AMD: Polaris 10 targets mainstream desktop & high-end notebook

[Abwx]

If we go on the smart numbers I'd say 75W card draw for the P10 on the left and 40W card draw for the P11 on the right. The translation appears to suggest P10 has a 6-pin now but it'll be seen without one eventually.

Those PCs drain typically 45W at idle (4790K + 4 DIMMs) including the GPU idling, if CPU usage is 10W this yield 23W and 52W for the GFXs at the 12V rail level.

 

[RussianSensation]

Looks like Pitcairn and Cape Verde all over again. At 113W system draw it's much more like the 7850 than the 7870 so I'd expect that to be the cut-down P10.

How would you get ~ 980 level of performance with that? Sorcery?

In games 980 peaks at 184W (reference card!). That's only the GPU alone.

113W system draw implies this card is not even 75W. Not a chance if Polaris 10 is a sub-75W GPU that it will end up as fast as a 184W GTX980.

980 has roughly 2X perf/watt of a 290X.

That means if a 75W Polaris card is as fast as a 184W 980, that means Polaris 10 would have almost 4-4.5X the perf/watt of a 290X. Welcome to 7nm GPUs.

Anyway, I am bummed out that AMD won't have anything exciting until 2017 in the dGPU space for me and NV's cards are crap for crypto-currency to make $ with.

 

[IllogicalGlory]

Why should the 2.5X performance/watt improvement be compared to Hawaii? It's most likely compared to Tonga, which going by the 380X is on par with the 270X in terms of efficiency.

This is a lower end card as well, so efficiency should be high compared to high-end cards like Hawaii.

I'm not necessarily agreeing that a system that draws ~110W total is going to provide us a graphics card with 980 performance, just that the comparison isn't really fair. Hawaii's efficiency is low compared to other GCN cards, GCN cards in the same performance bracket as Polaris.

 

[Abwx]

How would you get ~ 980 level of performance with that? Sorcery?

In games 980 peaks at 184W (reference card!). That's only the GPU alone.



113W system draw implies this card is not even 75W. Not a chance if Polaris 10 is a sub-75W GPU that it will end up as fast as a 184W GTX980.

980 has roughly 2X perf/watt of a 290X.

That means if a 75W Polaris card is as fast as a 184W 980, that means Polaris 10 would have almost 4-4.5X the perf/watt of a 290X. Welcome to 7nm GPUs.

Anyway, I am bummed out that AMD won't have anything exciting until 2017 in the dGPU space for me and NV's cards are crap for crypto-currency to make $ with.

To start with i checked the review of the Extreme gaming 980TI, if frequency is 21% higher than the stock 980TI then the GPU must consume 21% more assuming it s the same voltage for boths cards, here i see 5% difference, wich is totally impossible.

https://www.techpowerup.com/reviews/Gigabyte/GTX_980_Ti_XtremeGaming/21.html

If GPU voltage is the same these are made up numbers.

Second is what can be expected from GF s 14nm LPP, according to their published numbers the improvement in perf/Watt in respect of their own 28nm SLP (wich is better than TSMC 28nm) is 3.5x, to summarize a 2560 SPs Polaris (assuming it s a die shrink of a 290) would clock at 1.4GHz and at 135W TDP, and that s without fully accounting for the lower leakage..

Edit : i corrected the TDP number to 135W.

 

[JDG1980]

113W system draw implies this card is not even 75W.

Not necessarily. It depends on how powerful the CPU is, and how hard it's being pushed. The Polaris 11 demo shown to the press used a i7-4790K CPU according to the slide, but it was locked at 60 FPS so it's safe to assume that neither CPU nor GPU were being pegged at 100%. We have no idea what was used in this latest demo; nothing is stopping them from using a 35W CPU if that made the numbers look better. They are clearly pushing perf/watt in their marketing for Polaris.

Not a chance if Polaris 10 is a sub-75W GPU that it will end up as fast as a 184W GTX980.

Sure it could. Keep in mind that Polaris 10 is about the same die size as Pitcairn, and the Pitcairn-based 7850 - part of AMD's very first wave of 28nm products - was very energy efficient. According to TPU, it only averaged 87W in gaming scenarios.

Polaris is supposed to have about 2.5x the perf/watt of "28nm products". So let's figure out what that would mean if we use Pitcairn as our benchmark. At 1080p, the 7850 has about 37% of the performance of the GTX 980. That was when the 980 first came out; newer games probably show AMD cards in a better light on average. 37 x 2.5 = 92.5% of the GTX 980's performance at 7850 level power consumption. Sounds about right for the cut-down SKU with a focus on perf/watt. Presumably the full Polaris 10 SKU (480X?) will run higher clocks with a focus on performance - no reason it can't go up to 150W; it will still fit on a single 6-pin connector.

 

[Silverforce11]

@JDG1980
And when Pitcairn arrived, the 7870 beat the 6970 easily and came within 5-10% of the prior big-chip flagship, the GTX 580.

Polaris is indeed Pitcairn, small die size, perf/w focus. Does not mean it lacks performance. Everything about it screams "work smarter, not harder". Imagine not rendering at all ~1/3 of the frame due to discarding unseen objects. The implications are huge, and thus there is potential there for a major upset/shock.

 

[JDG1980]

Regarding the original post - when AMD says that Polaris 11 is "targeting the notebook market", I don't think that means there will be no desktop cards. It would be silly not to release desktop cards (which they've already developed and demonstrated!) given the excellent sales of products like the GTX 750 Ti and GTX 950. Rather, I interpret that statement as meaning that Polaris 11 was designed with laptops in mind - focused on perf/watt with a low power target. That's not surprising; even Nvidia's Maxwell (GM107 in particular) was developed as a mobile-first design. I do think that mobile GPUs will end up getting the best bins (as is often the case), and there's also a good possibility that we will see the desktop card only have the cut-down SKU at first, not the full chip (like what happened with Tonga).

 

[casiofx]

Regarding the original post - when AMD says that Polaris 11 is "targeting the notebook market", I don't think that means there will be no desktop cards. It would be silly not to release desktop cards (which they've already developed and demonstrated!) given the excellent sales of products like the GTX 750 Ti and GTX 950. Rather, I interpret that statement as meaning that Polaris 11 was designed with laptops in mind - focused on perf/watt with a low power target. That's not surprising; even Nvidia's Maxwell (GM107 in particular) was developed as a mobile-first design. I do think that mobile GPUs will end up getting the best bins (as is often the case), and there's also a good possibility that we will see the desktop card only have the cut-down SKU at first, not the full chip (like what happened with Tonga).

Polaris is pretty significant GPU for AMD. With more efficient design, the APU combination of Zen+Polaris is pretty awesome.

If Zen can deliver then AMD may even wrestle Apple's market from Intel. Having a single APU with higher performance GPU to accelerate OpenCL apps in Mac will greatly help with performance + smaller motherboard design for macbook Pros.

 

[AtenRa]

How would you get ~ 980 level of performance with that? Sorcery?

In games 980 peaks at 184W (reference card!). That's only the GPU alone.

Well, if the Polaris 10 demo was FPS Caped, then we cannot compare to GTX 980 peak power consumption.

You will need both cards at fps cap or both at peak performance to compare power consumption.

 

[RussianSensation]

Sure it could. .... 37 x 2.5 = 92.5% of the GTX 980's performance at 7850 level power consumption. Sounds about right for the cut-down SKU with a focus on perf/watt. Presumably the full Polaris 10 SKU (480X?) will run higher clocks with a focus on performance - no reason it can't go up to 150W; it will still fit on a single 6-pin connector.

Ya, so you weren't reading my post at all were you? Read again what I said. 7850 is nowhere close to a 75W card either.

Well, if the Polaris 10 demo was FPS Caped, then we cannot compare to GTX 980 peak power consumption.

You will need both cards at fps cap or both at peak performance to compare power consumption.

That's now what I am saying. I am saying Polaris 10 may have a chance to match or beat 980's performance but not a 75W Polaris 10. I am not saying anything about a 150W Polaris 10 not being able to match a 980.

 

[AtenRa]

That's now what I am saying. I am saying Polaris 10 may have a chance to match or beat 980's performance but not a 75W Polaris 10. I am not saying anything about a 150W Polaris 10 not being able to match a 980.

Well im pretty confident Polaris 10 at ~100W could be faster than GTX 980. But in DX-12 games, I wouldnt be surprised if 75W Polaris 10 be faster than GTX 980.

 

[R0H1T]

Well im pretty confident Polaris 10 at ~100W could be faster than GTX 980. But in DX-12 games, I wouldnt be surprised if 75W Polaris 10 be faster than GTX 980.

That;s highly suspect, but I'm guessing it could be plausible in compute heavy games where DX12 punishes the 980, there doesn't seem to be any other way IMO D:

 

[AtenRa]

That;s highly suspect, but I'm guessing it could be plausible in compute heavy games where DX12 punishes the 980, there doesn't seem to be any other way IMO D:

Just take TPUs GTX 980 peak power at ~180W, now add 2x perf/watt on the same architecture (Maxwell) and you get the same performance of GTX 980 at 90W. That will be a GM204 on 16nm give or take.

Since we all know that GCN is faster than Maxwell at DX-12, it will not be a surprise to anyone to see a 75W TDP Polaris be as fast or faster than GTX 980 in DX-12 games.

 

[airfathaaaaa]

Not necessarily. It depends on how powerful the CPU is, and how hard it's being pushed. The Polaris 11 demo shown to the press used a i7-4790K CPU according to the slide, but it was locked at 60 FPS so it's safe to assume that neither CPU nor GPU were being pegged at 100%. We have no idea what was used in this latest demo; nothing is stopping them from using a 35W CPU if that made the numbers look better. They are clearly pushing perf/watt in their marketing for Polaris.



Sure it could. Keep in mind that Polaris 10 is about the same die size as Pitcairn, and the Pitcairn-based 7850 - part of AMD's very first wave of 28nm products - was very energy efficient. According to TPU, it only averaged 87W in gaming scenarios.

Polaris is supposed to have about 2.5x the perf/watt of "28nm products". So let's figure out what that would mean if we use Pitcairn as our benchmark. At 1080p, the 7850 has about 37% of the performance of the GTX 980. That was when the 980 first came out; newer games probably show AMD cards in a better light on average. 37 x 2.5 = 92.5% of the GTX 980's performance at 7850 level power consumption. Sounds about right for the cut-down SKU with a focus on perf/watt. Presumably the full Polaris 10 SKU (480X?) will run higher clocks with a focus on performance - no reason it can't go up to 150W; it will still fit on a single 6-pin connector.

actually they said polaris will be 2.5x better only from the die shrink and we cant account the updates of the uarch itself since we dont know anything about it expect from the power gating

 

[AtenRa]

I believe the 2.5x perf/watt is at the same TDP over 28nm products and not specifically over one product like Hawaii R9 390X.

So example will be a 75W TDP Polaris vs 75W TDP 28nm GCN.
100W Polaris vs 100W 28nm GCN (HD7770)
115W Polaris vs 115W R7 260X

etc etc

 

[Adored]

Well I said just under 980, so take that as 5-10% slower.

I don't see the point in comparing perf/Watt to Hawaii given that none of the new cards will lack compression. For me it'll be Tonga for Polaris based cards, Fiji for Vega or what AtenRa says above makes sense as well.

 

[RussianSensation]

actually they said polaris will be 2.5x better only from the die shrink and we cant account the updates of the uarch itself since we dont know anything about it expect from the power gating

False. AMD themselves estimated 70% comes from a die shrink, 30% from improved architecture. It's hard to have reasonable discussions when people start making things up.

Too many AMD supporters overhype AMD CPUs and GPUs every release. This annoys objective PC gamers who have a more cautious outlook. The former sets up AMD to fail and leads to massive disappointments. This is the same stuff we read online how if AMD cannot make Zen as fast as a 5960X, they should close shop. The irony in that statement is that 5960X has 60-70% higher IPC over Vishera products, not 40%.

Now in this thread we see people wanted to set aside Hawaii as it doesn't fit their expectations for Polaris 10, but 380X Tonga and Hawaii have very similar performance/watt.

https://www.techpowerup.com/reviews/ASUS/R9_380X_Strix/24.html

Also, knowing the past, AMD had never met its perf/watt targets. That means new GPUs may only be 2.5X faster in perf/watt only in some games, not on average.

Neither NV with Maxwell, nor AMD with Nano/Fiji came anywhere close to having cards with 2X perf/watt over their last gen cards, despite their marketing claims. This point seems to evade some people.

It's far more realistic to expect a 110-120W Polaris 10 to match a 980, rather than a 75-90W one since now you start going way outside of realistic range and getting into last 5% probability of it happening. Why would anyone do that and completely set up the chip to fail?

 

[maddie]

Unlike the speculative (if this, then that) posts, we do seem to be going a bit overboard with certain claims.

 

[airfathaaaaa]

False. AMD themselves estimated 70% comes from a die shrink, 30% from improved architecture. It's hard to have reasonable discussions when people start making things up.

Too many AMD supporters overhype AMD CPUs and GPUs every release. This annoys objective PC gamers who have a more cautious outlook. The former sets up AMD to fail and leads to massive disappointments. This is the same stuff we read online how if AMD cannot make Zen as fast as a 5960X, they should close shop. The irony in that statement is that 5960X has 60-70% higher IPC over Vishera products, not 40%.

Now in this thread we see people wanted to set aside Hawaii as it doesn't fit their expectations for Polaris 10, but 380X Tonga and Hawaii have very similar performance/watt.

https://www.techpowerup.com/reviews/ASUS/R9_380X_Strix/24.html

Also, knowing the past, AMD had never met its perf/watt targets. That means new GPUs may only be 2.5X faster in perf/watt only in some games, not on average.

Neither NV with Maxwell, nor AMD with Nano/Fiji came anywhere close to having cards with 2X perf/watt over their last gen cards, despite their marketing claims. This point seems to evade some people.

It's far more realistic to expect a 110-120W Polaris 10 to match a 980, rather than a 75-90W one since now you start going way outside of realistic range and getting into last 5% probability of it happening. Why would anyone do that and completely set up the chip to fail?

the first 2.5x claim came from the glofo and tsmc not amd (later amd jumped on and said that its 2.5x from 28nm and 1.9x from 16nm)
glofo and tsmc have the right to claim it because the mobile soc they created for samsung/apple are showing at the very best a 45-50% increase on perf and at the very least (20 vs 14) a 34 to 44% increase that cant be account only from the uarch

 

[RussianSensation]

Unlike the speculative (if this, then that) posts, we do seem to be going a bit overboard with certain claims.

I even thought Fury X would have 53% more shader/TMU power and 60% more memory bandwidth over a 1Ghz 290X. That means on paper it should have easily outperformed the 980Ti and the Titan X. Almost all the pre-launch leaks showed Fury X beating 980Ti.

After that, I am not going to believe any AMD marketing claims at 100% face value and I won't believe any leaked online chart unless it's an actual video or a game recorded benchmark with GPU-Z and benchmark score side-by-side with proof.

We have guys with Hawaii cards dropping voltage 87-100mV and capping PowerTune to -15/20% getting 1050-1060mhz clocks while using 900-1000W of power in 4X Hawaii chips. You go out and buy a 390 now and not a single one of those will be using 230W of power @ 1050-1060 clocks out of the box.

AMD needs to deliver. Enough with the empty marketing claims and cherry-picked Hitman benchmarks. Besides, Polaris competitors will be GP106/107, possibly 1060Ti GP104.

Easily outperforming 950 in controlled tests doesn't prove Polaris 11 is great because you need the 2nd piece -> how good is GP106/107 in the $129-149 price bracket?

The next generation will last until 2018, and NV will have lower tier Pascal chips replacing 950/960/980/980. That means Polaris 11/10 will face cards that are vastly improved from those Maxwell SKUs.

Trying to compare Polaris to Tonga or Hawaii still doesn't tell us how good NV's Pascal cards will be. Until we know that, it's impossible to make any claims as to how amazing or successful Polaris will be. For example, NV could improve perf/watt by 60-70% from 980 since they will also benefit from FinFET. That means it could be a 0-20% delta between Polaris and Pascal but we don't even know their overclocking headroom, their VRAM amount (4GB vs. 8GB), or their prices.

 

[raghu78]

Too many AMD supporters overhype AMD CPUs and GPUs every release. This annoys objective PC gamers who have a more cautious outlook. The former sets up AMD to fail and leads to massive disappointments. This is the same stuff we read online how if AMD cannot make Zen as fast as a 5960X, they should close shop. The irony in that statement is that 5960X has 60-70% higher IPC over Vishera products, not 40%.

I can say that if AMD want to really succeed then they need to under promise and overdeliver. I don't know if they are doing that with Polaris and Zen. But I hope they do so.

Now in this thread we see people wanted to set aside Hawaii as it doesn't fit their expectations for Polaris 10, but 380X Tonga and Hawaii have very similar performance/watt.

https://www.techpowerup.com/reviews/ASUS/R9_380X_Strix/24.html

Also, knowing the past, AMD had never met its perf/watt targets. That means new GPUs may only be 2.5X faster in perf/watt only in some games, not on average.

RS its all interpretation. You chose to compare Polaris 10 with Hawaii while others compare with Fury X when interpreting the 2.5x perf/watt gains. Fury X was the latest AMD 28nm GPU and I think its reasonable to expect that Polaris 10 is being compared with it. btw I think you should try and reconcile your expectations with Raja's statements that Polaris brings significant architectural enhancements to improve throughput and efficiency. By saying that a full Polaris 10(2560sp) at a likely GPU clock of 1150 Mhz matches R9 390X(2816 sp) at 1050Mhz at 120w you are saying that perf/sp has literally remained static. Don't you think thats absurd.

By that extension since you expect a Polaris 10 to match GTX 980 at 120w a GTX 1080 should beat Polaris 10 by > 75% (stock GTX 980 Ti is 30-35% faster than stock GTX 980 and GTX 1080 is expected to be 35% faster than 980 Ti. So in effect a GTX 1080 is going to be 75-80% faster than Polaris 10.

http://www.hardwarecanucks.com/foru...2-nvidia-gtx-980ti-performance-review-19.html

Do you think AMD are just incapable of major architectural improvements and that only Nvidia can do that?

Neither NV with Maxwell, nor AMD with Nano/Fiji came anywhere close to having cards with 2X perf/watt over their last gen cards, despite their marketing claims. This point seems to evade some people.

It's far more realistic to expect a 110-120W Polaris 10 to match a 980, rather than a 75-90W one since now you start going way outside of realistic range and getting into last 5% probability of it happening. Why would anyone do that and completely set up the chip to fail?

yeah according to you AMD did nothing with 4th gen GCN to improve perf/sp . Well you are doing the same thing that you accuse others of. If they are overhyping you sure are underhyping.

 

[el etro]

To start with i checked the review of the Extreme gaming 980TI, if frequency is 21% higher than the stock 980TI then the GPU must consume 21% more assuming it s the same voltage for boths cards, here i see 5% difference, wich is totally impossible.

https://www.techpowerup.com/reviews/Gigabyte/GTX_980_Ti_XtremeGaming/21.html

If GPU voltage is the same these are made up numbers.

Second is what can be expected from GF s 14nm LPP, according to their published numbers the improvement in perf/Watt in respect of their own 28nm SLP (wich is better than TSMC 28nm) is 3.5x, to summarize a 2560 SPs Polaris (assuming it s a die shrink of a 290) would clock at 1.4GHz and at 135W TDP, and that s without fully accounting for the lower leakage..

Edit : i corrected the TDP number to 135W.

Pretty great observation. Indicates that Gigabyte sent a golden (high ASIC) bin to TPU to review. I conclude this once GM204 cards explode their power consumption once overclocked, as shown with cards reviewed in TPU.
Retail cards not having the same efficiency indicates that is really a very good ASIC bin.

 

[PPB]

False. AMD themselves estimated 70% comes from a die shrink, 30% from improved architecture. It's hard to have reasonable discussions when people start making things up.

Too many AMD supporters overhype AMD CPUs and GPUs every release. This annoys objective PC gamers who have a more cautious outlook. The former sets up AMD to fail and leads to massive disappointments. This is the same stuff we read online how if AMD cannot make Zen as fast as a 5960X, they should close shop. The irony in that statement is that 5960X has 60-70% higher IPC over Vishera products, not 40%.

Now in this thread we see people wanted to set aside Hawaii as it doesn't fit their expectations for Polaris 10, but 380X Tonga and Hawaii have very similar performance/watt.

https://www.techpowerup.com/reviews/ASUS/R9_380X_Strix/24.html

Also, knowing the past, AMD had never met its perf/watt targets. That means new GPUs may only be 2.5X faster in perf/watt only in some games, not on average.

Neither NV with Maxwell, nor AMD with Nano/Fiji came anywhere close to having cards with 2X perf/watt over their last gen cards, despite their marketing claims. This point seems to evade some people.

It's far more realistic to expect a 110-120W Polaris 10 to match a 980, rather than a 75-90W one since now you start going way outside of realistic range and getting into last 5% probability of it happening. Why would anyone do that and completely set up the chip to fail?

Objective pc gamers are too informed to factory fiji, a bottlenecked design with experimental memory interface, in AMD's claims regarding polaris.

And AMD's claims dont mean jack because of that: the whole perf/w and perf/mm2 curve in their current stack is too schizo to give you a clear idea of what those 2 parts are going to perform judging by what we know of polaris (50w and 100-120w gpus, 11x and 232mm2 die size). So you can come and argue that p10 is clearly inferior to gp104 and use tonga/fiji as base (both designs really underperforming in perf/mm2 front) and i can come with hawaii and pitcarin as base and come to totally different conclusions (pitcarin and hawaii being their best perf/mm2 designs which dont have anything to envy to maxwell in that front). Another poster can come too and say polaris will be great using R9nano as base for perf/w, or another dismiss it using fury air as base.

The informed people would look at architectural details that may factor what can be gained at both fronts compared to the current stack, doing some data normalization (taking hawaii's 512b gddr5phy vs fiji's hbm into account for perf/w estimations considering p10 is going 256b gddr5) inbetween.

Arguing about baseless BS like "that would leave too big of a performance gap between p10 and vega" in a market accoustomed to 55% perf delta between 960 and 970,is just that, whishy washy BS with no technical data behind it to back it up.

What we really should consider is Zlatan's comments about polaris and get to the idea that the performance profile of polaris will be totally different to what gcn was until now. There will be scenarios where Polaris will fly and will probably compete head on with GP104 and scenarios where it will behave more like what the shader count would suggest, even after factoring nominal IPC gains that would apply to 90% scenarios.

Sent from my XT1040 using Tapatalk

 

[tajoh111]

False. AMD themselves estimated 70% comes from a die shrink, 30% from improved architecture. It's hard to have reasonable discussions when people start making things up.

Too many AMD supporters overhype AMD CPUs and GPUs every release. This annoys objective PC gamers who have a more cautious outlook. The former sets up AMD to fail and leads to massive disappointments. This is the same stuff we read online how if AMD cannot make Zen as fast as a 5960X, they should close shop. The irony in that statement is that 5960X has 60-70% higher IPC over Vishera products, not 40%.

Now in this thread we see people wanted to set aside Hawaii as it doesn't fit their expectations for Polaris 10, but 380X Tonga and Hawaii have very similar performance/watt.

https://www.techpowerup.com/reviews/ASUS/R9_380X_Strix/24.html

Also, knowing the past, AMD had never met its perf/watt targets. That means new GPUs may only be 2.5X faster in perf/watt only in some games, not on average.

Neither NV with Maxwell, nor AMD with Nano/Fiji came anywhere close to having cards with 2X perf/watt over their last gen cards, despite their marketing claims. This point seems to evade some people.

It's far more realistic to expect a 110-120W Polaris 10 to match a 980, rather than a 75-90W one since now you start going way outside of realistic range and getting into last 5% probability of it happening. Why would anyone do that and completely set up the chip to fail?

AMD has updated their projections to 2x times performance per watt and no longer 2.5x. And this is over their mainstream products and not stuff like nano or fury. So if AMD makes the same improvements as Nvidia since they both claim 2x increase in performance per watt, they are going to be in the same position as now.

 

[Abwx]

AMD has updated their projections to 2x times performance per watt and no longer 2.5x.

This mean that they decided to dump some perf/Watt to get higher perfs and still be competitive enough for their taste, i guess that they got some early infos about their competition actual perf/Watt..

For instance if their perf/Watt is 20% better for a given segment than their competitor they can increase frequency (and hence perfs) by 10% and still be at the same level as said competitor in perf/Watt...

And this is over their mainstream products and not stuff like nano or fury. So if AMD makes the same improvements as Nvidia since they both claim 2x increase in performance per watt, they are going to be in the same position as now.

That s not true because perf/Watt is dependent of the absolute perf, we dont know where they did set the bar, but according to the first part of your post they did set it at comfortable absolute perf improvement...