NVIDIA says Fermi yields are fine, promises 512-core version

[Chumbucket843]

More then likely samples for reviews. That's part of the reason why I mentioned the 7900 GTX. Or can we include the 8800 Ultra?

there have been several cards from both ATi and nvidia that were the fastest of their time but virtually impossible to buy. the 512sp version would only offer ~6% more performance over a 480 at the same clocks. that's something to keep in mind. clockspeeds and power consumption are more important. these are the areas that can and need the most improvement.

[Manicdan]

wernt clockspeeds pretty good? faster than other nvidia cards

and if you take an 5870 and max it out at stock volts, and do the same with a 480, i think the 480 has alot more potential, and can help reduce alot of that perf/watt suckage people complain about.

[Manicdan]

woot for one click, double posting, lol

[Eastcoasthandle]

there have been several cards from both ATi and nvidia that were the fastest of their time but virtually impossible to buy. the 512sp version would only offer ~6% more performance over a 480 at the same clocks. that's something to keep in mind. clockspeeds and power consumption are more important. these are the areas that can and need the most improvement.

Hmm, no they weren't. Not in the context that you are replying to my post. The history regarding the 7900 GTX and 8800 Ultra is apparent.

[stinkas]

wernt clockspeeds pretty good? faster than other nvidia cards

and if you take an 5870 and max it out at stock volts, and do the same with a 480, i think the 480 has alot more potential, and can help reduce alot of that perf/watt suckage people complain about.

But does it need to be on water?
Water is a pain in the @ss and expensive...

[Olivon]

Drew Henry just tales truisms (Down with SKYMTL and annihilat0r) ... 'better said 768 CC on 28 nm ...

Nice to see Face again on XS

[Caparroz]

The architecture itself is designed with 512 cores. Saying there is a chance for a 512 core version is like saying the sky is blue...

wasnt the original yield number supposed to be based on higher clocks, the cores from the 480 or better and being lower voltage, so that 20% could well be true but they lowered the cores, lowered the clocks and upped the voltage so if u do any of those it would change yields.

not trying to be combative but when u change what u will use it greatly changes what your yields are.

More then likely samples for reviews. That's part of the reason why I mentioned the 7900 GTX. Or can we include the 8800 Ultra?

^^This.

[Sn0wm@n]

I love it how people keep blowing heat issues out of their asses. Unlocking all of the die won't do a thing to the cooling needed if yields are improving and they can do a better job cherry picking.

in order to improve yield they would need to tweak the process not wait out of luck.... + enabling the disabled core would raise the power draw.... duh.....

[annihilat0r]

in order to improve yield they would need to tweak the process not wait out of luck....

What? No. A process will increase in yields "by itself" with time.

[Sly Fox]

Yields so great they couldn't release the 512 version right off the bat!!!

Right Nvidia... right....

[HuffPCair]

Yields so great they couldn't release the 512 version right off the bat!!!

Right Nvidia... right....

Yeah just like how fast the 470 and 480 came out.

I am sure nvidia is on top of their shat

[Chumbucket843]

What? No. A process will increase in yields "by itself" with time.

its both but its hard to say which is more important. tsmc gives clients the tools to design a chip and then they manufacture it. being fabless doesnt mean that the process side is a free lunch either.

vias are being problematic. apparently ATi found a way around it.
http://www.semiconductor.net/article...ia_Defects.php

[tajoh111]

Yeah just like how fast the 470 and 480 came out.

I am sure nvidia is on top of their shat

I think the problem with 512 at this point is not only yields but additionally power requirements.

I.E. The 512 although using only 6 shaders, will use significanly more power than a 480(just look at the difference between the 470 and 480), additionally to justify the another product it will need higher clocks than the 480 itself, as a 6 percent difference hardly needs a new product.

512 shaders + 750mhz clock with the current process means a card that needs a waterblock and 2 8 pin power.

[Sn0wm@n]

What? No. A process will increase in yields "by itself" with time.

the yield are due in part to complexity + if the process helps... so its a yes

[HuffPCair]

I think the problem with 512 at this point is not only yields but additionally power requirements.

I.E. The 512 although using only 6 shaders, will use significanly more power than a 480(just look at the difference between the 470 and 480), additionally to justify the another product it will need higher clocks than the 480 itself, as a 6 percent difference hardly needs a new product.

512 shaders + 750mhz clock with the current process means a card that needs a waterblock and 2 8 pin power.

Yeah I agree. Am waiting to see how much BS they throw out in a few months.

[Starscream]

Yields so great they couldn't release the 512 version right off the bat!!!

Right Nvidia... right....

Dont forget they need to save something for the refresh.

Im shure the yields werent good enough for 512 at launch. But another reason for not launching a 512 version could be the refresh that at some point wil come.

[Apokalipse]

What? No. A process will increase in yields "by itself" with time.

I really don't think so.

If TSMC works to improve yields, yields will improve. Or if Nvidia tweaks the chip it might get better yields.

AMD improved yields by doubling the number of vias, and making some other tweaks to the design of Cypress (and of course the smaller die also naturally gets higher yields).

I don't see how it would magically improve by itself.

[LiquidReactor]

Is nVidia going to do a B1 respin? Kyle from [H] does not seem to think so, at least not anytime soon and I believe Charlie said the same thing. So, B1 respin in time for Christmas 2010 season

Just sticking to nVidias own schedule for Fermi here. Remember Fermi was supposed to be here "very soon" back in September 2009. "Very soon" in nVidia lingo equates out to 6-8 months in real world lingo

[Zonker]

Does anyone even know what this 'yield' number represents? Is it the percentage of chips that can be used as 480/470, or the percentage that can be used for 480/470/460?

[Apokalipse]

Does anyone even know what this 'yield' number represents? Is it the percentage of chips that can be used as 480/470, or the percentage that can be used for 480/470/460?

that's the thing, they could easily shift the goalposts just to say they have higher yields.

[Chumbucket843]

Is nVidia going to do a B1 respin? Kyle from [H] does not seem to think so, at least not anytime soon and I believe Charlie said the same thing. So, B1 respin in time for Christmas 2010 season

Just sticking to nVidias own schedule for Fermi here. Remember Fermi was supposed to be here "very soon" back in September 2009. "Very soon" in nVidia lingo equates out to 6-8 months in real world lingo

that would take 6 months... if they stayed on schedule. its best to spend more time on what they will be doing on 28nm.

[Dainas]

in order to improve yield they would need to tweak the process not wait out of luck.... + enabling the disabled core would raise the power draw.... duh.....

duh yourself

The thing is even ignoring tweaking refinements, not all cores are the same, even on the same wafer there can be big differences in TDP. Some dead, barely usable, some good, some very good and some excellent. Nvidia had the 480 configured so they could get numbers were not phantoms cards, this full 512 version looks like it could be a phantom. Creme of the creme these are the ultra rare perfect examples.

[BoredByLife]

Oh, come on. Which idiot thinks there will never be a 512 core Fermi? The GF100 architecture was designed for 512 cores. It will eventually have 512 cores. Or do you guys think Nvidia is going to abandon the Fermi architecture altogether because Charlie "told you so" that it was "unmanufacturable"?

To improve yields, you don't only need a new metal layer scheme, but a complete new scheme. Not just going from A4 to A5, but from A4 to B1, this costs a lot of time. So in my opinion, Nvidia better starts designing the upgrade from Fermi, because fixing it takes too much time.

[LordEC911]

I really don't think so.

If TSMC works to improve yields, yields will improve. Or if Nvidia tweaks the chip it might get better yields.

AMD improved yields by doubling the number of vias, and making some other tweaks to the design of Cypress (and of course the smaller die also naturally gets higher yields).

I don't see how it would magically improve by itself.

As the process matures, small tweaks are possible that will slightly increase yields, not much but a little.

Is nVidia going to do a B1 respin? Kyle from [H] does not seem to think so, at least not anytime soon and I believe Charlie said the same thing. So, B1 respin in time for Christmas 2010 season

Just sticking to nVidias own schedule for Fermi here. Remember Fermi was supposed to be here "very soon" back in September 2009. "Very soon" in nVidia lingo equates out to 6-8 months in real world lingo

Bx silicon seemed to be likely but that rumor changed from possibly seeing it in Q3, to Fermi II in Q4.

Does anyone even know what this 'yield' number represents? Is it the percentage of chips that can be used as 480/470, or the percentage that can be used for 480/470/460?

Yield, in this context, means total number of usable dice.
Those are then split into different bins, GTX480/470/460 depending on clocks/voltage/TDP/defects.

[Apokalipse]

As the process matures, small tweaks are possible that will slightly increase yields, not much but a little.

In other words, it doesn't happen by itself. It happens because tweaks are made as they use it and learn more about it.