What will be AMD'S next Move?

[VirtualLarry]

970/980 are clocked at about 1200Mhz on average, do the maths.

I stand by the perf/watt number, if perf/watt was intrinsicaly better, that is due to architectural superiority, then it would have the same max throughput for half the power comsumption but there s only 17% difference at peak throughput, what you fail to understand is that its efficency is not at a fixed value, it varies hugely with the GPU loading, it s not like an efficency gained from a node shrink where the power comsumption curve would be simply translated 30% lower, in this case the GPU more or less manage to gate off its non functional parts at very high speed within the flow of datas but if the flow is sustained and is close to max throughput then the GPU has to let all the parts supplied to process the bottlenecked datas in the waiting, comsumption will then reach what it is supposed to be at this node level.

Thg used the word compression as an analogy, what is compressed actualy is the size of the GPU in function of the computation needs, in a game this will translate in gated off unities when the scene is not demanding, and games are rarely, if ever, getting close to max throughput.

This chip adressed a few of the remaining thing to improve in Hawai s pipelines but it cant be considered next gen.

This makes sense, thanks. Good post.

So a full distributed-computing load, would basically revert Maxwell to being quite a bit more inefficient, due to all computational units being utilized, and not being allowed to power-gate them.

 

[parvadomus]

Tonga was not a major improvement in perf/w, I never trust TH for any hardware review, its not time to start using them as a reliable source.

Tonga's Perf/W

Tonga's average power use in gaming load

970 average power use in gaming load

Tonga is not better than Tahiti in perf/w, in some cases, worse. That isn't from one site either, most of the reviews found the same thing for Tonga & Maxwell. TH is the only one I've seen that seems to think R285 uses massively less power than the R280/7970, which we know to be false.

ps. I don't care that GPU compute puts Maxwell load higher, its not relevant to me, nor would it be if you show Mining power load for radeons and claim its absurd. Gaming load matters for gaming cards.

You can use TPUs power consumption all you want, but it shows tonga in its worst possbile way (comparing to other reviews). First they use a Nvidia game to get power consumption. Its a Physx game, how do we know if some of the physics are handled by GPU? TPU power consumption is BS from my pov.

I can also cherry pick the best possible power consumption for the 285, isolated from CPU consumption:
http://ht4u.net/reviews/2014/sapphire_r9_285_itx_compact-edition_im_test/index45.php

The best way to get power consumption is to isolate it from the rest of the system, if you dont do it you get nothing. Maybe AMD drivers are garbage and they are a CPU cycles hog, and the difference in power consumption comes from the CPU not the GPU. We will know when DX12 arrives, and CPUs stop being a bottleneck. Or maybe physx games affect power consumption, or maybe AMD cards sucks at power consumption (but ironically some use 6+6 pin and still consumes the same as NVs 6+8pin ones as TPU show )

 

[JDG1980]

The power comparison charts run by Tom's Hardware in the Maxwell review are garbage. You can't start using a new, completely different method of measuring power draw and then put those numbers alongside numbers derived from a different method and expect the chart to mean anything.

What matters is the card's DC power draw over the PCI bus and PCI-E connectors over relevant time frames (i.e. not microseconds). If tests run using this method show power draw over rated TDP during GPGPU, then it's worth noting. But no one else has demonstrated any such thing. I think that the reviewer for Tom's Hardware simply did not know how to properly handle the raw oscilloscope data.

For what it's worth, this blog contains some tests run on cryptocurrency mining on a GTX 980. At stock speeds, it stayed within TDP, despite the 100% GPGPU load. I suppose it's possible that the sensors polled by GPU-Z could be lying about the power draw, but I doubt it.

 

[Abwx]

so all the reviews about GTX 970 per watt is false?

No, technicaly they are about all right, what they fail to mention or notice is that the 970 perf/watt is not constant in respect of throughput, at low throughput it start very efficiently and as the throughput increase efficency converge to its intrinsical value, about the same as the 780/780TI.

The efficency curves of thoses two latter is different, it start at mediocre value and slowly converge to its intrinsical value as throughput increase up to its max value.

The efficency in function of throughput curve of the former will look like a square root function that would be stopped at (1,1) while the latter would be a linear function, in principle the efficency difference maxima occur at about 50% of the thoughput, wich is about the throughput one can expect in games, hence the high efficency measured by most sites.

This makes sense, thanks. Good post.

So a full distributed-computing load, would basically revert Maxwell to being quite a bit more inefficient, due to all computational units being utilized, and not being allowed to power-gate them.

Yes, i explained above why their scheme make good sense for games but is totaly useless in task that manage to use close to max throughput, the BOINC crowd has nothing to await from these cards other than getting used 780/780ti at dirt prices in the used marked....

 

[positivedoppler]

The power comparison charts run by Tom's Hardware in the Maxwell review are garbage. You can't start using a new, completely different method of measuring power draw and then put those numbers alongside numbers derived from a different method and expect the chart to mean anything.

What matters is the card's DC power draw over the PCI bus and PCI-E connectors over relevant time frames (i.e. not microseconds). If tests run using this method show power draw over rated TDP during GPGPU, then it's worth noting. But no one else has demonstrated any such thing. I think that the reviewer for Tom's Hardware simply did not know how to properly handle the raw oscilloscope data.

For what it's worth, this blog contains some tests run on cryptocurrency mining on a GTX 980. At stock speeds, it stayed within TDP, despite the 100% GPGPU load. I suppose it's possible that the sensors polled by GPU-Z could be lying about the power draw, but I doubt it.

Why is it garbage? His 970 and 980 numbers are in line with everybody else

 

[Gloomy]

They used the same for the 290/290X, you can go back and check. The readings were done more coarsely, though. The only difference between this and the 290/X readings is that they have more data for the 980. The final readings don't change no matter how often/less often you sample, though, because it's an average. (EDIT: I mean, they change, but they have enough samples for both for a decent enough average)

 

[el etro]

Cameron, i'm not responding to you on this, ok?

Nvidia exclusive features = AMD exclusive features: there's no better one here;

API Support: Maxwell is ahead, but it come first in the next-gen battle. If Dx 11.3 games launch in a short period of time then will be a advantage, if them don't come until GCN2 everything will be tied;

Display Interface Support: Maxwell is future-proof, but to mainstream users some of this features will no mean much;

Video Codec Support: Comparing tonga versus Maxwell here, Maxwell is sightly ahead. Not really much a generational step ahead here;

Efficiency: Very much ahead here;

Performance on gaming: Ahead, but ahead of a one year old GPU;

Performance on compute: Not ahead;

Memory: Both are already on GDDR5 limits, there's no much to improve here.

Memory Bandwidth: It seems Nvidia is ahead here. To make comparison between Tonga and Maxwell memory techs we would need to benchmark efficiency of compression techs(amount of compressed data and perceptual of image quality lost by compression)from both GPU vendors.

 

[garagisti]

GTX 970 is very, very good. As good as for me to want to get one even. Then again, all this is getting a bit too silly, with suggestions that AMD may go out of business etc. All that noise about the stock coolers still persists, while aftermarket solutions are more silent than the useless noise on forums everywhere. Pretty sure AMD isn't going to stop making graphics card. I mean they were the first to the market with a card with support for DX10 proper, same for DX11 and will support full features of another DX in their new card. If HD2900 didn't sink ATI/ AMD, surely they will be fine.

It may be launch of their APU based firepro's potentially on the 25th. Well, that is my guess.

 

[96Firebird]

No, technicaly they are about all right, what they fail to mention or notice is that the 970 perf/watt is not constant in respect of throughput, at low throughput it start very efficiently and as the throughput increase efficency converge to its intrinsical value, about the same as the 780/780TI.

The efficency curves of thoses two latter is different, it start at mediocre value and slowly converge to its intrinsical value as throughput increase up to its max value.

The efficency in function of throughput curve of the former will look like a square root function that would be stopped at (1,1) while the latter would be a linear function, in principle the efficency difference maxima occur at about 50% of the thoughput, wich is about the throughput one can expect in games, hence the high efficency measured by most sites.

Maybe I missed it, how did you come to this conclusion?

 

[Silverforce11]

Maybe I missed it, how did you come to this conclusion?

He's basing it on compute workloads, which he knows is not relevant for gamers.

 

[96Firebird]

I've read through Tom's review, where do they mention what programs they use for their GPGPU Power Consumption "torture" tests?

 

[Silverforce11]

You can use TPUs power consumption all you want, but it shows tonga in its worst possbile way (comparing to other reviews).

Uses more power than R280X.
http://hardocp.com/article/2014/09/02/msi_radeon_r9_285_gaming_oc_video_card_review/10#.VB5T1PmSyfU

Uses similar power to 7970.
http://www.guru3d.com/articles_pages/amd_radeon_r9_285_review,5.html

Tonga is not more efficient than Tahiti. Many sites show this, sites that are much more reputable than Toms.

No one here should be under the delusion that Toms is the go-to for hardware review accuracy. They haven't been for many years now.

If you want to directly compare Tonga to Maxwell, you can be generous and say both use similar power in gaming (which isn't so, Tonga uses a bit more), but then compare the performance. GCN 1.1.1 is not a competitor to Maxwell, there's no denying it.

AMD's true response will be GCN 2 and who knows when thats out. Could be a long time.

 

[JDG1980]

Why is it garbage? His 970 and 980 numbers are in line with everybody else

No, they aren't. Tom's Hardware claims that the reference GTX 980 draws 285 watts under a "torture test". Everyone else reports figures near the rated TDP (e.g. TechPowerUp's review cites a maximum power consumption of 190 watts, which is close enough to the TDP that slight measurement errors could account for it). The figures given by Tom's Hardware are way off from everyone else, and they're essentially accusing Nvidia of lying. I think it's more likely that the reviewer at Tom's simply doesn't know or understand what they're doing with their new toy.

 

[Abwx]

Maybe I missed it, how did you come to this conclusion?

I used the exemple of the square root curve because it allow easy understanding, the real curve could be rather a logarithmic one but only accurate measurements could find its actual shape, you would have to measure comsumption for a significant numbers of given fixed throughputs to get its precise shape.

As to how i managed to get to this conclusion is simply by checking and reading THG article, it is clear that at max throughput the card efficency is basicaly the same as 780 series, so that it s not architectural improvements that are at work (moreover when comparing the throughputs in function of frequency), the improvement in perf/watt is hence restricted to low/average throughput situations, that is to cases were the power comsumption is intrinsicaly low/average, this is typical of power gating behaviour, you get improvements only when all ressources are not needed.

 

[el etro]

Tonga is not more efficient than Tahiti.

At shows same power to both Tonga and Tahiti:

At is one of the best reviews in the net, very impartial IMO.

TPU i don't trust more, they do everything they can make in their reach to make Nvidia cards look better....

 

[desprado]

At shows same power to both Tonga and Tahiti:

At is one of the best reviews in the net, very impartial IMO.

TPU i don't trust more, they do everything they can make in their reach to make Nvidia cards look better....

I also dont trust other though.Guru3d and Anandtech are the best for reviews.

 

[Silverforce11]

At shows same power to both Tonga and Tahiti:

At is one of the best reviews in the net, very impartial IMO.

TPU i don't trust more, they do everything they can make in their reach to make Nvidia cards look better....

Exactly, R285 performs like an R280 in that bench (Crysis 3), the power is extremely close also.

Thus, it's NOT more efficient than Tahiti.

The evolution of GCN from 1.0, to 1.2 did not focus on efficiency, only features & a smaller bus to make PCBs cheaper.

 

[positivedoppler]

No, they aren't. Tom's Hardware claims that the reference GTX 980 draws 285 watts under a "torture test". Everyone else reports figures near the rated TDP (e.g. TechPowerUp's review cites a maximum power consumption of 190 watts, which is close enough to the TDP that slight measurement errors could account for it). The figures given by Tom's Hardware are way off from everyone else, and they're essentially accusing Nvidia of lying. I think it's more likely that the reviewer at Tom's simply doesn't know or understand what they're doing with their new toy.

I assume the torture test means hashing or folding in which the gpu is kept at 100% utilization. This effectively nullifies nvidia's highly refined gpu boost and amd's less steller powertune. There is already a thread that talks about it i think.

 

[AtenRa]

No, they aren't. Tom's Hardware claims that the reference GTX 980 draws 285 watts under a "torture test". Everyone else reports figures near the rated TDP (e.g. TechPowerUp's review cites a maximum power consumption of 190 watts, which is close enough to the TDP that slight measurement errors could account for it). The figures given by Tom's Hardware are way off from everyone else, and they're essentially accusing Nvidia of lying. I think it's more likely that the reviewer at Tom's simply doesn't know or understand what they're doing with their new toy.

Not everyone test the same software or using the same apparatus.

from TPU

Maximum: Furmark Stability Test at 1280x1024, 0xAA. This results in a very high no-game power-consumption that can typically be reached only with stress-testing applications. We report the highest single reading after a short startup period. Initial bursts during startup are not included, as they are too short to be relevant.

 

[AtenRa]

This is a claim that every so often is mentioned here.

I assume you are talking about desktop cards, and Nvidia making 85-90% of their income on sub $300 cards.
Sometime ppl will mention $200 or $250 mark. You are going with safer number, props for that, but its still a number pulled out of a thin air.

I suppose there was much truth to that claim in olden times when there was no SoCs, with huge OEMs shipments and only select few rocking high-end gaming setups.
But also I'm pretty sure this is not the case today, where high-end gaming is a viable money generator. Matter of fact this is how NV differentiates itself from PC slumber.
During last 2-3 years there were no other catalyst for them, so it HAS TO BE the rise of high-margin PC gaming which insulates them from PC downfall.


Funny fact: Nvidia's most popular chip on Steam is GK104, not GK106/7/8.
And this is going by pure chip VOLUME, and not by profit from your statement, which is several times bigger on more expensive cards.

Without you providing any data i could also say you are wrong. Just because we are at 2010 GPU volume, doesnt mean High-End $300+ GPU volume is more than 10-15% today. It was at 10% in 2010, lets say it is 15% today, well AMD can make profit without even having a GPU in that segment.
As for margins, just to remind you all that Tonga is smaller than GM204, one is sold at $249 with 2GB Memory and the other at $329 with 4GB Memory. So margins could be very well be the same. And i bet $249 GPU volumes are higher than $329 segment.

 

[sontin]

And why do you compare Tonga with GM204 when you can buy a GTX760 for $40 less than a r9 285?

And there are rumors that nVidia will introduce another cut down version of GM204. It's clear that Tonga is DoA.

 

[DominionSeraph]

well AMD can make profit without even having a GPU in that segment.
As for margins, just to remind you all that Tonga is smaller than GM204, one is sold at $249 with 2GB Memory and the other at $329 with 4GB Memory. So margins could be very well be the same. And i bet $249 GPU volumes are higher than $329 segment.

They'll only have volume there if customers are stupid enough to buy a card that's worth under $200 in a $329 GTX 970 world at $249.

And why do you compare Tonga with GM204 when you can buy a GTX760 for $40 less than a r9 285?

$56 less, actually.
http://www.newegg.com/Product/Produc...D=3938566&SID=

 

[AtenRa]

And why do you compare Tonga with GM204 when you can buy a GTX760 for $40 less than a r9 285?

The comparison was to show that Tonga has the same or higher margins and higher volume. Also Tonga is faster than GTX760 and slower than GTX970, the price is between the two.

And there are rumors that nVidia will introduce another cut down version of GM204. It's clear that Tonga is DoA.

So a 400mm2 die at even lower price than $329 will be fine for NVIDIA but Tonga that is smaller will be DoA. ???

They'll only have volume there if customers are stupid enough to buy a card that's worth under $200 in a $329 GTX 970 world at $249.

Currently R9 285 is faster than $200 GTX760 and slower than $329 GTX970. This is the real world price and performance metric as of Today.

$56 less, actually.
http://www.newegg.com/Product/Produc...D=3938566&SID=

Lowest GTX760 at newegg = $189,99 AR

Lowest R9 285 at new egg = $239,99

That is $50 when you count AR price, but you also get Game bundle with R9 285.

 

[jj109]

This makes sense, thanks. Good post.

So a full distributed-computing load, would basically revert Maxwell to being quite a bit more inefficient, due to all computational units being utilized, and not being allowed to power-gate them.

If the shaders are power gated and doing no work, then utilizing them will also increase throughput, unless the schedulers aren't optimal and the shaders just spin their wheels. Are we looking at low power usage (low power) or throughput/watt (efficiency)?

If we're interested in the second one, then the questions to answer are "what did Tom's run?" and "what was the throughput?" Without knowing those two, then trying to talk about efficiency is a waste of time.

 

[DominionSeraph]

The comparison was to show that Tonga has the same or higher margins and higher volume.

I don't see a single "R9 285 owner's thread" here. OTOH we seem to have a rather active one on the for the 970/980.
If the 285 couldn't get buzz at $250 when 780 Ti performance cost $600, what makes you think it's getting it now when you can get 780 Ti performance at $330?

Currently R9 285 is faster than $200 GTX760 and slower than $329 GTX970. This is the real world price and performance metric as of Today.

Hey folks, remember:
- If the 5770 is $180.
- The 5850 is $310
- And the GTX 460 is released at $200.
... everything is appropriately priced and there's no way this can start a price war because the GTX 460 wasn't released UNDER the 5770's price! AMD will surely continue to sell all the 5770's and 5850's they can!