[Official] GTX 580 Overclocking Club - Page 271

Quote:

Originally Posted by juano 

Quote:

Originally Posted by friend'scatdied 

The 'binned best of the best' is still inaccurate unless you call an 832MHz clock speed best-of-the-best. I haven't really heard of any GTX 580s that weren't able to break 832MHz.

I'd call it better than all but 3 or so other GF100s (gigabyte SoC and the two EVGA classies). But the point I was making is that it's a full 580, i.e. it isn't downgraded to a 570 or 560ti 448, and it's also the highest in MSIs binning (as well as the world, save for the 3 others mentioned) in that it isn't a TFII card, it isn't a reference board OC card, and it isn't a completely reference board. So it's the top of 3 tiers in terms of Nvidia binning and the top in MSI's binning of 4 tiers for GF100s. All of these things while not guaranteeing, do increase your chances of getting a good chip by the simple fact that it removes the cards that don't pass these tests from the pool and puts them in lower tiers. I don't know if you don't understand this or just feel like nitpicking me but I'm done 'debating' with you.

I'm going to have to agree with juano here, they are binned, and they are the "best of the best" from a QA and probability/statistics standpoint.

Quote:

Originally Posted by friend'scatdied 

Quote:

Originally Posted by juano 

I'd call it better than all but 3 or so other GF100s (gigabyte SoC and the two EVGA classies). But the point I was making is that it's a full 580, i.e. it isn't downgraded to a 570 or 560ti 448, and it's also the highest in MSIs binning (as well as the world, save for the 3 others mentioned) in that it isn't a TFII card, it isn't a reference board OC card, and it isn't a completely reference board. So it's the top of 3 tiers in terms of Nvidia binning and the top in MSI's binning of 4 tiers for GF100s. All of these things while not guaranteeing, do increase your chances of getting a good chip by the simple fact that it removes the cards that don't pass these tests from the pool and puts them in lower tiers. I don't know if you don't understand this or just feel like nitpicking me but I'm done 'debating' with you.

Yup, it's binned in the 580 sense (the only bin above being for the GTX 590).

I don't agree. The 832MHz clock is far too conservative to be considered a meaningful binning. Until I see a GTX 580 that can't hit 832MHz stable, I'll hold to that opinion. I'm willing to bet the clock is conservative enough so that virtually all of the cards they test qualify for the Lightning SKU until they meet the quota for Lightning production, which minimizes production costs.

Seems like they're twisting QC into marketing.

You are incorrect, friend'scatdied.

Proof:

( Link of quote: http://www.overclock.net/t/1201776/evga-geforce-gtx-580-classified-ultra-or-zotac-geforce-gtx-580-3gb-amp-2-ed#post_16226030 )

Quote:

Originally Posted by EVGA-JacobF 

580 Classified is a hand selected GPU, custom PCB, custom VRM and new cooler designed for OC... also supports memory voltage and PCIE voltage increases via software.

Clocks over 1GHz are almost a guarantee. So if you are interested in overclocking I would definitely suggest it.

Hand selected GPU is by definition "binned" as we like to call it. What's funny is that you posted three posts down from that very quote in that thread. I'm not sure how you missed it, catdied...

In another thread, Jacob says the same type of thing:

Proof:

( Link of quote: http://www.overclock.net/t/1199084/evga-580s#post_16184305 )

Quote:

Originally Posted by EVGA-JacobF 

VRAM is not binned, GPU is tested and certified at rated speed. Possible it may OC higher, but your mileage may vary.

Only exception is 580 Classified Ultra, this card has hand selected GPU's that in most cases scale beyond 1GHz (and sometimes more if you cool it good)

For gaming today and tomorrow I would recommend a 3GB card, as with MSAA in some titles we are getting pretty close to the 1.5GB limit.

Again, no binning is occurring in the VRAM but heavy GPU binning happening in the classified ultras and even though they only ship from the factory at a 900 MHz clock, he's saying that they scale beyond 1 GHz in most cases.

How that can't be "best of the best" is beyond me. I can't help but wonder if you're just arguing to be contradictory. In your defense, you did not post in this specific thread, but you did post on the thread after he posted in the previous quote.

friend'scatedied: You don't seem to understand how QA works in the manufacturing industry and I'm not sure that you have a handle on what standard deviations and confidence intervals mean to a manufacturing industry in terms of product reliability and guarantees.

There are two points that you are overlooking:

First, factory binning will never produce results as good as an overclocker at home tweaking his own system. The factory is testing based on a highly polished quality assurance statistical analysis method that virtually guarantees that whatever they're sending out the door will work at the speeds and voltages they set it at in the firmware on any system that meets the CPU, PCIe Bus, and Power Supply supply requirements of the card. Moreover, they're using statistical Confidence Intervals that are almost certainly 95% (if not 98% or even 99.5%) in their testing process to verify that the voltage they chose for that clock will work in any system that meets the requirements.

In a nutshell, nearly every card that comes off of the line has at least some OC headroom. It's entirely possible that you'll get lucky and get a card that's in the top Standard Deviation for your GPU, even if it's just a vanilla reference 580 and you'll end up 24/7 stable at 950 MHz. The factory, however, cannot waste the time it takes to discover each individual card and guarantee that they will perform that well in every system. (Your same card might crash and burn at 925 MHz in a slightly different system configuration.)

So, when the factory takes that stock 772 MHz and guarantees a 832 or even 900 MHz core clock, they're stating that "in any system that meets minimum specifications, we claim that you will never need to RMA the card as we shipped it to you, with any current or future driver or firmware changes with a 95% (or 98% or 99.5% or whatever) Confidence Interval. We make no claims on any changes you make to your system other than approved driver updates and approved firmware/BIOS updates."

Second, when you buy that stock 772 MHz reference card, they're making the same statement, but at 60 to 128 MHz less on the core.

In a nutshell, the "promise" that they're giving you with the reference card is a guarantee of a much lower quality on the core speed.

Thus, it is a fair statement to say that any card above the reference frequency is a "better" card than a vanilla reference card, from a production QA statistical standpoint. The manufacturer is tacitly saying so as well.

The fact that so many vanilla reference 580s are reaching 832 MHz with no problem at all when a knowledgeable overclocker tweaks his card with third party software individually for his own system simply speaks to the quality of the design of the Fermi architecture itself and doesn't reflect poorly on the high-end cards that guarantee this from the factory. Nor does it make the people who buy said cards foolish.

The EVGA Product Manager has confirmed (on these very forums) that EVGA bins the GPUs for the Classifieds as they are "hand-selected" but that VRAM is not binned. The MSI Lightning is essentially the direct competition to the Classifieds and I'd bet that they're binning their GPUs for it the same as EVGA is. I have no clue what Gigabyte has in that tier of cards but I'll bet that with their name and OC reputation that they have a similar binned product offering as well in a highly modified configuration.

EDIT:

Here's a graph of a 95% confidence interval:

If you put the core frequency on the X axis and the number of cards on the Y axis, you get a normal distribution. The confidence interval (in this graph) simply rejects whatever cards are too far below the norm or too far above the norm. The factory will most likely reject anything below the lower limit as unusable and not sell it at all. In the case of video cards, the cards that exceed the upper limit are candidates for "binning" to go into the super high end product offerings that they sell.

For example, if you set your lower limit (the 8 on this graph) at 772 MHz and your upper limit (the 12 on this graph) at 832 MHz, then you have a 95% CI on every reference card you sell, you reject anything that doesn't meet the 772 MHz reference, and you "bin" anything that exceeds the 832 MHz mark and put it into the "hand-selected" category for the Ultra Classifieds and such. (Or you don't bin the stuff on the right and you end up with a 97.5% CI on the product you sell.)

Nothing says that you can't place different constraints on your upper or lower limit or sud-divide your graph into various product lines for the 797 MHz Superclocked, the whatever MHz FTW edition, the Classified, and then the Classified Ultra as well. Further degrees of binning may occur in those cases, in that they may make the upper limit area (right side of the graph) larger, by decreasing the confidence interval, and then do further QA on those cards to produce another CI graph and pull the best of those for the top-end cards, then use the rest of that lot of cards as the "superclocked" or "overclocked" reference circuit designs.

They also use a similar process to set the voltages for a given card. The voltages will be on the X axis and the number of cards will be on the Y axis. They will put the target voltage as their "reference" voltage at the norm of the graph and test every card at 772 MHz. The ones that fall to the left of center get a higher voltage (to keep them stable so that they don't fail) and the ones that fall to the right of center get a lower voltage. (The voltage may be run in reverse on the graph with low voltages on the left and high voltages on the right; it's irrelevant which way they place them.) This ensures stability with a 95% CI in the shaded area. (Perhaps they might bin anything on one side or the other side of the graph, or perhaps they may not.)

Nothing says that you can't make that a 98.5% CI or even a 99.5% CI and virtually eliminate failure. This is what companies with excellent quality control may do and they usually have the reputation (and warranty service) to show for it. There's a reason that MSI, EVGA, and Gigabyte have such good reputations in the GPU industry, and it's their quality control, plain and simple.

If you go over to HWBot and check out the 3DMark entries for the 580 series, the top 20 scores are full of Lightnings, with a few scattered Asus Matrix & DCUII, Gigabyte SOC and such. Really dont see too many straight 580's like the zotacs and such or evga vanilla's...

As far as "binning" is concerned, it would have to depend on what you define as binning...

But it seems they either all the factory OC'd non reference cards clock higher, or benchers dont buy vanilla cards...

Could be either because these cards are "better" in the sense that they are on better PCB's with better components