Polaris: 2x Perf/Watt of Current AMD mainstream GPUs

[exar333]

I hope it doesn't become 1.5x performance/watt by the time Polaris actually launches. Knowing AMD's history, it is actually quite possible.

Could just be that they are upping the clocks to add more performance and the perf/watt could be impacted somewhat. Just a thought...that is possible if AMD is trying to squeeze some extra headroom out of Polaris.

 

[JDG1980]

It seems that despite AMD's first showing of Polaris and rumors that Nvidia wasn't doing well in yields with Pascal, so far Pascal to Polaris looks just like last generation.

I would rather not jump to a definitive conclusion with no data.

 

[JDG1980]

2x perf/watt over "mainstream" 28nm GCN chips could correspond to roughly Hawaii levels of performance at roughly Pitcairn levels of power consumption. For Polaris 10, that lines up with about what I'd expect. As I said, though, I'd rather wait for actual data rather than reading too much into AMD's marketing materials. Honestly, I think we're analyzing these PowerPoint slides with much more attention to detail than was ever put into creating them in the first place.

For all the people worrying about competitiveness versus Nvidia... GTX 1080 has the best perf/watt of any current card, but that's because all the competitors are currently 28nm. It actually isn't as big a jump as you might expect for a node shrink. According to TPU's chart, GTX 980 is 64% as efficient as GTX 1080 at 1440p. Flip that around and this means GP104 has 1.56x the perf/watt of GM204. Even if AMD does "only" 2x, that would still be a bigger jump, and would result in perf/watt on par with Pascal.

 

[maddie]

18k views meh. Nvidia gonna be fine. You think people are going to stop founding the 1080 because of some small confusion? No, they'll defend it. I'd give you evidence but it's so abundantly clear and I don't want to put any members in an awkward spot. For supporting the founders edition.
Everyone wants to found something, and if you want to found the 1080 God willing found away. Without founders how would a card ever even be produced?

Like I've said, expect Nvidia to dynamically price better. Did you think Uber, airbnb, etc. Pricing is coming out of thin air. Dynamic pricing is the most efficient way of capturing value all along the demand curve instead of at a fixed price point.

Amd, they need a marketing department founders card edition for vega. "please help us fund a real marketing department!" might work with some fans.....

Don't. That video link might have ruffled some feathers.

 

[R0H1T]

Lisa Su: Card
RTG: Chip?

I mean, power usage is more than the chip. You need to take board power into consideration, and that includes memory chips.

It seems that despite AMD's first showing of Polaris and rumors that Nvidia wasn't doing well in yields with Pascal, so far Pascal to Polaris looks just like last generation.

Great point, I did mention this previously & depending on the type of memory used the card's perf/w can range anywhere from 2~3x when compared against a number of skus, especially for sub 200W cards.

For instance the Fury X HBM draws less than half the amount of power that GDDR5 does on 290x ~ http://www.anandtech.com/show/9969/jedec-publishes-hbm2-specification

GDDR5x itself consumes two thirds the amount of an equivalent GDDR5 bandwidth system ~ http://www.anandtech.com/show/10193/micron-begins-to-sample-gddr5x-memory

Then you have to take into account the pcb size, memory controller, VRM et al & the gaming resolution itself.

So none of the previous claims are scandalous as the OP was trying to highlight!

 

[Ma_Deuce]

I hope it doesn't become 1.5x performance/watt by the time Polaris actually launches. Knowing AMD's history, it is actually quite possible.

If they say you are getting 2X, you better expect 1.5X because you are only actually going to get 1.25X :sneaky:

 

[3DVagabond]

I think AMD already answered that question with their demo of small Polaris.

They did not demoed only chip but whole card.

It was very low clocks though. @ 850MHz it might be 2.5X perf/W, but at 1.27GHz it might only be 2X?

 

[3DVagabond]

No it would not be faster than GTX 1080.

First of all a 2x improvement in perf/W would not be 50 GFLOPS/W, it would only be about 42 GFLOPS/W, since a 380X only does 21 GFLOPS/W (3973.1 GFLOPS, 190W TDP).

Secondly GCN generally has worse utilization than Maxwell. For instance the 390X and 980 are more or less similar in gaming performance, but the 390 has almost 20% higher processing power (5914 GFLOPS vs. 4981 GFLOPS).

So a 200W Polaris with a 2x improvement in perf/W over 380X, would be 8400 GFLOPS, which gaming performance wise would be equal to about 7100 GFLOPS for a Maxwell (and presumably the same for a Pascal card). The 1080 is roughly 9100 GFLOPS or about 28% more.

So all in all this hypothetical 200W Polaris GPU would be about equal to a Titan X, which isn't too bad honestly, especially if the price is good. By extension, a 150W Polaris would be about equal to a 980, and a 100W polaris about equal to a 380X.

If AMD manages to hit a 2.5x improvement instead of 2x, then they will basically be neck and neck with Pascal, efficiency wise. So fingers crossed I guess.

I think a lot of the "inefficiency" is due to DX11. There are dedicated compute units (ACE's for example) that go completely unused. There's also the DSP for TrueAudio that don't get used (nothing to do with DX11 of course). That's supposed to change with VR. Then there's the whole serial execution with DX11 that doesn't run as well on GCN.

 

[zentan]

Even 2X is not bad at all if dx12 remains favorable as it is touted to be. That can easily change the equation in its favor but we have to see.
From Polaris 11 end all they need is to snatch the mid-range gaming notebook designs 800-1100$ range which generally went to gtx860m and 960m. That alone would do a lot of good considering their presence was very limited there since the times of their VLIW GPUs. Polaris 10 might compete with gp106 SKUs or may be even higher/lower,who knows at this point? But will Polaris 11 have some challenge at the low end desktop SKUs unless of course either GP106 can compete with the same power-performance envelope or else if nvidia would have gp107 ready to compete by then ? That will be interesting to watch. It could be as popular as gtx 750/ti series came to be.

 

[antihelten]

I'm thinking that is one of the targeted improvements with all of the new Polaris IP blocks. New command processor, shader cores, primitive discard unit, etc.

We can't use the past except to use it as the absolute performance floor. Your perf/W comparison will be the lowest. A 150W P10 could be around 6500 GFOPS in the old AMD tradition but with better utilization than in the past, equal, or dare I say, exceed the same 6500 GFLOPS 1070. it depends on how much the efficiency has improved.

Having said that, I can't see a 1080 match in any realistic scenario. 1070 however, could be fair game.

I'm also wondering if AMD is waiting out Nvidia for the 1070 tests to go live before displaying P10 details. This will be the short term battleground between the two of them. No other overlap should exist for now.

It's worth noting here that when AMD said 2-2.5x perf/W improvements, they probably meant gaming performance per watt and not pure processing power per watt. As such improvements in utilization would be included in the 2-2.5x number, and thus already accounted for in my numbers. Whether processing power improves by 2x or it improves by 1.6x and utilization by 1.25x doesn't really matter here, since you still get the same end results, 2x improvement in gaming performance.

It is however also worth mentioning that a 2x improvement in processing power combined with a 1.25x improvement in utilization, would be equal to a 2.5x improvement in gaming performance. So it might be that the RTG have been quoting gaming performance numbers, and the heads of AMD have been quoting pure processing power numbers, thus explaining the difference in the numbers quoted (2.5x vs. 2x).

I think a lot of the "inefficiency" is due to DX11. There are dedicated compute units (ACE's for example) that go completely unused. There's also the DSP for TrueAudio that don't get used (nothing to do with DX11 of course). That's supposed to change with VR. Then there's the whole serial execution with DX11 that doesn't run as well on GCN.

DX12 should help improve the utilization of GCN, especially with the usage of async compute. Furthermore this is almost certainly not included in their 2-2.5x number, since they would be comparing Polaris versus GCN 3 in like for like scenarios, i.e. dx12 vs. dx12 and dx11 vs. dx11 (at least I would hope that this is what they do).

But this improvement in utilization would benefit both older generations of GCN (except perhaps gen 1, given it's lack of ACEs), and the new GCN 4, and as such won't necessarily lead to a jump in efficiency (perf/W) for GCN 4 over older GCN.

Also higher utilization from stuff like async compute probably also leads to higher power usage (you are lighting up more parts of the GPU after all), so it probably won't lead to significant efficiency gains compared to Maxwell/Pascal either. It could however lead to absolute performance gains versus Maxwell/Pascal, which might be even more important, seeing as Polaris 10 targets the mainstream segment and GP104 the performance segment. Thus the gap in performance could be smaller than expected, which is arguably more important than Polaris 10 beating GP104 in efficiency (depends on the target market of course).

 

[cusideabelincoln]

And the disappointing news starts to slowly set it in. Is perf/w the be end all metric? No. Is perf/w telling of the architecture vs. the competition? Yes.

If Polaris is 2x perf/w vs. Tonga, then AMD will be in the same situation that it was in at the start of Kepler and at the start of Maxwell - losing in perf/w, perf/mm2, and perf/transistor. They will once again be relegated to low margin prices just to stay competitive and will still be facing an uphill battle since no new parts released this year are faster than their previous best and they are releasing behind Nvidia.

In non-DX12 games this might be true.

 

[airfathaaaaa]

Mainstream is always 199-249$ bracket.

So R9 380 and 380X.

We are looking at 50 GFLOPs/Watt for Polaris.

mainstream TAM as per amd is up to 390/x(mainstream vr anyway since this is what they are targeting)
anything lower cant achieve their goal to bring the current gen down a scale and call it "mainstream"
its something we know for quite a long time i dont know why people keep changing its meaning

 

[Cookie Monster]

Also I want to add another perspective into this discussion, that when you increase utilization of a said GPU, the power consumption will obviously be higher as parts of the GPU that is idle are now doing work. (In return for increased performance)

edit - it'd actually be interesting to see perf/watt figures for GCN under DX11 and DX12 titles (or the same game under different DX paths).

 

[Flapdrol1337]

Performance and powerconsumption should go up, but efficiency should go up as well. Idle hardware is pure waste.

With cards that always clock up to their max tdp it could be dramatically more efficient, in games where the hardware is underused they will just push the clock and voltage for little gain. If the hardware is kept busy it might be forced to run at a more sensible and efficient speed.

 

[nenforcer]

From my understanding Polaris 11 at ~50W is going to replace the Radeon R7 370 2/4GB (R7 270) in the sub $200 category as the Radeon R7 470.

I'm not sure which card the Polaris 10 is intended to replace - either the Radeon R9 290/390 as the Radeon R9 490 8GB GDDR5X and then the Radeon R9 480 8GB GDDR5?

The Radeon R7 370 is already off of the market while the Radeon R9 290/390 "Hawaii/Grenada" is nearly 3 years old now. The Radeon R9 285 / 380 "Tonga/Antigua" is only slighly newer at nearly 2 years old.

If the Pascal GTX 1070 comes in at roughly $380 - $400 and outperforms Radeon Fury / Nano then we might see a price drop of $100 or more on those cards - down to $350-400 unless AMD is stubborn.

This puts Polaris 10 in the midrange once again where it is intended to probably be in the $200-300 range with the Radeon R9 390 slotted just above it at $300-350.

I'm also curious if AMD is going to put in Crossfire limits on these cards just like nVidia did limiting GTX 1070 / 1080 to Dual SLI only.

 

[Glo.]

It was very low clocks though. @ 850MHz it might be 2.5X perf/W, but at 1.27GHz it might only be 2X?

It would have no sense. 420 MHz higher core clock would bring much higher compute power compared to 850 MHz GPU.

The problem is power consumption. To get to 2 times higher perf/watt it would need with that core clock increase by 100W. It makes absolutely no sense.


Nothing so far from my calculations for every CU configuration between 3072 GCN cores, and 2048 GCNC makes sense, with 2 times higher perf/watt. Only 2.5 times higher - that makes sense with respect to what we know about core config and thermals.

 

[Glo.]

The picture in the slide is either Nano or Fury x. I know that you may consider mainstream 380 380x due to price but remember AMD now considers fury as mainstream from their latest road map.

It is Fury X. And it has 31 GFLOPs/Watt of compute power. Increasing it by 2x brings it to 60GFLOPs/Watt.

3072 GCN core GPU with 125W and 1.25 GHz core clock brings 61GFLOPs/Watt.

It is also around 2.5 mark from other GPUs.

 

[William Gaatjes]

I would not be surprised if polaris is launched with settings that have low power consumption at a lower voltage/lower clock. I think just like the fury nano, it will be possible to be set to a higher use of power consumption / higher voltage/ higher clock speed. That is, if i read the reviews about the fury nano correctly.
Remember, the fury nano can reach 1GHz but usually maxes out on 750MHz because of the power settings.

Leaving the choice with the user.

I am guessing here, but i think it makes sense that polaris will have that ability too. An automated boost clock mechanism where the user can set the maximum boost clock to a by AMD defined maximum setting of a power profile option. Having a slider or something. If true, then the interesting part will be how high that by AMD defined maximum setting will be.

 

[railven]

Following the Zen thread over at CPU Sub-section, it seems AMD is making a few different claims. They're changing based on what they are comparing it too.

Is GloFo also making Zen for AMD? Wonder if GloFo is prepping to drop the ball.

 

[airfathaaaaa]

Lisa Su: Card
RTG: Chip?

I mean, power usage is more than the chip. You need to take board power into consideration, and that includes memory chips.

It seems that despite AMD's first showing of Polaris and rumors that Nvidia wasn't doing well in yields with Pascal, so far Pascal to Polaris looks just like last generation.

so by saying launching on mid 2016 since december and actually doing it is translated into "worse yields" ?

 

[zlejedi]

2x perf/watt over "mainstream" 28nm GCN chips could correspond to roughly Hawaii levels of performance at roughly Pitcairn levels of power consumption. For Polaris 10, that lines up with about what I'd expect. As I said, though, I'd rather wait for actual data rather than reading too much into AMD's marketing materials. Honestly, I think we're analyzing these PowerPoint slides with much more attention to detail than was ever put into creating them in the first place.

For all the people worrying about competitiveness versus Nvidia... GTX 1080 has the best perf/watt of any current card, but that's because all the competitors are currently 28nm. It actually isn't as big a jump as you might expect for a node shrink. According to TPU's chart, GTX 980 is 64% as efficient as GTX 1080 at 1440p. Flip that around and this means GP104 has 1.56x the perf/watt of GM204. Even if AMD does "only" 2x, that would still be a bigger jump, and would result in perf/watt on par with Pascal.

What's the point of comparing percentage gains when starting level is so different ?

GTX 1080 might be "only" 1,56x GTX980 but that was most power efficient chip of previous generation. Same chart in tpu database shows 390 at 36% of gtx 1080 so AMD would need to make nearly 3x increase to match it.

 

[Glo.]

What's the point of comparing percentage gains when starting level is so different ?

GTX 1080 might be "only" 1,56x GTX980 but that was most power efficient chip of previous generation. Same chart in tpu database shows 390 at 36% of gtx 1080 so AMD would need to make nearly 3x increase to match it.

If you look at gaming benchmarks then yes. Not when you look at compute performance per watt. Even Fury X was more efficient than GTX 980 in that factor. Most power efficient GPU of previous gen. was Fury Nano.

 

[Paul98]

The only card on here that is similar to what AMD considers "Mainstream" is the R9 270x.

 

[airfathaaaaa]

again i will just leave this here............

do you see what amd means ?

 

[Flapdrol1337]

do you see what amd means ?

I do not, as the image doesn't show.