Opinions on Nvidia's new GPU binning system?

[cbn]

http://www.anandtech.com/show/3809/nvidias-geforce-gtx-460-the-200-king/16

As we’ve discussed in previous articles, with the Fermi family GPUs no longer are binned for operation at a single voltage, rather they’re assigned whatever level of voltage is required for them to operate at the desired clockspeeds. As a result any two otherwise identical cards can have a different core voltage, which muddies the situation some.

Does anyone have opinions good or bad on Nvidia's new GPU binning system?

I personally think this was great innovation and a true progression towards an individual knowing the exact quality of the chip he or she has. I only wish Nvidia would stamp the voltage/frequency certification on the heatspreader as well.

 

[taltamir]

I personally think this was great innovation and a true progression towards an individual knowing the exact quality of the chip he or she has.

That isn't what I see.
it used to be that you could just look up the card model (say, GTX260), and be told the voltage, since all GTX260 cards ran in the same voltage.
This is no longer true. Now you buy it and it can be ANY voltage in a wide range. the only way to tell is if you use a tool (like gpuz) to see what voltage is the currently installed card. You could do that before as well.

 

[GaiaHunter]

It isn't that great IMO.

The only reason they do it is to increase yields - basically cards that would be trash or used to make cheaper SKU are now OC from factory to reach stock clocks!

Before your cards could OC more or OC less, but at least you knew what was the expected power consumption.

Now you don't.

And a card that requires extra voltage compared to another to reach stock speeds probably won't be a great overclocker.

 

[Sylvanas]

This is probably due solely to improving yields. I would say that they can't get to where they'd like to be on their 40nm process anytime soon, but they can salvage that by selling the same physical chip that may require a higher voltage to operate- not ideal, but better for the bottom line and getting the product to market for a hard launch.

I think it's great for us overclockers. Some may get lucky with a lower VID which *might* lead to more headroom (relative to other cards and higher VIDs) it appears the 460 does not support voltage adjustments via the software just yet::

OnSemi's NCP5388 is a reasonable priced voltage regulator, unfortunately it does not support I2C interface for voltage control. NVIDIA however exposes an API for voltage changes via VID in their NVAPI.

It's up to developers to add support, I'm sure one of Unwinder's utilities will get it going soon.

 

[taltamir]

It isn't that great IMO.

The only reason they do it is to increase yields - basically cards that would be trash or used to make cheaper SKU are now OC from factory to reach stock clocks!

Before your cards could OC more or OC less, but at least you knew what was the expected power consumption.

Now you don't.

And a card that requires extra voltage compared to another to reach stock speeds probably won't be a great overclocker.

correct on every single issue.

I think it's great for us overclockers. Some may get lucky with a lower VID which *might* lead to more headroom. it appears the 460 does not support voltage adjustments via the software just yet::
It's up to developers to add support, I'm sure one of Unwinder's utilities will get it going soon.

I don't understand, why? You could still get an excellent clocking chip before, it would just be set to a higher stock voltage. I had a friend who scored an amazing 8800GTS which overclocked higher then the 8800 ultra (I know, insane!)... if anything, I think its worse for overclockers, because chips that used to not cut it are now mixed into the supply with a higher default voltage. So your chances of getting a great overclocker are lower.

EX: lets say you have 3 chips, chip A requires 0.9v to run at stock speed, chip B requires 1.0v, and chip C requires 1.1v.
used to be, nvidia would set the voltage on chip A and B to 1.0v, stock speed, and sell them to you. they consumed the same amount of power, and if you lucked out you got chip A.
Chip C would be clocked lower, and sold for less money as a lower end card.

Now nvidia simply sets chip A to run at 0.9v by default, chip B to be at 1.0v, and chip C at 1.1v, and sells all three under the same brand name. You can still luck out and get chip A, but it will OC exactly the same as if you would have gotten it under the other system. You can't know the exact power consumption before-hand... if you don't OC at all, chip A would be slightly more power efficient (since its at a lower voltage)... your chances of getting chip A went down, from 50% to 33% in this example. And you might be paying full price for chip C, which would have sold for less money before, and is a terrible overclocker.

overall the consumer loses under the new system.

 

[sandorski]

Probably just a temporary fix until TSMC gets their act together. If it's a permanent policy, it will have negative Rep repercussions.

 

[Sylvanas]

I don't understand, why? You could still get an excellent clocking chip before, it would just be set to a higher stock voltage. I had a friend who scored an amazing 8800GTS which overclocked higher then the 8800 ultra (I know, insane!)... if anything, I think its worse for overclockers, because chips that used to not cut it are now mixed into the supply with a higher default voltage. So your chances of getting a great overclocker are lower.

Well the nature of overclocking is still the same, it's hit and miss. But we may end up with a situation similar to what happened with the Q6600 G0 stepping (there was a better XS thread, can't seem to find it). Whereby some people with CPUs with a lower VID were able to achieve higher clocks at a given voltage level than those who started out with a CPU whose default VID was higher to begin with.

Yes, it would not be ideal if you got a card with a high stock VID to begin with but that does not ultimately determine the maximum OC, there are more variables than that. So I can see the situation my not be ideal to everyone, but from Nvidia's point of view, it's better to get the cards on the shelves and into the hands of people sooner rather than a shortages and a soft launch like what happened with the 470s/80's.

EDIT: I amended my original post with " ...more headroom relative to other cards and higher VIDs"
Damn forum maintenance

 

[Sylvanas]

EX: lets say you have 3 chips, chip A requires 0.9v to run at stock speed, chip B requires 1.0v, and chip C requires 1.1v.
used to be, nvidia would set the voltage on chip A and B to 1.0v, stock speed, and sell them to you. they consumed the same amount of power, and if you lucked out you got chip A.
Chip C would be clocked lower, and sold for less money as a lower end card.

Now nvidia simply sets chip A to run at 0.9v by default, chip B to be at 1.0v, and chip C at 1.1v, and sells all three under the same brand name. You can still luck out and get chip A, but it will OC exactly the same as if you would have gotten it under the other system. You can't know the exact power consumption before-hand... if you don't OC at all, chip A would be slightly more power efficient (since its at a lower voltage)... your chances of getting chip A went down, from 50% to 33% in this example. And you might be paying full price for chip C, which would have sold for less money before, and is a terrible overclocker.

With regard to your example, the maximum headroom depends on whether you want to adjust the voltage when OCing. I doubt chip A and C will achieve the same maximum frequency at their respective VIDs. Chip C, at a higher stock VID, may in fact be good for a range of +80 above stock, Chip A at stock frequency may be a lower power chip, but it may only have +40mhz headroom over stock frequency as it is already reaching it's limits as superior silicon. If you wanted to adjust voltages, and bump chip A to 1.0v then you may be good for another +70 on what you already achieved. As always, it depends

 

[cbn]

I don't understand, why? You could still get an excellent clocking chip before, it would just be set to a higher stock voltage. I had a friend who scored an amazing 8800GTS which overclocked higher then the 8800 ultra (I know, insane!)... if anything, I think its worse for overclockers, because chips that used to not cut it are now mixed into the supply with a higher default voltage. So your chances of getting a great overclocker are lower.

EX: lets say you have 3 chips, chip A requires 0.9v to run at stock speed, chip B requires 1.0v, and chip C requires 1.1v.
used to be, nvidia would set the voltage on chip A and B to 1.0v, stock speed, and sell them to you. they consumed the same amount of power, and if you lucked out you got chip A.
Chip C would be clocked lower, and sold for less money as a lower end card.

Now nvidia simply sets chip A to run at 0.9v by default, chip B to be at 1.0v, and chip C at 1.1v, and sells all three under the same brand name. You can still luck out and get chip A, but it will OC exactly the same as if you would have gotten it under the other system. You can't know the exact power consumption before-hand... if you don't OC at all, chip A would be slightly more power efficient (since its at a lower voltage)... your chances of getting chip A went down, from 50% to 33% in this example. And you might be paying full price for chip C, which would have sold for less money before, and is a terrible overclocker.

overall the consumer loses under the new system.

Wouldn't a system you are proposing with three tiers of chips (ie, .9v, 1.0v, 1.1v) lower average prices compared to a system selling only two tiers of chips (.9v, 1.0v)?

By definition the bin is now larger right?

 

[Phynaz]

This is standard practice in CPU's, I would expect it in GPU's.

 

[Sylvanas]

Wouldn't a system you are proposing with three tiers of chips (ie, .9v, 1.0v, 1.1v) lower average prices compared to a system selling only two tiers of chips (.9v, 1.0v)?

By definition the bin is now larger right?

One would hope so. Supply and demand, inventory is now larger to meet initial demand- but it's up to Nvidia if they choose to pass that onto the consumer.

 

[cbn]

With regard to your example, the maximum headroom depends on whether you want to adjust the voltage when OCing. I doubt chip A and C will achieve the same maximum frequency at their respective VIDs. Chip C, at a higher stock VID, may in fact be good for a range of +80 above stock, Chip A at stock frequency may be a lower power chip, but it may only have +40mhz headroom over stock frequency as it is already reaching it's limits as superior silicon. If you wanted to adjust voltages, and bump chip A to 1.0v then you may be good for another +70 on what you already achieved. As always, it depends

You are saying just because a certain 460 GTX has a low VID, that doesn't necessarily mean it will have a higher overclock ceiling right?

Hmmmm.....That is pretty interesting. Is this what happens with Mobile GPUs and CPUs? Certain chips that make great low voltage bins wouldn't necessarily scale well at desktop levels?

 

[HendrixFan]

This is standard practice in CPU's, I would expect it in GPU's.

Which CPUs have different voltages for the same model number? I've never heard of that.

 

[GaiaHunter]

The thing Sylvanas is that is actually worse now (for OC potential if no voltages tools available) for the better chips.

In taltamir example, the A chip, under old policy would come with a "voltage bump" so it would typically OC more compared to the same chip under the new policy (without bumping voltage).

The chip B would be the same and the chip C would be absent under old policy-

Look at that Q6600 GO example - it is exactly the same.

The chip A at 1.0v would OC more than chip B at 1.0v since it actually only requires 0.9v to run stable at stock speed. But now it appears at 0.9v so you will have to bump up it voltage to have the same OC headroom has it had before at 1.0v. If good voltage tools appear, this new policy is the same for chip A OC and better for people that don't OC the chip A.

But it is much worse for anyone that gets a chip C, chip D, etc.

 

[HendrixFan]

Hmmmm.....That is pretty interesting. Is this what happens with Mobile GPUs and CPUs? Certain chips that make great low voltage bins wouldn't necessarily scale well at desktop levels?

Going back some years I seem to recall the lower voltage chips being great overclockers, but I haven't stayed on top of those kinds of things to say if it is still the same.

You have two chips that run the same frequency but one can do it with less power, it is likely to be a better overclocker. This is assuming they are getting good yields and the majority of the low voltage bins are full of chips that run better than expected. Given the supposed poor yields from the Fermi chips, it is likely that they aren't binning low voltage chips as much as they are pumping up the voltage on chips they would otherwise throw out so they can put cards on the market.

 

[GaiaHunter]

Hmmmm.....That is pretty interesting. Is this what happens with Mobile GPUs and CPUs? Certain chips that make great low voltage bins wouldn't necessarily scale well at desktop levels?

Well the Mobile GPUs/CPUs don't operate at same stock speeds.

The baseline would be - chips that take lower voltage to reach a given speed have more OC room.

 

[cbn]

In taltamir example, the A chip, under old policy would come with a "voltage bump" so it would typically OC more compared to the same chip under the new policy.

I get what you are saying. If all chips with the bin are set to 1.1 volts default that would mean the "A chips" only needing .9 actual volts to run would be much better stock voltage overclockers.

However, lets assume, voltage tools become available for 460 GTX. Wouldn't you want to know what a chips actual voltage rating was rather than going through all the hassle of testing each chip to see what the best overclocker was? People used to do this with Opteron 165 lots, but to me that seems like a real hassle and far less accurate to boot.

 

[cbn]

Going back some years I seem to recall the lower voltage chips being great overclockers, but I haven't stayed on top of those kinds of things to say if it is still the same.

You have two chips that run the same frequency but one can do it with less power, it is likely to be a better overclocker. This is assuming they are getting good yields and the majority of the low voltage bins are full of chips that run better than expected. Given the supposed poor yields from the Fermi chips, it is likely that they aren't binning low voltage chips as much as they are pumping up the voltage on chips they would otherwise throw out so they can put cards on the market.

Yep, that is what I have always thought. It's just that I thought Sylvanas was saying something different. Maybe he was trying to explain a caveat to me? If so, I am all ears.

 

[GaiaHunter]

I get what you are saying. If all chips with the bin are set to 1.1 volts default that would mean the "A chips" only needing .9 actual volts to run would be much better stock voltage overclockers.

However, lets assume, voltage tools become available for 460 GTX. Wouldn't you want to know what a chips actual voltage rating was rather than going through all the hassle of testing each chip to see what the best overclocker was? People used to do this with Opteron 165 lots, but to me that seems like a real hassle and far less accurate to boot.

A few thoughts.

1) It's also a lot easier for the AIB partners to know - look at the amount of OC GTX 460 models coming out already. Maybe it is because the partners had these chips for a long time and were just waiting for GF100 inventories to go down. Or maybe it is because it is a lot simpler. Or both.

2) You have to have a cooperative retailer for that.

3) While this might bring the prices down, the GTX 460 chip pool is also worse. It will be interesting to hear for the next few weeks if people can achieve those OCs (and power consumptions/temperature/noise at stock) we saw in the reviews with regular GTX 470 as opposed to special OC'ed editions.

 

[Sylvanas]

You are saying just because a certain 460 GTX has a low VID, that doesn't necessarily mean it will have a higher overclock ceiling right?

Hmmmm.....That is pretty interesting. Is this what happens with Mobile GPUs and CPUs? Certain chips that make great low voltage bins wouldn't necessarily scale well at desktop levels?

Yes. At a given voltage a CPU/GPU has a range of operation (freq) that it can tolerate without producing errors or becoming unstable due to minute defects in the silicon at a molecular level. Two GPUs both at 1.0v will usually not have an identical maximum OC frequency because of the subtle differences in their fabrication. Lets say stock frequency is 100mhz. GPU 1's range might be +30mhz above stock and GPU2 might be only 10mhz above stock. Both GPUs qualify for sale as they are both capable of 100mhz at 1.0v, but each is subtly different and getting GPU1 is ideal. But, change that to 1.1v and the ranges may be +50 and +60 respectively (perhaps an embellishment but you get the idea) as the situation changes due to the nature of the silicon and it's changed tolerances and properties (Electromigration may play role aswell).

@GaiaHunter
And you would be correct in your example. The policy change is good and bad, good with regard to more availability/ lower prices and bad with regard to lower ceilings on lower default VID cards provided there remains no voltage tuning support. We also assume that higher VID=more ceiling and lower VID= less, but this is not a practical rule and is only a general guideline that was observed with Q6600's. That will inevitably change as MSI Afterburner and EVGA Precision gain support so it's a moot point really.

 

[Voo]

EX: lets say you have 3 chips, chip A requires 0.9v to run at stock speed, chip B requires 1.0v, and chip C requires 1.1v.
used to be, nvidia would set the voltage on chip A and B to 1.0v, stock speed, and sell them to you. they consumed the same amount of power, and if you lucked out you got chip A.
Chip C would be clocked lower, and sold for less money as a lower end card.

Or they could just run chips A and B at 1.1v and sell all three as whatever they want, which is probably what they did already. I mean why wouldn't they?

Don't see much difference to before, only thing is, if you're lucky you'll end up with a lower voltage chip that's a bit cooler and needs less power. But it's not as if all chips OCed equally good beforehand.

 

[Nox51]

Between variable VIDs the chips are using, the disabled clusters on the chips themselves and, from the anandtech article, a soft launch/delayed availability* for the 1gb card I think that Nvidia is still having major yield issues. That is somewhat of a concern regarding both the architectural design and TSMC's ability to deliver on the process.


* I am not entirely familiar with US retail availability so I am just going by the article.

 

[3DVagabond]

Probably just a temporary fix until TSMC gets their act together. If it's a permanent policy, it will have negative Rep repercussions.

TSMC doesn't only make chips for NV. Only NV seems to resort to these solutions with their product.

 

[Lonyo]

TSMC doesn't only make chips for NV. Only NV seems to resort to these solutions with their product.

Intel "resort" to these solutions too, and they have much more binning.
To me it seems like a sensible move.
The only people who can really complain are overclockers, but then
1) When you get the chip you have a better idea what it's going to be like because you have an accurate voltage
2) These chips are going to be factory overclocked so you could get one of them instead.

Really the only difference is possibly a higher chance of a "dud" (one that requires high volts for stock), but then that's quite likely anyway since partners will be selecting the best chips for their overclocker models, and overclocking is always YMMV anyway.

 

[Cookie Monster]

Which CPUs have different voltages for the same model number? I've never heard of that.

Think almost all CPUs have different "stock" voltage levels, and instead have a defined voltage range instead of a single fixed voltage value. For instance my Q9550 has a VID of 1.1375V. This is the voltage identifier that tells the motherboard VRM (via BIOs I presume?) that this amount of voltage is required to run that processor at the desired clock frequency.

So theoretically, the lower the VID the better since it gives you a bigger headroom.