View Single Post
Old 05-16-2012, 07:00 PM   #21
Tempered81
Diamond Member
 
Tempered81's Avatar
 
Join Date: Jan 2007
Posts: 6,358
Default

It doesn't mean much. They use a value found on each chip denoting what they call "quality" during the binning process. It really has no reflection on overclock relative to voltage & temperature, since people with 60% and 80% can clock higher than others with 71% and 84%, while some 78% and 68% are the bottom of the barrel. See what I mean, it has no relationship.

The final stable overclock of your board greatly depends on the quality of the board and the quality & quantity and spec of components used. Take 100 reference 580 boards, or 100 reference 7970 boards as an example. Last I checked gpu-z's ASIC quality check feature didn't even work for most Nvidia cards, while the number reported is useless for AMD cards. W1zzard however tell us something like lower = better overclocking with regular cooling while higher = better oc with better cooling.

So the chips are binned in complicated process that I don't have all the details on. I'm not an EE in charge of asic validations. Chips have a measured leakage and a temperature operational range, those values are measured and combined to form the hard-coded readable data for 'asic qual'. The chips are assigned a default voltage ID based on the results of leakage testing. For each product in AMD's case there is a range for which the voltages must be set. Here's where it starts getting tricky.

Say the TahitiXT 7970 board has a possible 4 default VID states on reference designed boards in a range of 1005-1180mv. It might not be exactly that, but it's close. Now 100 dies pop out of the wafer fully functional at the desired 925mhz within that range and the desired temperature & leakage range. The too leaky/too hot parts are discarded or binned into a lower product. Each of the 100 samples will overclock at a different speed using a different voltage releasing a different quantity of heat because of various quantities of leakage. The result may be that the worst one does 950 @ 1150mv near the maximum temperature range, while the best one does 1250 @ 1025mv near the bottom of the temperature range. The 'best' chip in the above example is considered "rare" and would be used for the dual-gpu/mobile segments where the temperature quals are more strict - just a higher bin.

But in our example, all 100 chips are planned to be used as 7970 products - lets say because they are all fully functional, and even worse or even better samples exist for different skus for the sake of the example. The unbelievable aspect (for me at least) is that each of these 100 samples are assigned a VID in the reverse order that you would expect them to be. I would think the rare cherry sample would have the 1005mv VID, but that is not the case. Contrarily that best rare cherry sample ships out the door with the HIGHEST VID. The VID choice is based mostly on the leakage & temperature metrics. The least leaky, coolest running most efficient chips are given a higher VID to fit a temperature range and to create consistency in the SKU. You might be asking why would they artificially limit something like this? The answer is because even the worst pieces of silicon still need to become 7970 SKUs to keep yield up.

So the binning process is going to measure a lot of things.
Speed. What mhz will it run while completing the gamut of tests (926-1450mhz+) ?
Heat. Is the temperature maintainable for the hottest dies assigned the lowest VIDs?
Leakage. Are the worst silicons still going to achieve the desired binning requirements to become 7970 products?
And tons of other more precise metrics that I do not understand.

Of all the 100 dies, and their measured quantities of leakage they will all clock to various heights but not all will be solely limited by leakage. Although leakage would be the primary deciding factor, because it is a measurement of efficiency and a poor sample usually requires more voltage to maintain a similar overclock as a lower leakage part, while putting out more heat. But the complex nature of the binning process to make a larger lot of SKUs means for some bass-ackwards methods of assigning default voltages. You also can't always say that your 7970 is a better or worse overclocker because it was given one of the 4 default voltage lookup tables that is more/less than another sample. The ASIC quality, together with the default voltage, and the known temperature of the card when overclocking all combine to give you some sort of relative measurement of its overclocking ability.

This is where the quality of the board comes into play. Each of the individual logics on the board are also binned just like the great big die is! The memory ASICs, the mosfets, chokes, inductors, master and slave buck controllers even each capacitor and resistor are all binned and specced when they are manufactured. There is a lottery involved when assembling the PCB as well. You would hope to win the lottery on the PCB & the die! A combination of the top 100 die sample & top 100 PCB sample results in the best video card overclocks that you see. A combination of the worst die and the worst board results in the lowest clocking samples. Say Shamino & Kingpin are overclocking 100 samples (these samples have to be 580's for KP but the same metrics apply). Kingpin's skills are going to be in use at each benching session. Say he isn't even using LN2 yet and for our examples sake, he's only going to measure each card's limit with its stock HSF.

Now each of the 100 GTX580s is going to yield a different result to the most consistent seasoned VGA overclocker in the world. Each GF110 die and reference GTX580 PCB combination is going to vary in quality with a number between 0-99. That end result is not wholly measured by gpu-z's ASIC quality number, even though it may be relative in one part of the binning process. I don't know off hand what a GTX580's air overclocking range is but let's just say 795-1050mhz. Evenly space out 100 different results within that range, and there you have it. One card will be the best clocker with that stock HSF, and the even stranger part, is that one particular best card might NOT be the best card for reaching 1500mhz on LN2. That might happen to be the 76th ranked card in the line-up. Which gives credence to the ASIC qual explanations from W1zzard.

Now you would be surprised to see that his best sample had a crazy ASIC quality number like 68%, and you expect that number to be 95-99%.
Tempered81 is offline   Reply With Quote