[VC]NVIDIA GeForce GTX 980, GTX 980 SLI, GTX 970, 3DMark performance

[realibrad]

The likelihood of a defect forming is uniform across the wafer. With larger dies, you have a greater percentage of the dies containing a defect. The simple example is there's one defect: with 100 dies per wafer, you have a 1% failure rate. With 10 dies per wafer, you have a 10% failure rate.

I'm not sure what your point is. I have not taken a stance on big vs small, only that its a cost/benefit that the person did not have data on to make any statement on what should be done.

 

[3DVagabond]

So bigger more loosely packed chips is the way to go? What the hell are these companies doing then trying to shrink and make more dense their Arch's? /sarc

 

[Idontcare]

I won't even bother. Not sure that I can anyways. It's just a matter of common sense and physics. If there is a 100mm^2 die that has 1 billion transistors, and another 100mm^2 die that has 500 million transistors, both dies use the same ratio of transistor types, the die with more transistors per mm^2 is going to be prone to more defects.

This is true.

And it comes down to the probability of a defect causing yield loss or not causing yield loss, which comes down to the size of the defect and how much "free space" is available within the IC for random defects to exist without critically disturbing the circuits.

Have a less dense layout, less dense xtors and less dense wire routing, and your IC has more places for defects to exist which won't result in the chip becoming non-functional.

In fab parlance, you are always worried about your D0 (defect density) and D0 is particle size dependent as well.

The maximum size a defect can be before causing functional yield loss will be node-specific as it depends on the minimum poly (or fin) pitch and minimum metal pitch.

Being less aggressive on the pitches (less dense layout) will result in higher yields when you have a high D0. As your node matures and the D0 becomes less and less, the yield hit from the D0 will reduce and as such the benefits to having a less dense layout will also be reduced.

However, the reason I suspect you all are at odds here in this conversation is because it appears to me that neither party are speaking to the same issue.

The flip-side of the discussion, using tviceman's quotes example, is that you could have "100mm^2 die that has 1 billion transistors, and a 50mm^2 die that has 500 million transistors" (note the same xtor density now) which would result in approximately twice the number of chips per wafer.

100mm^2 = 640 chips per 300mm wafer
50mm^2 = 1319 chips per 300mm wafer (never perfectly 2x)

In this case, since both of these theoretical dies would have the same density the D0 impact is such that not only will the smaller die have higher functional yield (percentage-wise), but the total number of chips per wafer (NUBs or net units built) will be all the higher than that of the larger die.

Thus, in practice the only time you would intentionally layout your chip to have low density for the sake of functional yield is if you are expecting (because you start this process 2-3 yrs before taking the chip to production) the D0 for your production node and fab to be absurdly high.

But there are other reasons, for the sake of parametric yield, for which you would intentionally go about designing less dense circuits and ICs. Sram for L1/L2/L3$'s is but one very easy to identify example of this. Operating voltage is yet another, as well as heat dissipation and power consumption (leakage), etc.

 

[Skurge]

970 will be about U$450
980 will be about U$650

 

[FatherMurphy]

No way there is $200 between them, assuming the 970 is the same chip but with disabled bits and pieces. Not sure there is any precedent for that, either, but I could be wrong.

 

[jpiniero]

970 will be about U$450
980 will be about U$650

I'd say it's pretty likely it'll be $499 for the 980 and $399 for the 970. The 290X is already down to ~$490 and the 290 for ~$370. Obviously there will be a price preminum because it's nVidia; but there are limits to it. I imagine the 290 and 290X will just be cut a bit more.

The low TDP of the 980 would make for a nice "990" (980x2) if they decided to do that. Maybe later on.

 

[Skurge]

I'd say it's pretty likely it'll be $499 for the 980 and $399 for the 970. The 290X is already down to ~$490 and the 290 for ~$370. Obviously there will be a price preminum because it's nVidia; but there are limits to it. I imagine the 290 and 290X will just be cut a bit more.

The low TDP of the 980 would make for a nice "990" (980x2) if they decided to do that. Maybe later on.

I wasn't speculating.

If you look at the 290 and 290x speculation pages. I also gave pricing. Now pricing won't be right to the dollar, but it should be close and the gap between the two should be accurate.

 

[FatherMurphy]

So you are saying that there will absolutely be a $200 gap between GTX 980 and 970?

 

[Grooveriding]

I wasn't speculating.

If you look at the 290 and 290x speculation pages. I also gave pricing. Now pricing won't be right to the dollar, but it should be close and the gap between the two should be accurate.

Well that is interesting. I expect more performance than any of the leaks have been showing with those price points. $650 is 780ti pricing and I would guess if the card is taking that position it will be appreciably faster... I hope.

 

[DooKey]

Well that is interesting. I expect more performance than any of the leaks have been showing with those price points. $650 is 780ti pricing and I would guess if the card is taking that position it will be appreciably faster... I hope.

Yep, if he is correct or even close then the 980 will be at least 20% faster than a 780Ti IMO.

 

[Skurge]

I have no idea how fast it will be though, I can confirm it is 256bit and 4GB for both models.

The 970s will have customs at launch. Not sure about the 980 though.

 

[jpiniero]

I wasn't speculating.

If you look at the 290 and 290x speculation pages. I also gave pricing. Now pricing won't be right to the dollar, but it should be close and the gap between the two should be accurate.

Huh. AMD must be breathing a sigh of relief. The 290 looks a lot more viable given it's pricing, even if you would be giving up a lot on power draw.

 

[tviceman]

970 will be about U$450
980 will be about U$650

If true, disappointing on all levels.

 

[96Firebird]

I wasn't speculating.

If you look at the 290 and 290x speculation pages. I also gave pricing. Now pricing won't be right to the dollar, but it should be close and the gap between the two should be accurate.

290 Will be U$449, So I'm guess 290X will be U$549/U$599

Should I subtract the $50 now, or wait...

 

[alcoholbob]

970 will be about U$450
980 will be about U$650

This must be just temporary, no? The 980 is significantly cheaper to produce than the 780 (being a much smaller die), which retails at $499.

 

[S.H.O.D.A.N.]

Well that is interesting. I expect more performance than any of the leaks have been showing with those price points. $650 is 780ti pricing and I would guess if the card is taking that position it will be appreciably faster... I hope.

Based on what? I mean, other than common sense?

Nvidia recently released a 3kUSD card beaten by a 1.5k AMD offering. AMD is releasing the 285 at a price and performance point that doesn't seem to make any sense, even in relation to their own product, not even mentioning the competition.

Let's not delude ourselves with reasonable expectations. Best not to have expectations at all.

 

[tviceman]

This is true.

And it comes down to the probability of a defect causing yield loss or not causing yield loss, which comes down to the size of the defect and how much "free space" is available within the IC for random defects to exist without critically disturbing the circuits.

Have a less dense layout, less dense xtors and less dense wire routing, and your IC has more places for defects to exist which won't result in the chip becoming non-functional.

In fab parlance, you are always worried about your D0 (defect density) and D0 is particle size dependent as well.

The maximum size a defect can be before causing functional yield loss will be node-specific as it depends on the minimum poly (or fin) pitch and minimum metal pitch.

Being less aggressive on the pitches (less dense layout) will result in higher yields when you have a high D0. As your node matures and the D0 becomes less and less, the yield hit from the D0 will reduce and as such the benefits to having a less dense layout will also be reduced.

However, the reason I suspect you all are at odds here in this conversation is because it appears to me that neither party are speaking to the same issue.

The flip-side of the discussion, using tviceman's quotes example, is that you could have "100mm^2 die that has 1 billion transistors, and a 50mm^2 die that has 500 million transistors" (note the same xtor density now) which would result in approximately twice the number of chips per wafer.

100mm^2 = 640 chips per 300mm wafer
50mm^2 = 1319 chips per 300mm wafer (never perfectly 2x)

In this case, since both of these theoretical dies would have the same density the D0 impact is such that not only will the smaller die have higher functional yield (percentage-wise), but the total number of chips per wafer (NUBs or net units built) will be all the higher than that of the larger die.

Thus, in practice the only time you would intentionally layout your chip to have low density for the sake of functional yield is if you are expecting (because you start this process 2-3 yrs before taking the chip to production) the D0 for your production node and fab to be absurdly high.

But there are other reasons, for the sake of parametric yield, for which you would intentionally go about designing less dense circuits and ICs. Sram for L1/L2/L3$'s is but one very easy to identify example of this. Operating voltage is yet another, as well as heat dissipation and power consumption (leakage), etc.

Thanks so much for this extremely informative post! It sounds like density AND die size both have affects on overall yields.

 

[Silverforce11]

Thanks so much for this extremely informative post! It sounds like density AND die size both have affects on overall yields.

From AMD's stand point and what I saw from their engineers saying on public, a denser design for them is all about reducing the die size to keep costs down, the downside that they mentioned is more leakage resulting in higher power use.

Thus, its expected that 28nm at TSMC is very high yielding that errors wouldn't be a concern for them to go with this approach.

As to other people commenting about the absurd price at >$600, NV doesn't give a shat about "oh, its a smaller die, it costs us less therefore we should price it less", that's purely wishful thinking from a consumer point of view. To them, IF they can get away with $650 for a cheaper mid-range maxwell than selling a full blown kepler, its more profit! Profit = good.

Currently there's no reason for NV to price it any less than the 780ti if it has similar performance. Why should they when 780 and 780ti sell just fine? This is a refresh, you get similar performance at less power use, at similar prices. Unless AMD puts out pressure on the top end, they have zero incentive to price it reasonable for consumers.

At this junction, I'd like to remind all the loyal fans from both sides, you never wish total dominance of one company over another UNLESS you are not a consumer but a shareholder. As a consumer, you should want strong competition, always. Else you get shafted for price and lack of innovation or stagnation.

 

[DominionSeraph]

This must be just temporary, no? The 980 is significantly cheaper to produce than the 780 (being a much smaller die), which retails at $499.

The Sandy, Ivy, and Haswell i7's are all a much smaller die than the FX-4300. You don't see Intel selling them for $100.

The last price war was brought on by Nvidia after the GTX 460 launch. A few months after launch a 768MB GTX 460 (launch $200) could be had for $90. After that, prices actually went up, and since then both parties seem content to let each new halo release push up the price of the high end, reducing the rate of depreciation for everything else. 2.5 years after launch the GK104 is still selling for more than the Fermi GF104 did on the day of release.

AMD didn't have to price gouge on the 5000 series after launch. They did.
Nvidia didn't have to release the GTX 680 at $500. They did.
Having the ability to reduce prices says nothing about whether they will.

 

[3DVagabond]

At this junction, I'd like to remind all the loyal fans from both sides, you never wish total dominance of one company over another UNLESS you are not a consumer but a shareholder. As a consumer, you should want strong competition, always. Else you get shafted for price and lack of innovation or stagnation.

With all of the speculation going on I would just like to say that ^^^THIS^^^ is not speculation, but fact. We should all want both companies to compete and push to make the best GPU's they can.

 

[Borealis7]

so apple just announced the A8 CPU, 89mm^2, 2B transistors @ 20nm.
that's what we get instead of better GPUs.

i guess more people want to game on a 4.7" 1334x750 screen than on a 4K monitor.

damn't Apple, let other companies get 20nm chips too!

 

[Skurge]

Should I subtract the $50 now, or wait...

... Hold me to it.

 

[RussianSensation]

Based on what? I mean, other than common sense?

Nvidia recently released a 3kUSD card beaten by a 1.5k AMD offering. AMD is releasing the 285 at a price and performance point that doesn't seem to make any sense, even in relation to their own product, not even mentioning the competition.

Let's not delude ourselves with reasonable expectations. Best not to have expectations at all.

That's true. NV can easily charge $100 more for similar performance to AMD ($399 260 vs. $299 4870; $360-380 770 4GB vs. $260-280 280X; $450 680 vs. $350 7970 Ghz; $450-550 780 vs. $350-450 R9 290).

If we go by that logic NV can release a card 10% faster than 780Ti and price it at $699 all day long since: (1) it's faster, (2) it uses less power, (3) it has better features - 4K decode, 4GB VRAM.

However, if NV prices it at $499-549 it will get a large wave of 580/680 users upgrading. At $649-699 many will feel too ripped off paying that much for a midrange Maxwell. As far as I am concerned GM204 replaces GK104 so $499 for 4GB and $549 8GB is a direct replacement of 680 2/4GB. In my eyes to justify > $549, 980 needs to beat 780Ti by more than 680 beat 580 (so > 35%).

Otherwise, NV can release every next fastest card at $699 forever, as long as it's faster than the preceding card. But that's not how the GPU industry works. By now we should get 690 level of performance for half.

 

[tviceman]

IMO, anything other than $499 will be a disappointment if its only 10% faster. If its 20% faster, then anything more than $549 will be a disappointment.

Personally, I think it'll be 10% faster and $549. The gtx 970 will be $449, and the gtx 970 will be $349, and GK104 will get rebadged as the gtx 960 at $249.

 

[wand3r3r]

970 will be about U$450
980 will be about U$650

Very disappointing NV (if true)! :thumbsdown:

560 ti 2nd gen at $650? What a joke. NV is ridiculous, I won't be surprised if it's slower than the 780 ti and still $650, based on the titan-z the worlds 2nd fastest GPU for 2-3x the price. When is the price gouging going to stop? I won't be supporting this.

NV fans better prepare for $1500 titans in a year, followed by $1300 high end (780 ti successor).