- Tomshardware.com, Chris Angelini
Why quadruple the number of CUDA cores and double the other resources? Kepler’s shaders run at the processor’s frequency (1:1). Previous-generation architectures (everything since G80, that is) operated the shaders two times faster than the core (2:1). Thus, doubling shader throughput at a given clock rate requires four times as many cores running at half-speed.
The question then becomes: Why on earth would Nvidia throttle back its shader clock in the first place? It’s all about the delicate balance of performance, power, and die space, baby. Fermi allowed Nvidia’s architects to optimize for area. Fewer cores take up less space, after all. But running them twice as fast required much higher clock power. Kepler, on the other hand, is tuned for efficiency. Halving the shader clock slashes power consumption. However, comparable performance necessitates two times as many data paths. The result is that Kepler trades off die size for some reduction in power on the logic side, and more significant savings from clocking.
To answer your question, essentially, EVGA's OC Scanner is misleading you. I don't think it's been properly updated for the 680s unique hardware. As far as I'm aware this is the first dedicated GPU where the shader clock and the core clock run 1:1. I think there have been a few "on die" graphics solutions that have either cut way down on the shaders and run them 1:1 to save power or most of them simply offload the shader computations to the CPU itself. Up until now 3D Shader clocks have always been 2:1 or even 4:1 compared to Core clocks.
Basically, Kepler has been engineered from the ground that Fermi stood on to be even more efficient. I've played a variety of games and had several where the 680 barely ever moved off the 324 "idle" clock (e.g. - Sins of the Solar Empire). I mean even during "massive" armada battles, Kepler only bounced a lil up to the 540 Mhz for a few mins. The 680 was downright bored with Sins. I've played games all along the scale up to Crysis with it churning out the clocks it needed to give me a good game experience (with very few moments where it couldn't quite keep up) without upping the electric bill if it didn't have to. Just played some Supreme Commander: Forged Alliance tonight and it was all over the place (tho never at the 324 clocks, but everything else from 540 to 1267. Up til now, GPUs have pretty much been either in "idle" clocks, or "Givin it all she's got Cap'n." Other then gauging performance, do we really need our GPUs cranking out 200+ FPS on every game? Of course not, it's a waste. Most moniters only display 60 Hz or FPS, and the best ones really only go to 120 (with a few old timers using CRTs @ 75 or 85 Hz/FPS). Anything over that is just for bragging rights, poor visual quality, and migraines.
As someone who isn't rich and has to pay his own electric bill, I deeply appreciate what Nvidia has done with Kepler. Bottom line, is in fact, the bottom line, if the 680 doesn't need to flex it's muscle to give you a quality performance, it won't (and saves you money), but if you need it, the power is yours and Kepler will indeed reply, "I'm givin it all she's got Cap'n."
As for over clocking the 680, there seems to be a ton of assumptions, wild hairs, and morons spouting a lot of voodoo or just plain nonsense.
The Kepler has been designed around efficiency (as we've already covered). So, OCing it requires some understanding of a few things. First, the normal Vcore on your Mobo is about 1.2 V that supplies power thru the PCIe bus, and the powers that be (that designed the architecture) only allow about 75W (1.1) or 100W (2.0 & 3.0) to be drawn thru the motherboard to your card (Tho everyone sticks with the 75W limit for backwards compatability. Forcing this voltage higher can do VERY BAD THINGS
. (Remember volts* amps = watts, and factors for resistance - so 75W / 1.2 V = (roughly, Ohm's Law haunts us again) 60a, and that's a lot of current running running thru your Mobo generating heat. In the case of the 680, doing so will also downclock performance by about 25ish% (yes, I've tried it), there seems to be an actual voltage limiter built into the card. I suspect it's not just to frustrate overclockers, but to protect the chip (we are dealing with a 28 nm process now) and the motherboard itself. The added benefit by not going all the way to 1.2V is that it generates less heat (easier to cool, less noise, less wasted power due to resistance, more bang for the buck).
This leads us to the Power Target
setting. Seems to be a lot of misunderstanding as to what this does. We've already discussed the 75 Watts thru the PCIe bus, the additional power comes thru the 6 pin connectors off your PSU. 6 Pin connectors are also rated for 75W each. That's a total of 225W max power draw. However, the engineers at Nvidia designed the cooling on the 680 for a max thermal load (remember most power used by compters gets turned into heat) of 195W at what they determined to be a noise level that didn't replicate the PR nightmare that was the "dustbuster."
Engineers also have this habit of a "safety net", and since not all chips measure up to what designed on paper, they reduced this by roughly 10% and rated the card at 175W. 175W * 132% = 231W ... a little over the supposed 225W max draw, but lets face it, you don't buy a $500 GPU and slap it into a bare minimum set up. In case you didn't follow that, the 132% from the aforementioned equation is the max setting for the Power Target slider on your OC application. However, most of you have noticed that it never seems to even get close to 132% of TDP (or Thermal Design Power) Side Note: I've gotten to 127.2 but only for a moment . My card seems to like running @ 115%-120% most of the time. The closer you can keep the card to 1.175V, the more power it's using, the higher it's TDP. With TDP, you'll notice the first word is Thermal. Which leads us to Temps.
There seem to be a number of thermal "steps" that correspond with voltage and clock "steps". We've already seen a number of these discussed here (69.5oC, 70.5oC, 79.5oC, 80.5oC... up to 90.5). There seem to 13 voltage steps that depend on temp, with the closer you get to either 90.5oC or 0 load the closer it goes from 1.175 to 0.987V. This is Overclockers. net, I don't feel the need to explain voltage and temp. Basically, the closer gets to 90.5oC the more it steps the voltage down (reducing leakage, power, heat) to cool the card. We get that.
There are also voltage steps associated with Load, if the card isn't running @ 98+% it may step voltage down. The other odd quirk is that even if a monitor program of some sort tells you it's high load there may be parts of the card that aren't under load (warp schedulers, dispatch units, shaders, register filers, etc.) and the card will reduce voltage because it doesn't need the power for those parts of the chip. It actually turns the light out when it leaves the room, unlike my wife.
To use an automotive analogy, the Kepler isn't a dragster or a muscle car. It's not a straight line race - add more power, keep it cool, get more FPS. It's more of a LeMans European Coupe, sure it's got tons of horsepower, but it was also designed to handle the curves and give you great fuel economy so you can win. Let's not forget, there is a Speed Limit per se with Graphics Processing, we gain nothing by going over 120. Kepler's not just raw power, it's a thing of elegance.