7970 Lightning now available

[railven]

That's how we have always counted any product with dynamic OC --> see CPUs.
Please take 10 min to read the PureOverclock 1 page, it explains it much better than I do.

GTX680 has less overclocking room because dynamic OC already overclocks part of the way there. That has nothing to do with GTX680 having poor overclocking scaling. What it rather means is 680's have lower overclocking room to gain performance from dynamic base clocks.

GTX680 has 1006 factory clocks.
Its operational clocks are 1110mhz.
Manual overclock is 1300mhz.

How much is that overclock in the real world? 1300mhz / 1110mhz = 17%

Therefore, you should see at most a 17% performance increase from stock benchmarks, which is exactly what's being shown.

It's the same as we've always counted it with CPUs. This applies to any product with dynamic OC functionality built in. For example, while Core i7 860's stockspeed is 2.8ghz, it turbos at a minimum to 2.93ghz on all 4 cores. If I were to run all my 4 threaded benchmarks with i7 860 and get my Base scores and assign them 100% reference, then compared to those scores a manually overclocked 860 @ 3.2ghz is not 18% faster! It would correct to say that my CPU was overclocked by 14% from "stock clock of 2.8ghz" but it would be INCORRECT to compare benchmarks and look for a 14% growth since the stock 860 never operates at 2.8ghz. It operates at a minimum of 2.93ghz. The largest increase I would see is 9%. Alternatively, imagine if you ran a single threaded bench such as SuperPi. A stock i7 860 would beat a manually overclocked 3.2ghz i7 860. Using your logic, you'd get negative CPU scaling. That's because it's incorrect to use 2.8ghz as the base clock for a Turbo enabled CPU.

For mobile CPUs, the CPU boost is even more. You can have a mobile CPU boost 1ghz from 2ghz. Are you going to say that if you overclocked your CPU to 3.2ghz, that's a 1.2ghz overclock from stock? NO. That's only a 200mhz overclock from factory operational clocks.

So if you are going to look at frequency scaling, the base clock = highest frequency scaling out of the box vs. overclocked highest frequency.

We wouldn't say that a 4.0ghz 2600K is 18% faster clocked than a reference 2600K with Turbo. We would say a 4.0ghz 2600k has 18% higher frequency than 2600K operating at 3.4ghz. This is because 2600K can turboboost to 3.8ghz. So obviously you'll never see 18% scaling on 2600K @ 4.0ghz over a 2600K with Boost up to 3.8ghz.

The benchmarks are showing this exactly. The base = 100% is likely at minimum 1110mhz, not 1006. So it's 100% incorrect to compare reference benchmarks and then say it's a 30% overclock from 1006 to reference #s because the GPU never operated at 1006mhz. You would say GTX680 has less overclocking headroom since Dynamic OCing has eaten into that headroom with a 7-10% factory dynamic OC.

There is no apparent GPU scaling problem with Kepler, it's simply incorrect use of mathematics using an incorrect 1006mhz as the base for the benchmarks. What Kepler has is a lower overclocking headroom in % terms vs. 7970.

I had to read this three times, because essentially in it you said what I said, just three times, if not more.

Basically with such a moving target (note CPU's don't have such a wide moving target) one can't really say where their reference point is. Either we use the reference clock (and stress that) or we don't use anything and let the range be so wide, that people's results will vary on how hot their room for any given moment.

All we can do is make a range - and in that range, anything can be argued, again my two examples of using a box versus a good case. The Boost will vary in both situations.

EDIT: Your source, http://www.pureoverclock.com/review.php?id=1461&page=19, says the same thing, yet they don't disclose the difference minus the GPU boost:

Ultimately, our card achieved a 1245MHz overclock speed., which roughly translates to a 24% overclock from the Base Clock, which is excellent. Keep in mind though, the actual OC increase is only 145MHz, because our card already provided about a 100MHz boost out of the box. And this is also why overclocking Kepler, while simple, provides diminishing returns, because the GPU Boost is essentially already overclocking the card for you. You can certainly OC the card, but you won't be gaining too much horsepower beyond what's already being offered for "free". The overall framerate improvements in games won't be very much, perhaps 5-7% on average, depending on the Boost Potential of your particular card. That said, the GeForce GTX 680 overclocks like a monster, and seeing a 24% overclock is outstanding.

24% OC over Base (1006) they note the 100+ GPU Boost, but don't subtract it. Had they subtracted, it only be 12% OC, but again - it isn't written there. Murky waters.

 

[RussianSensation]

Basically with such a moving target (note CPU's don't have such a wide moving target) one can't really say where their reference point is.

It's much more accurate to assume that the reference benchmarks have a reference point somewhere between 1058 and 1110mhz than to assume them to be 1006mhz. Assuming 1006mhz for benchmarks as a reference point is actually the most incorrect assumption of all.

However, we don't even need to assume much. Right now, it's pretty clear that @ 1200mhz or higher, HD7970 is slightly faster than an overclocked GTX680. However, to get that overclock, you more or less have to spend $80-100 for premium Sapphire Dual-X, Asus Direct CUII, or MSI Lightning or Gigabyte Windforce 3X versions.

With performance being so close, I am not seeing much sense in spending $80-100 more for what is maybe a 5% faster card in a 1200-1250mhz 7970 over a 1200-1300mhz 680. I've stuck to my guns when I recommended HD6970/GTX570 over 15% faster 580 (a savings of $150 last generation) and there is no way I can recommend a 5% faster card for $80-100 more, especially since it requires overclocking that's not even guaranteed. GTX580 had guaranteed 15% faster performance over 570 out of the box and most of us still deemed it to be poor value. In this case, outside of bitcoin mining, it's very hard to make a case for the 7970 for gaming at current pricing structure. A gamer would likely be better off pocketing $80-100 today, then selling 680 in 6 months and reinvesting that $80-100 savings into GTX685, or w/e else faster card is around. Alternatively, $80-100 saved for 5% performance difference will net a lot more performance in the future.

Also, HD7970 is only close in performance at 2560x1600. At 1080P, 680 easily leads, and esp. in BF3, Dragon Age 2, SKYRIM, Dirt 3, Batman AC, WOW, SC2 (most likely Diablo 3). Those are some popular games right there. The choice for gamers has never been easier. The only notable games where 7970 puts up a strong showing are Metro 2033 and Crysis 1/Warhead.

24% OC over Base (1006) they note the 100+ GPU Boost, but don't subtract it. Had they subtracted, it only be 12% OC, but again - it isn't written there. Murky waters.

Ya, they established the base as 1006mhz. The % is simply (New - Old) / Old x 100%. Old value = 1006mhz (base). In regard to benchmarks for reference 680's is the base 1006mhz? No it's not, since GTX680 does not operate at 1006mhz but at reference speeds runs at 1110mhz. So the "Old" value is really something like 1110mhz (but even this isnt' accurate since GPU speed varies on a game-by-game basis).

I stated it 3x because it seems you keep talking about % overclock while if you had read the posts more clearly, the issue is measuring GPU scaling in % terms from reference scores. You are discussing semantics of % overclock from base vs. boosted clocked. No one is arguing that a 1300mhz overclock is not a 30% overclock, of course it is. The the issue pointed out is using reference benchmarks and assigning them a stock clock of 1006 when measuring GPU scaling at 1300mhz. The "base performance" for 680 in benchmarks is NOT a card that operates 1006mhz, but a card that runs at overclocked frequencies in the first place. Therefore, the 1006mhz base clock is a meaningless metric in deriving 1300mhz GPU scaling and concluding that 680 doesn't scale as well with GPU overclocks as does HD7970. All it means is the dynamic OC eats away the overclocking room that normally would be achieved with a manual overclock. It doesn't mean that 680 GPU has poor scaling, what it means is GTX680 has lower % overclocking headroom. Those are not the same thing. There was no 30% overclock from reference benchmark scores because the benchmark scores were never run at 1006 clocks.

Therefore, the GPU scaling in no way is to be expected @ 30% since the GPU has only been overclocked to 1300mhz from 1110mhz, or 17%. If you want to state that a 1300mhz is a 30% overclock from a base clock of 1006, that statement is correct because you clearly identified a starting point. However, GPU scaling observations are being erroneously presented against 1006mhz clocks since the card hardly runs at that speed.

Either way, at this point in time GTX680 is more or less sold out. Right now this is the only thing saving 7970's (and bitcoin mining). If NV manages to ship enough 680's in the next month or so, AMD will need to lower prices or release faster clocked variants. GTX680 costs $80-100 less than aftermarket 7970s that can reach 1200mhz+ while maintaining quiet operation. Since most 7970s cannot easily reach 1250mhz on air, the 2 cards more or less trade blows at max overclocked speeds, leaving 7970 with consuming more power, being $80-100 more expensive and having a lower feature set. In the NA market, 7970 is no longer competitive and 7950 is vastly overpriced.