then the difference between a cut down 314mm2 chip vs a full 232mm2 P10 might be in huge favor for Polaris

314mm2 cut down version is relative power of a 235mm2 chip.
232mm2 chip adding 9% for density could equal a 252mm2 chip made at 16nm FinFET
If we add 10% for the LPE to LPP difference (samsung states 15% difference not in density but in performance) then it jumps to possible relative comparison to equal a 277mm2 16nm and that is way more than what 1070 is.

I am completely unaware if i can use such comparisons in reality but that is what i am reading the past month and what makes most sense to me

Meaning a P10 can and should beat 1070 unless AMD goes for much less power efficiency vs performance. While it is a 232mm2 chip and it should cost less than the 1070 (even before considering how much overpriced the 1070 is)

 

[showb1z]

Yep, that all makes sense to me. And this is why AMD better gives some solid info on Computex even if they're not launching until later.
1080 doesn't matter, it's another price bracket and people buying it now probably would've bought the Nvidia card anyway. But they can't let 1070 launch without at least a ballpark performance target on Polaris or they will lose a chunk of potential customers for no reason.

 

[richaron]

However isn't 1070 consisting of just 75% of the 314mm2 die in terms of performance & Cuda Cores. Well it keeps the size but is a cut down version.
That maybe makes it easier to cool down etc. But still Polaris 10 should easily be head to head to 1070 or even above it considering the slightly better density between the 16nm FinFET of the 314mm2 chip vs the 14nm FinFET
As we know from the apple chips the difference in the 2 nodes provides about 9% size difference . And when you add the fact that that comparison is between the 16nm FinFET Nvidia is using vs the 14nm FinFET LPE and not the 14nm FinFET LPP that AMD is using for Polaris then the difference between a cut down 314mm2 chip vs a full 232mm2 P10 might be in huge favor for Polaris

314mm2 cut down version is relative power of a 235mm2 chip.
232mm2 chip adding 9% for density could equal a 252mm2 chip made at 16nm FinFET
If we add 10% for the LPE to LPP difference (samsung states 15% difference not in density but in performance) then it jumps to possible relative comparison to equal a 277mm2 16nm and that is way more than what 1070 is.

I am completely unaware if i can use such comparisons in reality but that is what i am reading the past month and what makes most sense to me

Meaning a P10 can and should beat 1070 unless AMD goes for much less power efficiency vs performance. While it is a 232mm2 chip and it should cost less than the 1070 (even before considering how much overpriced the 1070 is)

The calculations you have done are really only relevant to see if Polaris can compete per clock. Which is extremely likely.

You've already talked about the higher density a few times, which will affect the chips thermally. And there are plenty of unknown arch' differences, with the nV chip being designed for high clocks.

So the real unknown is how much slower the Polaris chips will be clocked, and how far from peak efficiency AMD will push them in an attempt to compete.

 

[Erenhardt]

That's exactly what they've been doing for years? Other than the top-end name one segment where AMD doesn't deliver better price/performance.

It clearly is an issue of "I want AMD to deliver good products for cheap to force NV to lower prices so I can have my Geforce GPU cheaper."

People want nvidia cards. They want them cheap. They need amd to force nv to drop prices. Every generation it is harder to make nv drop prices, since they see the sales numbers going up regardless of worse perf/$ than their competition.

As to the benchmarks. AMD keeps the seal on Polaris tight.

 

[Head1985]

However isn't 1070 consisting of just 75% of the 314mm2 die in terms of performance & Cuda Cores. Well it keeps the size but is a cut down version.
That maybe makes it easier to cool down etc. But still Polaris 10 should easily be head to head to 1070 or even above it considering the slightly better density between the 16nm FinFET of the 314mm2 chip vs the 14nm FinFET
As we know from the apple chips the difference in the 2 nodes provides about 9% size difference . And when you add the fact that that comparison is between the 16nm FinFET Nvidia is using vs the 14nm FinFET LPE and not the 14nm FinFET LPP that AMD is using for Polaris then the difference between a cut down 314mm2 chip vs a full 232mm2 P10 might be in huge favor for Polaris

314mm2 cut down version is relative power of a 235mm2 chip.
232mm2 chip adding 9% for density could equal a 252mm2 chip made at 16nm FinFET
If we add 10% for the LPE to LPP difference (samsung states 15% difference not in density but in performance) then it jumps to possible relative comparison to equal a 277mm2 16nm and that is way more than what 1070 is.

I am completely unaware if i can use such comparisons in reality but that is what i am reading the past month and what makes most sense to me

Meaning a P10 can and should beat 1070 unless AMD goes for much less power efficiency vs performance. While it is a 232mm2 chip and it should cost less than the 1070 (even before considering how much overpriced the 1070 is)

Yeah 1070 is just 1060TI.Polaris 10 should compete with it.Or be within 10% of 1070.
GTX660TI was 9% faster than 7870.
GTX660Ti is today 1070 and 7870 will be polaris 10.

But 14nm is better than 16nm and pascal is just maxwell at 16nm + polaris will have many improvements.
It all depends on clocks.1266Mhz is super low it should be at 1400Mhz minimum.

 

[slavovid]

I was pointing out my thoughts in favor of the fact that Polaris 10 should be able to compete and even beat the cut down that is 1070. Considering the node differences and without mentioning any possible gains from GCN 4 architecture upgrades.

So the statement that it would be impossible for Polaris 10 to beat 1070 seems a bit off. While possible however it depends on how well AMD did with their architecture changes and much more.
It is better for the community that Polaris 10 full beats pascal cut down version as that will mean a lot of competition on the market and that is better for us in all aspects.
I am optimistic to think that since we saw that Hitman Demo so long ago AMD did not sit idle all this time and meanwhile did polish their product.

 

[gamervivek]

The calculations you have done are really only relevant to see if Polaris can compete per clock. Which is extremely likely.

You've already talked about the higher density a few times, which will affect the chips thermally. And there are plenty of unknown arch' differences, with the nV chip being designed for high clocks.

So the real unknown is how much slower the Polaris chips will be clocked, and how far from peak efficiency AMD will push them in an attempt to compete.

Indeed. The 1.27Ghz clocked Polaris 10 chip is better but not even close to enough. 1.5Ghz is the minimum that AMD need in order to compete unless they are going really overboard on density. Fiji had ~11% more transistors than GM200, Polaris will likely double that advantage at least.

 

[slavovid]

Indeed. The 1.27Ghz clocked Polaris 10 chip is better but not even close to enough. 1.5Ghz is the minimum that AMD need in order to compete unless they are going really overboard on density. Fiji had ~11% more transistors than GM200, Polaris will likely double that advantage at least.

Actually considering 1920 CC and 1683 clock speed on the 1070 then a 2304 CU on a P10 will need a clock speed of ~ 1400 to give us the same FLOPS rate.
Considering 16nm should provide similar or even less clocks jump than 14nm LPE and then the 14nm LPP is stated to provide even better power efficiency and clock rates than it's predecessor. Then even 1500 mhz clock shouldn't be hard to reach and that will put Polaris 10 on even higher Flops rate than 1070

Then we are left with the GCN4 changes to push up AMD's FLOPS performance closer to that of Nvidia.

 

[Riek]

Actually considering 1920 CC and 1683 clock speed on the 1070 then a 2304 CU on a P10 will need a clock speed of ~ 1400 to give us the same FLOPS rate.
Considering 16nm should provide similar or even less clocks jump than 14nm LPE and then the 14nm LPP is stated to provide even better power efficiency and clock rates than it's predecessor. Then even 1500 mhz clock shouldn't be hard to reach and that will put Polaris 10 on even higher Flops rate than 1070

Then we are left with the GCN4 changes to push up AMD's FLOPS performance closer to that of Nvidia.

1070 in the reviews is boosting to 1.9Ghz not 1.6xx. (thats 7.3 versus 6.4 according to nvidia marketing slides).

 

[richaron]

Actually considering 1920 CC and 1683 clock speed on the 1070 then a 2304 CU on a P10 will need a clock speed of ~ 1400 to give us the same FLOPS rate.
Considering 16nm should provide similar or even less clocks jump than 14nm LPE and then the 14nm LPP is stated to provide even better power efficiency and clock rates than it's predecessor. Then even 1500 mhz clock shouldn't be hard to reach and that will put Polaris 10 on even higher Flops rate than 1070

Then we are left with the GCN4 changes to push up AMD's FLOPS performance closer to that of Nvidia.

Sorry but your thinking is fundamentally flawed. Or more precisely it's way too simplistic, even though it may turn out to be close to the truth. Your posts build up hype and expectations but you methodology is just wrong. And I find if odd that a new poster, who obviously has covered all the arguments, would come in here to spread these expectations.

The main reason your calculations are wrong is simply the frequency vs IPC argument. There are deliberate trade-offs in designing a chip, and your calculations are assuming both companies made the same decisions. You are making claims like "1500 mhz clock shouldn't be hard to reach" but this is based on nVidia's architecture; not AMD's.

We simply don't know enough to come to the conclusions you have, but as i said, you might be right. AMD could also release Polaris 10 with 900mhz clocks and super IPC and blow away a 1080.

 

[slavovid]

Sorry but your thinking is fundamentally flawed. Or more precisely it's way too simplistic, even though it may turn out to be close to the truth. Your posts build up hype and expectations but you methodology is just wrong. And I find if odd that a new poster, who obviously has covered all the arguments, would come in here to spread these expectations.

The main reason your calculations are wrong is simply the frequency vs IPC argument. There are deliberate trade-offs in designing a chip, and your calculations are assuming both companies made the same decisions. You are making claims like "1500 mhz clock shouldn't be hard to reach" but this is based on nVidia's architecture; not AMD's.

We simply don't know enough to come to the conclusions you have, but as i said, you might be right. AMD could also release Polaris 10 with 900mhz clocks and super IPC and blow away a 1080.

My thoughts about 1500 not being impossible is based on the clocks jump from Nvidia's previous generation to this generation clock jumps and have in mind AMD's previous generation too and considering the even slightly better node jump.
That said i said that 1400 is enough to push the same FLOPS compared to 1070's base clock how much can each overclock is a different subject.

I also stated right there in the end that "Then we are left with the GCN4 changes to push up AMD's FLOPS performance closer to that of Nvidia."
And that alone shows that i am completely aware that there is difference in the architectures and it will require the exact same architecture for the above calculations to be that simple.

They are just that, simple calculations to show that 1070 is not out of the reach of Polaris 10 full. But still could be really far from the unknown reality 😀

And yes i am not only a fresh signup but also quite uneducated on hardware level. All this is based on the amounts of data i have managed to read on the subject in the past 1-2 months (subject being pascal and polaris)

 

[crisium]

If full P10 is not within 15% of GTX 1070 it will be a failure in my book. I'd like it to get even closer or match if possible, but I'm giving leeway.

My reasons are not technical, but simple desires based on observations.

660 Ti vs 7870, as already covered.

Nvidia's new 1920 Shader card beats its old 3072 Shader card: 62.5% of the shaders, less bandwidth, 67% of the ROPs and still faster.

This SP ratio means a 2304SP P10 should beat the Fury Air, and frankly since Fiji is fundamentally flawed, it should have enough leeway to tie or beat the Fury X. At 1440p DX11, Fury X is within 15% of 1070.

If Pascal is mostly Maxwell on steroids, and since AMD keeps tagging "New" all over its pictures of Polaris architecture, then how the heck can AMD not at least match Nvidia's gains?

Also, using the Nvidia improvement math, a die shrunk Hawaii would easily beat the Fury X. Would AMD really engineer all this "new" just to be worse than a theoretical die shrunk Hawaii? They'd be better of shrinking Hawaii, adding H265, HDMI 2.0, etc. then bothering with all that money and time on developing a slower Polaris 10.

I really hope I'm not disappointed.

 

[Sonikku]

So why is Nvidia far bigger if AMD has constantly had better price/performance cards?

Better marketing. They sell the idea that you're extra cool for putting a Nvidia brand gpu in your PC. So if you don't care about price/performance ratio you'll be an extra hip nerd for going with them instead of AMD. :thumbsup:

 

[tential]

I shouldn't be so happy about the ops post but I am. The op is the exact reason I have predicted Polaris to fall flat on its face. Targeting price/perf means that people actually are doing hard calculations. People rarely do that kind of math when making a purchase.

If the 1070 is 380 and p10 is 300 it won't matter if p10 is the better bang for your dollar. If the 1070 wins the general masses will buy a 1070. They are not going to read a review suite take the averages of games and realize that p10 is within x% of the 1070 but the 1070 costs y% more. I've seen like a small minority of the population.

This strategy amd has proposed works for the most analytical of people. Which is not what gamers use in purchasing decisions sorry. I mean we still pre-order as gamers....

But personally I don't care if the masses don't understand how to do math to figure out which card is the best bang for buck. Too many buy poor choice Nvidia cards, which leads to early price cuts for amd, which means even better bang for buck cards for me to use with freesync. So meh, I can't complain. I may pick up a stopgap used $300 nano if possible but probably not seeing the number in circulation.