Overclocking the 7970

[rdqlus]

Silicon design engineer reporting in ...

Just wanted to make an educated guess post regarding the seemingly low default clocks on the 7970. It's very likely that 28nm has a large process variation across wafer lots. AMD is probably using a very aggressive screen to catch marginal parts which is why everyone is having pretty good success in overclocking to the max allowed by CCC. There will still be some crap that gets through that they can't easily screen for which is why the defaults are low.

AMD probably has a pile of parts sitting in the fab or test house that don't pass their aggressive screen but pass nearly everything else. Once they find a way to screen this second batch of parts to find the ones that are truly bad then theyll release them to the board manufacturers. This means there is probably a batch of boards coming down the pipe that won't overclock nearly as well as the first cards that came out. This is the reason for the low default clocks - AMD probably thinks that those are the frequencies at which they can meet their desired yield numbers. They just don't have their production testplan sorted out yet to bin their parts. So factory overclocked boards (probably way more than the current 1GHz cards) are coming as well as reference cards at 925MHz that won't overclock nearly as well as the Jan 9th cards.

 

[dragonstongue]

well, I know from my previous experiences with AMD card being a 4870 and now 6870, mine and many others that have owned them either are "just" able to hit CCC limits, or able to exceed them quite well. My 4870 could do roughly 20 faster then limits on core and 50 faster on memory then CCC allowed, my 6870 is running 965/1135 well below the "reviewed" averages of what it can do, but the reason to this is below.

all I am hoping for is, when ASUS and these guys release thier cards that can do "voltage tweak" that it actually works the way it should, and WILL change said voltages as an allowed range through the actual Vreg and not some stupid software trying to impose changes on it that does not allow the voltage to be controlled AT ALL.

I dont mind that it can declock the voltage or hit the top of its pre-programmed limit, but when they claim up to 50% more because of voltage tweaking ability, that never works becuse said circuit is not allowed to change its voltages above the pre-programed amount i.e 1.17-1.18v that is supposed to go up to 1.3v but CANT, well that just pisses me off

Anyways, the OC of these chips is probably like you mentioned above, get the best initial samples out there for now, the other ones will probably be used for 7950s or even for dual chip 7990s, but then again, after sampling them for sale for a couple of months, they will either be better able to clock them, or at least tweak thier design a hair and "nerf" the clocking ability and such, so us users will have to go with the factory "cherry picked" cards if we want to OC them well. Its hard to say, simply, no cards released up until now seen as near high clock as these seem to hit on "average"

 

[Porter_]

interesting info OP. it will be interesting to see how the next few waves of cards overclock.

 

[SIRTIM3]

Makes me want to get one right now, rather then wait for 7950's.

 

[Quartz-1]

It's very likely that 28nm has a large process variation across wafer lots.

Could you possibly explain and elaborate on the term 'process variation'?

 

[next-Jin]

God know how those 1337 cards are going to perform.

 

[rdqlus]

Could you possibly explain and elaborate on the term 'process variation'?

A simple explanation would be that there are 3 process corners - slow, typical and fast. Slow silicon is exactly that, slow. It generally won't run at as high a frequency and usually consumes less power due in part to lower leakage current. Fast silicon is the opposite and typical silicon lies somewhere in the middle. The fab (TSMC in this case) guarantees that production wafer lots will fall somewhere in the range between the fast and slow corners.

When a company tapes out a chip they often ask the fab for slow and fast wafer lots. The fab then purposefully skews the process to produce slow and fast parts. AMD then does a pile of testing to make sure that their part works across the temperature and voltage ranges in their spec for the chip. If you're lucky, usually the difficult corners are:
1) slow silicon at high temperature and low voltage (setup problems)
2) fast silicon at low temperature and high voltage (hold problems)

If you're unlucky and screwed something up you may end up with "islands" of non-functionality (example: typical silicon that doesn't work at room temperature but then starts working once it heats up .... this sucks, hard).

Anyways, when I mentioned process variation I was referring to the amount of difference between slow and fast silicon. As that gap gets wider it gets difficult to get chips to work across voltage and temperature. AMD might be getting around this by testing for the fast silicon parts and only shipping them out. If so, they'll have a pile of slow and typical parts that, while functionally fine, won't overclock as well as the fast silicon.

All I know for sure is that I'd love to sit in on some of their engineering meetings. It would be really interesting to know what exactly is going on.