AMD Cayman info (or rumor)

[Manicdan]

im REALLY running out of reasons not to upgrade. just need a 6850 with a nice cooler, like the gigabyte SOC and it will be all over.

[eXceeded]

Surprised noones posted this yet:
http://translate.google.com/translat...2F1953715.html

Forum also heard today from chiphell.com a new spy photos, I believe that AMD's HD 6970 the future for graphics cards (Cayman XT HD 6870 core), this graphics card from the previously reported data, can be said that today's single- card performance of the king.

Radeon HD 6970 (Cayman XT) 3DMark Vantage test results floating around in the X12000, but P mode is P24499. 在Unigine Heaven (1920 1200 4AA+16AF)环境设置下运行速度为36.6帧。 In Unigine Heaven (1920 1200 4AA +16 AF) environmental setting speed was 36.6 frames.

HD 5870 3DMark Vantage获得P19337的成绩，在Unigine Heaven (1920 1200 4AA+16AF)运行帧数为17.3 。 HD 5870 3DMark Vantage P19337 results obtained in Unigine Heaven (1920 1200 4AA +16 AF) running 17.3 frames. 同平台下GTX 480 3DMark Vantage获得P21106和X9309的成绩，在Unigine Heaven (1920 1600 4AA+16AF）速度为29.5。 The same platform GTX 480 3DMark Vantage P21106 and X9309 results obtained in Unigine Heaven (1920 1600 4AA +16 AF) speed is 29.5.

毫无疑问Radeon HD 6970(Cayman XT)将会是单卡之王 ，在相同状况下超出GTX 480性能10%以上，从测试中来说具备压制性的实力。 There is no doubt Radeon HD 6970 (Cayman XT) will be the king of a single card in the same circumstances beyond the GTX 480 performance above 10%, from tests run with oppressive power.

不过这张卡依旧是内部样卡，板型类似5870，核心频率在800MHz－900MHz范围内，供电接口为6 +8，相信TDP应该在255W或者以上。 However, the card is still within the sample card, similar to plate 5870, the core at 800MHz-900MHz frequency range, power supply interface is 6 +8, I believe should be 255W TDP or more.

We love Engrish! But looks like (if these are reliable in anyway) that we are in for some nice Christmas presents. I'm still going to wait for Antilles, my machine is soo ancient now!

[hurrdurr]

Looks like it will be slower than 5970 in DX9-DX10 (according to Vantage data), but the Unigine results show an incredible jump in tessellation intensive applications. The rumored improvements for DX11 were not for the 6870 but for 6970, after all.

BTW anyone have any idea how they'll keep within the 300W limit for Antilles? 6870 showed nearly no performance / watt increase. With a larger GPU (like Cayman) I'm not sure they can reasonably increase the P/W without a node shrink. Yeah, redesign and all, but their nice redesign for Barts made no improvement in P/W over Cypress and I doubt Cayman would bring that kind of improvement.

The reason I am mentioning this is the fact that 6870 consumes 150W. Makes me think 6950 will be around 180W and 6970 will be no less than 210-220W. How can they produce a dual card from that, keeping within the 300W limit? They would lower voltages and clocks, obviously, but the bottom line is that when performance / watt stays the same, if you stay within 300W you can't produce a card much faster than 5970 which was 300W itself. And if you are going to downclock / trim the card so much that it'll go down from 420W to 300W, you'd better use something that would be 300W to begin with, which brings me to....

... Barts showed a significant increase in Crossfire scaling (around 10%). Keeping in mind that a single Barts is between a 5850 and 5870 in performance, and that one 5970 GPU was also between 5850 and 5870: If you slap two Barts cores together, which would be something like 280W due to single PCB, and increase clocks slightly to make it 300W; because of the increased clocks and better Crossfire scaling, you'd get a card that would be around 15% faster than the 5970.

So, if 6970 performs on par with the 5970; then 6990 would be 15% faster than 6970, which would be in line with the naming. 5970 was 40% faster than the 5870 and thus it was named 59xx instead of 58xx. But if they do what I said, the performance difference would be smaller, like 15%, which would be in line with the 6990 being in the same "family" as 6970, i.e. 69xx.

HELLO BY THE WAY

[Xoulz]

Welcome hurrdurr,

Antilles is not a dual card, but a dual GPU...

[hurrdurr]

Welcome hurrdurr,

Antilles is not a dual card, but a dual GPU...

Dual GPU? As in dual-core CPU's, and not like HD5970?

[EvilOne]

If 6870 is 150w, wouldnt a 6890(6870x2) be the sweet spot for a 300w card? With the shared components, should even be a little room for a little factory OC and still be under 300w.

[Xoulz]

Dual GPU? As in dual-core CPU's, and not like HD5970?

NO.. just 2 gpu's on one card.

[Macadamia]

^
Last month's rumor. Leaked pics are real but the benches have been *speculated* to be fake (Heaven at least, the text was not bolded and all)

Given that the leaks might not be true, I still believe that Cayman's gonna do something special with Tessellation though. Maybe not reaching Fermi rate but way enough for all 3D games.

[Manicdan]

the 5970 is 2x5870s but with 5850 clocks. which should put it around 180W each, but managed to keep it under 300W

it shouldnt be that hard to do something similar where they build it downclocked, and leave plenty of room to OC it, if you can handle the noise.

[hurrdurr]

the 5970 is 2x5870s but with 5850 clocks. which should put it around 180W each, but managed to keep it under 300W

it shouldnt be that hard to do something similar where they build it downclocked, and leave plenty of room to OC it, if you can handle the noise.

As long as performance / watt doesn't increase (which it hasn't in Barts) you can't get more performance at 300 watts. The only thing that would be increasing performance would be CF scalability, which did increase about 10% in Barts CF situations.

BTW a single 5850 is 150W whereas a 5870 is 180W. Two 5870's in 5850 clocks and voltage would be around 160W; which would make 320W, but because one PCB's worth of energy is saved it's 300W.

There is no way AMD is putting two Cayman's in the Antilles unless:

1. Performance/watt has increased incredibly, so incredibly that a 6950 consumes the same power as a 6870 while performing about 25% better.

OR,

2. They won't be caring about the 300W limit.

[Manicdan]

As long as performance / watt doesn't increase (which it hasn't in Barts) you can't get more performance at 300 watts. The only thing that would be increasing performance would be CF scalability, which did increase about 10% in Barts CF situations.

BTW a single 5850 is 150W whereas a 5870 is 180W. Two 5870's in 5850 clocks and voltage would be around 160W; which would make 320W, but because one PCB's worth of energy is saved it's 300W.

There is no way AMD is putting two Cayman's in the Antilles unless:

1. Performance/watt has increased incredibly, so incredibly that a 6950 consumes the same power as a 6870 while performing about 25% better.

OR,

2. They won't be caring about the 300W limit.

perf per watt has always favored lower clocked cards because you can reduce voltages too. mobile gpus can see 30+% better perf/watt because they run so low.

even if a 6970 is at 900mhz, the 6990 can be 600mhz if they need to be. and you want it to be so you can just turn it up to 900mhz when u get one

[hurrdurr]

even if a 6970 is at 900mhz, the 6990 can be 600mhz if they need to be. and you want it to be so you can just turn it up to 900mhz when u get one

Definitely. But such a low clock would make the stock card's performance suffer incredibly (to the point that it's only slightly faster than the single card), and on paper you would be buying a $700 card that's not really faster than a $400 card.

Sites do reviews on stock cards and do not really focus hugely on overclocking results. And I doubt they would do any differently for such a card. They will never be saying "Well we upped the voltage by X volts and the clocks by Y MHz (which would also depend from site to site and would clog the internet with hugely varying results) and conducted the review with the overclocked card, because we really do expect all consumers to up the voltages and the clock".

[hurrdurr]

BTW, the same Chinese site reports the TDP of 6970 at 255W. Ouch.

[marten_larsson]

For Antilles I was thinking maybe AMD is going the 5830 x2 route, use some cut down version of Cayman with a bit lower clocks to keep it <300W. This would also make the assumed 6990 name more reasonable IMO.

[Manicdan]

Definitely. But such a low clock would make the stock card's performance suffer incredibly (to the point that it's only slightly faster than the single card), and on paper you would be buying a $700 card that's not really faster than a $400 card.

Sites do reviews on stock cards and do not really focus hugely on overclocking results. And I doubt they would do any differently for such a card. They will never be saying "Well we upped the voltage by X volts and the clocks by Y MHz (which would also depend from site to site and would clog the internet with hugely varying results) and conducted the review with the overclocked card, because we really do expect all consumers to up the voltages and the clock".

that was an extreme case idea. i honestly do not expect the 6990 to be more than 70% faster than a 6950, and slower than a 6950xfire or 6970xfire, but still that would probably mean 40% faster than a 6970.

check out the powerpoint slides for the 5970, it was pushed out as an overclockers card. the stock cooling was rated for 400W, even though it only used 300W to start. such things are not really for average joe, but for the people who just want bragging rights on having the "best"

[iMacmatician]

BTW, the same Chinese site reports the TDP of 6970 at 255W. Ouch.

Wow. Although given the reports of >225 W for Cayman XT I can't say I'm surprised.

255 W is also higher than the (claimed) TDP of GTX 480.

[hurrdurr]

For Antilles I was thinking maybe AMD is going the 5830 x2 route, use some cut down version of Cayman with a bit lower clocks to keep it <300W. This would also make the assumed 6990 name more reasonable IMO.

Exactly what I wrote. If AMD is thinking of staying within 300W I think the route that would make the most sense would be to take two Barts cores. It would be 10-15% faster than 5970 thanks to the impressive CF scaling. And the naming would make sense too (6970 to 6990 would be a small increase in performance and price)

[Xoulz]

IMO.. the cayman has more width. So the speed doesn't effect throughput as much, thus 2 Cayman can be under-clocked, with less loss of performance than barts.

[Lanek]

Definitely. But such a low clock would make the stock card's performance suffer incredibly (to the point that it's only slightly faster than the single card), and on paper you would be buying a $700 card that's not really faster than a $400 card.

Sites do reviews on stock cards and do not really focus hugely on overclocking results. And I doubt they would do any differently for such a card. They will never be saying "Well we upped the voltage by X volts and the clocks by Y MHz (which would also depend from site to site and would clog the internet with hugely varying results) and conducted the review with the overclocked card, because we really do expect all consumers to up the voltages and the clock".

The crossfire scaling with 6870 is really far better of what is it with the 5xxx, it's more in the 90-95% range, so even if the clocks are reduced, they can surely maintain an overall performances of +60% vs a single 6970
(ofc if the 69xx scale as the 6870 )...
A second possibility will be to use 2 completes core of 6950 ( so with less amount of SP vs the 6970 ).... the 5970 was clocked like the 5850, but they was using 2x 1600SP ( vs 1440 for the 5850 )...

A 5870 is running with 180W, if the 6950 is in the same range, it's completely possible Antilles will come under the 300W...

( edited: look like my mind was thinking to the 6950 and writed 6850 = )

ofc this is only pure speculation ...

[SKYMTL]

Wow. Although given the reports of >225 W for Cayman XT I can't say I'm surprised.

255 W is also higher than the (claimed) TDP of GTX 480.

TDP does not mean max power consumption.


Remember, AMD lists Max Board Power while NVIDIA uses TDP. Those are two completely different metrics.


MBP = specification which an add-in partner needs to specify their board to handle if they want to mount a specific chip to it. It is very rare that a program will stress a GPU + mem enough to even come close to MBP. It can also be used as the absolute peak power consumption a given board design is meant to handle. Meanwhile, it is possible for GPU + mem power consumption to EXCEED TDP.

Hence why you don't see linear scaling between AMD's MBP figure and NVIDIA's TDP in more reviews' power conumption charts.

[blindbox]

Exactly what I wrote. If AMD is thinking of staying within 300W I think the route that would make the most sense would be to take two Barts cores. It would be 10-15% faster than 5970 thanks to the impressive CF scaling. And the naming would make sense too (6970 to 6990 would be a small increase in performance and price)

If we're to trust AMD's slides and 300W figures, it HAS to be a Cayman chip. Slides indicate that Antilles will have higher perf than 5970. Which Cayman chip and what clocks, we do not know.


For all we know, Barts XT might have some shader missing, saved for Antilles (oh yeah I speculate too! Everyone's doing it!)

[Calmatory]

TDP does not mean max power consumption.


Remember, AMD lists Max Board Power while NVIDIA uses TDP. Those are two completely different metrics.


MBP = specification which an add-in partner needs to specify their board to handle if they want to mount a specific chip to it. It is very rare that a program will stress a GPU + mem enough to even come close to MBP. It can also be used as the absolute peak power consumption a given board design is meant to handle. Meanwhile, it is possible for GPU + mem power consumption to EXCEED TDP.

Hence why you don't see linear scaling between AMD's MBP figure and NVIDIA's TDP in more reviews' power conumption charts.

And to add to this, TDP is related to the cooling solution only. TDP is the value the cooling solution has to be able to dissipate without the chip reaching it's maximum junction temperature. In other words, if a GPU has TDP of 130 W, the actual card can consume even 300 W in some cases.
1.) as the TDP is related to the cooling solution, other components such as VRM, VRAM, board may consume whatever they please and as long as the cooling solution isn't designed to cool them, it does not affect TDP.
2.) Under some workloads(e.g. Furmark), the chip may dissipate much more heat than the cooling solution is designed to handle, as in most cases TDP does not represent the maximum theoretical power draw of a given chip, but rather the typical power consumption under real world workloads for which the chip has been designed for.

Hence why TDP itself isn't good estimate of whole card's power consumption, and why MBP isn't good estimate of only GPU power consumption.

[SKYMTL]

And to add to this, TDP is related to the cooling solution only. TDP is the value the cooling solution has to be able to dissipate without the chip reaching it's maximum junction temperature. In other words, if a GPU has TDP of 130 W, the actual card can consume even 300 W in some cases.
1.) as the TDP is related to the cooling solution, other components such as VRM, VRAM, board may consume whatever they please and as long as the cooling solution isn't designed to cool them, it does not affect TDP.
2.) Under some workloads(e.g. Furmark), the chip may dissipate much more heat than the cooling solution is designed to handle, as in most cases TDP does not represent the maximum theoretical power draw of a given chip, but rather the typical power consumption under real world workloads for which the chip has been designed for.

Hence why TDP itself isn't good estimate of whole card's power consumption, and why MBP isn't good estimate of only GPU power consumption.

Exactly. In addition, neither takes into account the efficiency (or lack thereof) of the VRMs, etc which can have a siginificant impact upon overall power consumption of a given design.

[hurrdurr]

Also I believe Techpowerup had a review (of GTX 480 or 470?) that compared two exactly same clocked Fermi's, one with better custom cooling and one with stock cooling. The one with better custom cooling (and temperatures) consumed a lot less power, in spite of the fact that the two GPU's were identical.

[Manicdan]

Also I believe Techpowerup had a review (of GTX 480 or 470?) that compared two exactly same clocked Fermi's, one with better custom cooling and one with stock cooling. The one with better custom cooling (and temperatures) consumed a lot less power, in spite of the fact that the two GPU's were identical.

yup, even though your new, your definitely familiar

it was able to drop like 20-30W of power by removing 30C from the temps. we were all quite shocked by how much it helped.