AMD: Sea Islands R1100 (8*** series) Speculation/ Rumour Thread

[lanek]

Thx! So Turbo won't be dependant on the cooling solution employed, right?
What catched my attention from Anand's piece about Bulldozer was this part:
"The APM modules samples a number of performance counter signals and these samples are used to estimate dynamic power with 98% accuracy."

So, static power cannot be a factor in Power Management then? As must be the case with GCN-chips, since they're also using digitaly obtained estimates from activity counters. I am not really sure if I might be missing something, but wouldn't have the set values then contain worst-case guardsbands for the most leaky chip that'll ever get past their screening process?

There should be a limit or indirect impact with temperature... Temp allways influence the TDP. ( a chip at 1.2Vcore 900mhz / 80°C have an higher TDP of a chip at 1.2Vcore 900mhz / 50°C ). If you use watercooling, and have let say 45°C maximum with it at 900mhz stock... and compare the TDP with a standard air cooled card at same clock speed.. you will find 5-10W difference ( with the 580 it was more in the 15W ).. and this have a big impact if your turboboost, is power constrained. ( higher temperature = higher leakage ). So basically, lower is the temperature of the core, higher will go the Turboboost as the limit is not the same.

Basically both temperature and TDP are dependant of each other... if you set a max limit of TDP, and the temperature increase, the TDP will increase, so the software will throttle down the clock for going back to the TDP fixed. The temperature will go down, So you need something for be sure to dont get too much throttle up and down due to this.

 

[Silent_Buddha]

Since Dave is actively participating in this thread I'd like to take the chance and ask about how GPU-Boost is different from what Intel and AMD are employing in their CPUs. Based on TDP headroom, clocks are increased - shouldn't this also be subject to individual ASIC variation?

Intel clocks relatively conservatively, hence all GPUs of a given speed grade are virtually guaranteed of hitting their turbo clocks under the same conditions. Cooling solutions, how many cores in use, etc. They aren't dependant as much on the chip quality. Similar to how ATI sets up for the worst case scenario, Intel does so as well. Hence the relatively huge overclock potential on Intel CPUs.

Nvidia is cutting it much closer to wall. Hence not all cards are guaranteed to hit the same automatic boost levels under the same conditions as you see with Intel Turbo on their CPUs.

In other words there's not really a gamble with Intel CPUs unless you manually overclock. But with the GTX 680. Well, good luck on getting a good chip versus a marginal chip.

So while the mechanism might be similar. The end user result is vastly different. On the one hand you have consistent results across the board (Intel Turbo) on the other you have relatively wild variations from card to card (Nvidia Boost).

On the other hand, it is a good way to win benchmark reviews, especially if you make sure reviewers get cards that boost well.

Regards,
SB

 

[CarstenS]

In other words there's not really a gamble with Intel CPUs unless you manually overclock. But with the GTX 680. Well, good luck on getting a good chip versus a marginal chip.

So while the mechanism might be similar. The end user result is vastly different. On the one hand you have consistent results across the board (Intel Turbo) on the other you have relatively wild variations from card to card (Nvidia Boost).

You make it sound like there's a potentially huge penalty. Do you have an idea how much the chips would vary? With a 1006 base clock, Nvidia would guarantee 1058 MHz, right? That's ~5% and IMHLO not worth all the noise that's currently buzzing around this feature.

On the other hand, it is a good way to win benchmark reviews, especially if you make sure reviewers get cards that boost well.

Do you really think that's dependant on boost levels beyond the guaranteed minimum?

 

[Alexko]

Thx! So Turbo won't be dependant on the cooling solution employed, right?
What catched my attention from Anand's piece about Bulldozer was this part:
"The APM modules samples a number of performance counter signals and these samples are used to estimate dynamic power with 98% accuracy."

So, static power cannot be a factor in Power Management then? As must be the case with GCN-chips, since they're also using digitaly obtained estimates from activity counters. I am not really sure if I might be missing something, but wouldn't have the set values then contain worst-case guardsbands for the most leaky chip that'll ever get past their screening process?

Well, technically and as Dave mentioned, the kind of power management used by AMD that relies on digital activity counters can be set to be deterministic or not, whether you give the same parameters to all chips or not. If you do choose to be deterministic, they have to be set for the worst case, or you risk blowing your TDP.

So I got a bit carried away because I'm not 100% sure that AMD is doing that, but since they're using the same system on BD that they are on GCN, and since CPU customers tend to be much more demanding when it comes to reliability and reproducibility, it's very likely to be the case.

As far as Intel is concerned, Silent_Buddha is probably right.

There should be a limit or indirect impact with temperature... Temp allways influence the TDP. ( a chip at 1.2Vcore 900mhz / 80°C have an higher TDP of a chip at 1.2Vcore 900mhz / 50°C ). If you use watercooling, and have let say 45°C maximum with it at 900mhz stock... and compare the TDP with a standard air cooled card at same clock speed.. you will find 5-10W difference ( with the 580 it was more in the 15W ).. and this have a big impact if your turboboost, is power constrained. ( higher temperature = higher leakage ). So basically, lower is the temperature of the core, higher will go the Turboboost as the limit is not the same.

Basically both temperature and TDP are dependant of each other... if you set a max limit of TDP, and the temperature increase, the TDP will increase, so the software will throttle down the clock for going back to the TDP fixed. The temperature will go down, So you need something for be sure to dont get too much throttle up and down due to this.

Careful, TDP ≠ actual power draw.

 

[CarstenS]

There should be a limit or indirect impact with temperature... Temp allways influence the TDP. ( a chip at 1.2Vcore 900mhz / 80°C have an higher TDP of a chip at 1.2Vcore 900mhz / 50°C ). If you use watercooling, and have let say 45°C maximum with it at 900mhz stock... and compare the TDP with a standard air cooled card at same clock speed.. you will find 5-10W difference ( with the 580 it was more in the 15W ).. and this have a big impact if your turboboost, is power constrained. ( higher temperature = higher leakage ). So basically, lower is the temperature of the core, higher will go the Turboboost as the limit is not the same.

Basically both temperature and TDP are dependant of each other... if you set a max limit of TDP, and the temperature increase, the TDP will increase, so the software will throttle down the clock for going back to the TDP fixed. The temperature will go down, So you need something for be sure to dont get too much throttle up and down due to this.

I am not sure I understand what you're trying to tell me here. Basically, I was asking whether or not one would need to factor in worst-case leakage (after screening) for all chips on a certain SKU when estimating power based on activity only and not on measured power?

Maybe an example helps (number just purely for illustrating purposes - completely made up by me!):
For example you have two ASICs, both qualifying for use in HD 7970. One has a leakage of 0.93, the other of 1.0 (let's say, this is the upper boundary, just for argument's sake).

When you're doing DVFS based on activity, you'd have to cap power (i.e. set your "allowed 100% activity level) at the desired TDP of, say 200 watts for the 1.0-leakage part. The other one could potentially go higher, depending on how great a percentage of total power your leakage would be, right?

 

[lanek]

Careful, TDP ≠ actual power draw.

Ofc, but if it have an impact on one, it have forcibly an impact on the second.

 

[Silent_Buddha]

You make it sound like there's a potentially huge penalty. Do you have an idea how much the chips would vary? With a 1006 base clock, Nvidia would guarantee 1058 MHz, right? That's ~5% and IMHLO not worth all the noise that's currently buzzing around this feature.


Do you really think that's dependant on boost levels beyond the guaranteed minimum?

1058 isn't guaranteed however. 1058 is the "average" level of boost for cards out in the wild.

http://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review/4

Accordingly, the boost clock is intended to convey what kind of clockspeeds buyers can expect to see with the average GTX 680. Specifically, the boost clock is based on the average clockspeed of the average GTX 680 that NVIDIA has seen in their labs.

Hence, if 1058 is the average there must be cards that do better and there must be cards that do worse.

And while it's certainly easy enough to discount 5%, in a graphics card review the difference between a 10% lead versus a 5% lead over the competition looks relatively large. Or 15% versus 10%. Or if performance was similar it could lead to relative changes in positioning on a performance bar graph.

Hence, why it's far more useful as a tool to win benchmark review articles than anything else. Especially if you ensure that review sites get cards that boost particularly well.

So if 5% is at 1058, that could go up to 10% at 1110. Hence a well boosting card would influence relative positioning in a video card review that much more.

Regards,
SB

 

[jaredpace]

Turn Powertune into "boost".
Beef up driver features & game support
Cut out compute, go low DP ratio
Increase Z stencil fillrate / ROPs
Sound dampening fluid dynamic fan
Increase power/# of compute units
Focus on perf/mm^2
Faster MC and memory ICs
Still 28nm, refine for low power/higher clock
Learn from GK104's sm design

Give us a huge chip too plz!

 

[Lightman]

I am not sure I understand what you're trying to tell me here. Basically, I was asking whether or not one would need to factor in worst-case leakage (after screening) for all chips on a certain SKU when estimating power based on activity only and not on measured power?

Maybe an example helps (number just purely for illustrating purposes - completely made up by me!):
For example you have two ASICs, both qualifying for use in HD 7970. One has a leakage of 0.93, the other of 1.0 (let's say, this is the upper boundary, just for argument's sake).

When you're doing DVFS based on activity, you'd have to cap power (i.e. set your "allowed 100% activity level) at the desired TDP of, say 200 watts for the 1.0-leakage part. The other one could potentially go higher, depending on how great a percentage of total power your leakage would be, right?

I think this is a reason why AMD has 3(4?) different vGPU for HD7970. Basically by applying different default voltages to certain leakage thresholds they are keeping power consumption in check. This then lends to predictable DVFS behavior for wide spread of leakage in chips.

So your 1.0 leakage chip (assuming 1.0 is the leakiest part accepted) will get 1.1V vGPU and @925MHz will be capped using performance counters to 200W consumption.
Then 0.93 chip will get 1.175V default vGPU and @925MHz still will be capped to 200W based on same DVFS algorithms as more leaky chip.

This approach gives more predictable power consumption of final cards and gives AMD bigger pool of binned chips to choose from.

Now if AMD as example targeted 1.1GHz SKU with 250W TDP then 0.93 leakage chip would probably not make that bin due to clock limit. More leaky parts, lets say 0.95-0.97 would attain that clock with slight vGPU bump, for example to 1.175V and still stay within target 250W where 1.0 chips would go over that power limit. So this theoretical example shows why AMD might have chosen to clock conservatively without competition in sight. They simply could use much more binned chips as HD7970 initially.


I'm sure Dave Baumann have pretty diagram he can't share showing spread of leakage vs required vGPU vs power consumption vs clock.

Anyway all these numbers are pulled of my ... and this is only my understanding how it's done. If I'm wrong please correct me

 

[lanek]

Wow, you are going to far, i dont speak about standard clocked chips, but about overclocked one and Turbo boost and how the limit can vary ..


For what i have constate with 3 different 7970's, the Asic have his importance, but the margin is really large on each...

AMD have been really conservative with the 925mhz core clocks... really.. .. But i think AMD as Nvidia ( at least for Nvidia before they have seen the result of AMD ) have not predict the 28nm parts they was made will have a so good core speed capabality.. ( this can explain why all non retails 680 have a core clock a lot lower of 1006mhz ( including the cherrypicked one used by overclockers teams right now, who show in their GPU-Z screenshots and with last drivers, different default clock )

The test made on 7970 have all shown the cards was never pass over 197W in all games/bench situations . they could have constrain the Powertune at 0% and 220W, and set 1100mhz without any problem, whatever was the Asic: 1112mv, 115mV, and 1117mV ... and still annonce an average TDP of let say 200W.. Is why you can set the core to 1100mhz without touch the Powertune slider, .... you are far of the limit.

Lets not forget, AMD use the term of Max Board Power ( 250W ), and not average TDP in standard situation..

Was really amusing me, i have never see Nvidia and AMD card so close of each others... at 1100mhz ( the turbo boost seen in most games, bench ) , the 7970 is absolutely equal of the 680 ( faster ).. ofc some win on some bench and the other win in the other... TDP wise ( reported max is 195W for the 7970, who is the limit for the 680 ), Performance wise, clock to clock, the result are really similar...
( i dont know if it is a good thing if Nvidia have really something better to give us... but it seems at this moment: 680 vs 7970 this is not a bad competition .... we can only dream of an hypotetic GK110 ( some say August, other say end of the year, but yet we can only count on this fight for now ... all other marketing shit, who cares.. the cards sell at same price., other are only words).

 

[CarstenS]

1058 isn't guaranteed however. 1058 is the "average" level of boost for cards out in the wild.

http://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review/4



Hence, if 1058 is the average there must be cards that do better and there must be cards that do worse.

And while it's certainly easy enough to discount 5%, in a graphics card review the difference between a 10% lead versus a 5% lead over the competition looks relatively large. Or 15% versus 10%. Or if performance was similar it could lead to relative changes in positioning on a performance bar graph.

Hence, why it's far more useful as a tool to win benchmark review articles than anything else. Especially if you ensure that review sites get cards that boost particularly well.

So if 5% is at 1058, that could go up to 10% at 1110. Hence a well boosting card would influence relative positioning in a video card review that much more.

Regards,
SB

You're mentioning for the second time now the possibility that review sites are getting particularly boosty cards - any indications on whether or not that might be the case?

Anyway, the anandtech article you quoted has a quote itself from the reviewers guide, which does indeed sound a bit differently than how Ryan phrased it. It goes like this:

"The “Boost Clock” is the average clock frequency the GPU will run under load in many typical non-TDP apps that require less GPU power consumption. On average, the typical Boost Clock provided by GPU Boost in GeForce GTX 680 is 1058MHz, an improvement of just over 5%. The Boost Clock is a typical clock level achieved running a typical game in a typical environment"

Further down, Ryan re-tested their parcours with what he thinks effectively negates GPU-boost. If i calculated correctly, GPU-boost (above 1006, not above 1058 mhz) gave a tremendous gain of 3,34% on anandtech's benchmarks (Ryan said 3%).

Further, he shares this insight:
"NVIDIA has on-card sensors to measure power consumption at the rails leading into the GPU, and will only allow the video card to step up so long as it’s below the GPU Boost power target. This target isn’t published, but NVIDIA has told us that it’s 170W. Note that this is not the TDP of the card, which is 195W. Because NVIDIA doesn’t have a true throttling mechanism with Kepler, their TDP is higher than their boost target as heavy workloads can push power consumption well over 170W even at 1006MHz."
There's - going by Ryans words - a safety margin of roughly 25 watts for more demanding applications and - if you're right - more leaky chips, that should ensure, that they're mostly able to hit a little turbo now an then.
--
From what I've seen there and at other places including our own lab, where we saw power going into the 680 being contained to just above 170 watts (max. 174 watts in our testing games) plus a 680 sample that was obtained individually (and within 1 watt of the other card under load), I doubt that variability GPU-Boost has influenced the outcome of a review. Said Anandtech review sees 680 ~12% (individual percentages averaged) in front of 7970 at 2560 counting the gaming tests including portal 2 with supersampling.

--
I think we've both made our points. I for one will not re-iterate the same arguments over and over again just to keep the impression of having "won" an argument.

On computing parts: the 7970 is largely faster anyway of the 680 ( i have not say faster of Kepler )

That's what I find really amusing about the current situation. For years Nvidia's been telling people how they're investing in Compute and now they're trailing AMD in that area with GK104.

I'm just waiting for a semi-accurate article accusing Nvidia of loosing focus on compute - just as they did two years ago for Nvidia forcing down compute every gamers' throat with broken and unfixable, software-tessellating Fermi.

 

[jimbo75]

You're mentioning for the second time now the possibility that review sites are getting particularly boosty cards - any indications on whether or not that might be the case?

Hexus and Anandtech had samples over 1100 MHz in the games they saw. I haven't looked at the rest but I'll be pretty amazed to find they weren't performing over the "average". Tom's hit 1110 MHz in 2 games and a little below that in the rest - http://www.tomshardware.com/reviews/geforce-gtx-680-review-benchmark,3161-3.html

Given plenty of observable headroom, it gets our board’s maximum clock rate of 1110 MHz. Metro 2033 is another fairly low-power game, which also gets pegged at 1110 MHz. Crysis 2, DiRT 3, and even Battlefield 3 receive less consistent boosts. Still, though, they run a bit faster, too.

I haven't seen any game below the "average" yet but I haven't looked that hard.

As for the rest, Ryan benched the boost at 1600p, where it's performance gains are likely being more limited. Of course the really interesting thing would be to have benched it at 1080p seeing as that is where the card is being positioned as unbeatable but you'll have to ask him why he didn't do that. I have my own theory of course.

 

[silent_guy]

As for the rest, Ryan benched the boost at 1600p, where it's performance gains are likely being more limited. Of course the really interesting thing would be to have benched it at 1080p seeing as that is where the card is being positioned as unbeatable but you'll have to ask him why he didn't do that. I have my own theory of course.

Actually, I would really like to hear that theory! Ryan probably too, since he's a reader here.

 

[Mianca]

Learn from GK104's sm design

What should they learn? Their design is still better.

If Pitcairn was a ~300mm² chip with some 6GHz memory at its disposal, it would win against GK104 in total performance, perf/mm² and perf/W.

If Pitcairn was a ~300mm² chip with some 6GHz memory at its disposal AND a power determinisitc auto overclocking switch that keeps the chip right at the 200W mark when under gaming load - it would just kick GK104 out of the stadium.

GCN is a very sound arcitecture - just as PowerTune is a very sound high performance power management feature.

The problem with the GK104 vs. Tahiti comparison is that we compare a gaming performance optimized chip that comes with a stock power deterministic boost feature to a compute optimized chip that's clock deterministic by default - and very conservative at that.

GTX680 is a sleek sports car.
It's designed to be fast on the highway - and that's about it. It drives at least 100mph - but it will usually drive faster by default as long as fuel consumption doesn't exceed a certain limit. Problem with this appraoch is that the car tested in reviews might run on some high performance fuel which you aren't guaranteed to get as a regular customer. On top of that, some review sites might chose not to test it on roads that aren't explicitly specified it their owner's manual - but that's another story.

HD7970 is a sportive van.
It's designed to be fast when driving on a highway - but you can also use it to do some shopping or go on a weekend trip with your family. When on a highway, it goes into cruise control. It will drive 92.5mph. All the time. There is no hard limit in fuel consumption it will strive to exploit. It doesn't care if you're driving downhill or uphill. It will just drive 92.5mph. It might throttle a little when driving up to the snowy tops of Furmark mountain - but that's about it.

So, yeah. Let's see what happens when AMD actually builds a sleek sports car in the same cubic capacity range as GK104.

 

[silent_guy]

What should they learn? Their design is still better.

If Pitcairn was a ~300mm² chip with some 6GHz memory at its disposal, it would win against GK104 in total performance, perf/mm² and perf/W.

If Pitcairn was a ~300mm² chip with some 6GHz memory at its disposal AND a power determinisitc auto overclocking switch that keeps the chip right at the 200W mark when under gaming load - it would just kick GK104 out of the stadium.

GCN is a very sound arcitecture - just as PowerTune is a very sound high performance power management feature.

The problem with the GK104 vs. Tahiti comparison is that we compare a gaming performance optimized chip that comes with a stock power deterministic boost feature to a compute optimized chip that's clock deterministic by default - and very conservative at that.

GTX680 is a sleek sports car.
It's designed to be fast on the highway - and that's about it. It drives at least 100mph - but it will usually drive faster by default as long as fuel consumption doesn't exceed a certain limit. Problem with this appraoch is that the car tested in reviews might run on some high performance fuel which you aren't guaranteed to get as a regular customer. On top of that, some review sites might chose not to test it on roads that aren't explicitly specified it their owner's manual - but that's another story.

HD7970 is a sportive van.
It's designed to be fast when driving on a highway - but you can also use it to do some shopping or go on a weekend trip with your family. When on a highway, it goes into cruise control. It will drive 92.5mph. All the time. There is no hard limit in fuel consumption it will strive to exploit. It doesn't care if you're driving downhill or uphill. It will just drive 92.5mph. It might throttle a little when driving up to the snowy tops of Furmark mountain - but that's about it.

So, yeah. Let's see what happens when AMD actually builds a sleek sports car in the same cubic capacity range as GK104.

I love car metaphors because they enable getting ideas across in a very concise ways. (Sorry for quoting your whole post, but it was the easiest way to read it without those annoying colors that were added by some bot.)

 

[Mianca]

I love car metaphors because they enable getting ideas across in a very concise ways. (Sorry for quoting your whole post, but it was the easiest way to read it without those annoying colors that were added by some bot.)

Yeah, sorry, that's Mr. Colorson. He's a dear friend of mine. Keeps lurking right above my text fields when I type posts. I told him not to interfere with my texts and leave my metaphors alone. Many times. But he's such a cheeky rascal. Always finds a way to pull my leg.

Luckily, he isn't half as bad as Mr. Smiliespam.

 

[CarstenS]

Hexus and Anandtech had samples over 1100 MHz in the games they saw. I haven't looked at the rest but I'll be pretty amazed to find they weren't performing over the "average". Tom's hit 1110 MHz in 2 games and a little below that in the rest - http://www.tomshardware.com/reviews/geforce-gtx-680-review-benchmark,3161-3.html

I haven't seen any game below the "average" yet but I haven't looked that hard.

As for the rest, Ryan benched the boost at 1600p, where it's performance gains are likely being more limited. Of course the really interesting thing would be to have benched it at 1080p seeing as that is where the card is being positioned as unbeatable but you'll have to ask him why he didn't do that. I have my own theory of course.

You've got a point there - lower resolutions, when not running into a CPU limitations big time increase the sheer number of fps and thus oftentimes (at least in my experience) also increase power draw, thus heat, thus turbo-ability for GTX 680.

But then, there's also temperature as a factor as Ryan mentioned. As long as your card stays below 70 (°C i belieb he meant), your turbo will hit higher rates than above that. But that also implies, that better cooling will help your GPUs speed, doesn't it? Now, if Nvidia had adopted a very aggressive cooling solution turning up the noise at below that temperature threshold, I could believe they did that just in order to achieve marginal margins on some reviews.

What should they learn? Their design is still better.
If Pitcairn was a ~300mm² chip with some 6GHz memory at its disposal, it would win against GK104 in total performance, perf/mm² and perf/W.

If Pitcairn was a ~300mm² chip with some 6GHz memory at its disposal AND a power determinisitc auto overclocking switch that keeps the chip right at the 200W mark when under gaming load - it would just kick GK104 out of the stadium.

If I take Computerbase's ratings, GK104 is roughly 34 to 40 percent above Pitcairn, it's die size is 38,6% above - it'd be a close call I think, but probably nothing like one blowing the other out of some sports arena. And that's assuming area can be translated into performance on a 1:1 ratio. Extrapolated power budget headroom (114 watts measured for example in BF: BC2) to match GTX 680 would be ~16 watts above 1.386 x 114 watts (158 watts for a scaled Pitcairn).

For example: How large would memory controllers have to grow in order to get past AMDs binning at such speeds?

GCN is a very sound arcitecture - just as PowerTune is a very sound high performance power management feature.

True.

 

[Mianca]

If I take Computerbase's ratings, GK104 is roughly 34 to 40 percent above Pitcairn, it's die size is 38,6% above - it'd be a close call I think, but probably nothing like one blowing the other out of some sports arena.

It's all personal guesstimation, of course, but in order to evaluate the basic architectural performance efficiency, I basically compared GTX680 results minus 5-10% boost speed (i.e. 1Ghz base clock) to the results of a 1Ghz HD7870@6000MHz mem clock - and extrapolated from there.

With respect to power, it's not even close. According to W1zzard (whose power measurements are as good as they get - directly measuring via PCI-Express power connectors and PCI-Express bus slot) GTX680 draws 60% more power on average than HD7870. So Perf/W is nearly 30% better on average without even making some adjustments.

As for scaling - just look at at the perf/W and perf/mm² scaling of HD7770 vs. HD7870. Without all the compute related stuff, GCN scales very well: As a matter of fact, HD7870 even has 6% better average perf/W than HD7770, while perf/mm² is only slightly worse (Pitcairn is about 72% bigger than Cape Verde - with HD7870 performing 66% faster than HD7770).

So, yes. Maybe I exaggerated a bit. But I'd still bet on a gaming optimized 300mm² GCN GPU to consistently beat GK104 by a fair margin - even without AMD implementing all that Boost stuff @ stock settings.

The fact that we're even discussing this is an impressive testament to NVidia's very good work with GK104, though - and who knows: If they can keep improving efficiency like that, Maxwell will be VERY hard to beat.

 

[CarstenS]

Sorry to disappoint, but "average" is just Crysis 2:
"Average: Crysis 2 at 1920x1200, Extreme profile, representing a typical gaming power draw. Average of all readings (12 per second) while the benchmark was rendering (no title/loading screen)." (From your link)

The numbers I mentioned are also directly from the slot plus PSU connectors, so they're not guesstimates any more than what you've linked. Avg. Performance per watt correllating performance in many games and watts in just one is more a rough ballpark (albeit a very useful one!) than a number whose decimal's points I'd trust in for inter-arch comparisons.

Here's a few games with watts and fps at exactly the same spot:
http://www.pcgameshardware.de/aid,873907/Test-Geforce-GTX-680-Kepler-GK104/Grafikkarte/Test/?page=19
Unfortunately we've only had time to do 680, 580 and 7970 here (for 7970 we've also used the reviewer sample provided by AMD). But nevertheless, the results are way less advantageous as Nvidia would like you to think. 680 is beating 580 handily though.

 

[jimbo75]

Using Anands numbers on the 680 and 7870 power bench during OCCT...

They both idle at 112W, the 7870 draws 259W during OCCT and the 680 draws 333W. That's a difference of 221/147 = 50% higher for the 680.

I think TPU's perf/watt numbers are off however due to Crysis 2 scoring higher fps on the 7870 compared to the 680, which clearly isn't the case.

I'd guess the actual lead for the 7870 over the 680 in perf/watt is 10% -15% depending on the game. Some will be closer to 5% and some might be closer to 30% but the average will be around 15% I think.