AMD: Southern Islands (7*** series) Speculation/ Rumour Thread

Because you have to and because you can?

One thing I don't get is the relationship between threshold voltage variability and leakage. I'm not so sure that higher threshold voltages necessarily correspond to higher leakage. After all, leakage becomes a bigger problem as processes target narrower voltage swings between transistor on and off states to reduce dynamic power consumption.
 
IHVs do not run chips at higher voltage unnecessarily just to reduce power consumption variability. What possible benefit does that have as long as all chips come in under the rated TDP?
Chips don't come at a specified TDP they come a distribution of different speeds and leakages. The point of binning is to make sensible products with these varing parameters. Perviously GPU's used to have a single voltage per SKU, so that meant there was a high TDP variation between the highest and the lowest leakage part. By increasing the voltage on high leakage parts and decreasing the voltage on lower leakage parts all you are doing it incrasing that variability. If the high leakage/high voltage parts are running the maximum TDP then you are leaving a lot of power on the table for the the low leakage/low voltage parts and likely performance.

By running low leakage at high voltage and high leakage at low voltage you are narrowing the power band that they can operate in and increasing the overall maximum clock speed for the available TDP.
 
I have no idea what you're trying to say. All 7970's run at the same clock speed and therefore have the same performance. Increasing voltages unnecessarily does nothing to improve performance - all chips are still shipped at 925Mhz. When you say "leaves performance on the table" what are you talking about? Increasing the voltage of a high quality 925Mhz Tahiti die doesn't make it faster than a lower quality 925Mhz chip.

Maximum clock speeds are determined by voltage, not total power consumption. By your logic AMD should ship all 7970's at the same voltage to guarantee some maximum overclock. Obviously they don't do so.

I suspect you're confusing attempts to minimize manufacturing variances (Vt, leakage) with some imagined need to reduce measured TDP variation within an SKU at retail. It would be downright stupid to run a Tahiti chip at a higher than required voltage just to artificially increase its power consumption. How does that benefit either the manufacturer or consumer?
 
Actually, it does the opposite! We scale the voltage based on leakage, so the higher leakage parts use lower voltage and the lower leakage parts use a higher voltage - what this is does narrow the entire TDP range of the product.

Everything is qualified at worst case anyway; all the TDP calcs and the fan settings are completed on the wors case for the product range.

http://forum.beyond3d.com/showpost.php?p=1344008&postcount=29

No, the nominal voltage is 1.165V, however the higher leakage parts do not need as much voltage to run at those speeds.

http://forum.beyond3d.com/showpost.php?p=1344083&postcount=33
 
I have no idea what you're trying to say. All 7970's run at the same clock speed and therefore have the same performance. Increasing voltages unnecessarily does nothing to improve performance - all chips are still shipped at 925Mhz. When you say "leaves performance on the table" what are you talking about? Increasing the voltage of a high quality 925Mhz Tahiti die doesn't make it faster than a lower quality 925Mhz chip.

I suspect you're confusing attempts to minimize manufacturing variances (Vt, leakage) with some imagined need to reduce measured TDP variation within an SKU at retail. It would be downright stupid to run a Tahiti chip at a higher than required voltage just to artificially increase its power consumption. How does that benefit either the manufacturer or consumer?

Actually they're not, ASUS is already shipping 1GHz clocked ones (DCII, listed at least in Germany)
 

Thanks, so that answers my earlier question. Higher leakage parts require lower voltage to achieve the same frequency. So Speccy is right and AMD is artificially increasing voltages on lower leakage parts (would love to know why)

Also, we've seen the dual GPU cards from both vendors use chips binned for very low Vt. Does that mean dual GPU boards actually use the leakiest chips?
 
Interesting, so the range is roughly from 55% to just under 90% for people in that thread.

Makes me wonder if the 55-70% chips will start disappearing from the 7970 lineup once the 7950 launches. Those chips instead being used to fullfill demand for 7950 if they can't harvest enough chips to meet demand.

Regards,
SB
 
Actually they're not, ASUS is already shipping 1GHz clocked ones (DCII, listed at least in Germany)

That's an AIB overclocked part. We're talking about voltage binning of stock 7970's.

I would really like to understand how artificially narrowing the TDP band by over-volting low leakage chips benefits AMD or the consumer though.
 
Thanks, so that answers my earlier question. Higher leakage parts require lower voltage to achieve the same frequency. So Speccy is right and AMD is artificially increasing voltages on lower leakage parts (would love to know why)

AFAIK the higher leakage transistors switch faster than the low leakage transistors at the same voltage, this is why you want higher voltage in the low leakage transistors, to make them switch more or less at the same speed than the higher leakage transistors at lower voltage, and this way you can get a higher base frequency for your SKU.
 
trinibwoy said:
I would really like to understand how artificially narrowing the TDP band by over-volting low leakage chips benefits AMD or the consumer though.
Not only the TDP: by setting the voltage in function of the leakage (and silicon speed!), you also tighten the speed variation.

If you first tune the transistor speed to a certain desired point (by tuning voltage) and then do your chip qualification, you don't need to worry as much about some critical paths of slow silicon being dangerously close to not hitting the spec. It removes one dimension from the set of parameters to qualify for.
 
Thanks, guys. So it's not an artificial voltage increase at all - it's actually required for lower leakage parts to hit the target frequencies. Makes complete sense now!
 
Thanks, guys. So it's not an artificial voltage increase at all - it's actually required for lower leakage parts to hit the target frequencies. Makes complete sense now!

And this way you differentiate SKUs based on the active/disabled units and not based on the frequency, so you have less SKUs. Imagine that you need four SKUs with all the active units, four SKUs with one part disabled... this would create lots of confusion on the users.
 
One thing I don't get is the relationship between threshold voltage variability and leakage. I'm not so sure that higher threshold voltages necessarily correspond to higher leakage. After all, leakage becomes a bigger problem as processes target narrower voltage swings between transistor on and off states to reduce dynamic power consumption.

Leakage increases exponentialy as you lower threshold voltage.
 
Maximum clock speeds are determined by voltage, not total power consumption. By your logic AMD should ship all 7970's at the same voltage to guarantee some maximum overclock. Obviously they don't do so.
GPU clocks are rarely circuit limited (ie, with critical path, voltage etc.) these days. They are power limited.
 
Also, we've seen the dual GPU cards from both vendors use chips binned for very low Vt. Does that mean dual GPU boards actually use the leakiest chips?

My guess would be that the chips selected for dual cards are those that can hit high clocks with low Vt, so they have room for undervolting them.
 
Thanks, so that answers my earlier question. Higher leakage parts require lower voltage to achieve the same frequency. So Speccy is right and AMD is artificially increasing voltages on lower leakage parts (would love to know why)
As stated earlier, it is likely the highest voltage is the "nominal" voltage and the process is fine to operate at that level, meaning the low voltage parts have a lot of voltage headroom to OC with and with them being fast/leaky parts they should behave very well with it.

Also, we've seen the dual GPU cards from both vendors use chips binned for very low Vt. Does that mean dual GPU boards actually use the leakiest chips?
These are the most power bound to reach their specs. Likely they use the "rarest" of parts that are both relatively fast (so they don't need high voltage) and relatively low leakage. In other chips in the stack these types of parts go to the notebook SKU's but, commonly at these high end chips, no notebook target they can make dual chips boards with such parts.
 
Thanks, so that answers my earlier question. Higher leakage parts require lower voltage to achieve the same frequency. So Speccy is right and AMD is artificially increasing voltages on lower leakage parts (would love to know why)
Straight from the horse's mouth (wrt to Cypress back then, tbh):
http://forum.beyond3d.com/showpost.php?p=1344008&postcount=29
Actually, it does the opposite! We scale the voltage based on leakage, so the higher leakage parts use lower voltage and the lower leakage parts use a higher voltage - what this is does narrow the entire TDP range of the product.

Everything is qualified at worst case anyway; all the TDP calcs and the fan settings are completed on the wors case for the product range.

edit: damn, to late. :)
 
Last edited by a moderator:
lower leakege chips use more voltages.. higher leakage chips use lower voltage so that it can clock higher..

1.0250V is high leakege chips and they can be overclocked better..

So what I gather is that similar voltage requirements for the 5000 series still apply for the 7000 series (higher voltage for lower leakage parts). But does that always mean that such a card will do the highest OC that CCC will allow without 3rd party utilities?

Edit:
I just realized there is an update to GPUz which provides ASIC quality now.
http://www.techpowerup.com/159098/TechPowerUp-GPU-Z-0.5.8-Released.html
 
Last edited by a moderator:
So what I gather is that similar voltage requirements for the 5000 series still apply for the 7000 series (higher voltage for lower leakage parts). But does that always mean that such a card will do the highest OC that CCC will allow without 3rd party utilities?

It's generally yes it does but not everytime.. Leakage is not the whole thing it's just a part of it.. Sometimes higher VID CPUs clock better than lower VID equivalents.. It's rare but not impossible
 
Back
Top