Nvidia GTX 690 = 2 x Nvidia GTX 680!!

[RussianSensation]

Sell the chips with 2 faulty clusters? They probably have some, but I have no idea what % of their dies fall into that category. Perhaps it's not enough to launch a mid-range product or they're still saving up dies and selling their functional high-end chips to make the most money possible.

I may be 100% wrong, but I still think GK104 was meant to be GTX670/GTX670Ti. It was never meant to have many faulty clusters or to be a flagship. Imagine it's like asking NV to make 2 more smaller SKUs out of a GTX460? That's why it's only 30% faster than GTX580, has gimped GPGPU performance, static scheduling, 195W TDP. It's not 50-75% faster as we would expect from a new generation of NV chip. It only has 256-bit bus with same memory bandwidth as the 580 which is also not characteristic of next generation flagships. Also, why would NV be "relieved" when HD7970 launched since they haven't felt the pressure to have the single fastest GPU since 2006 when 8800GTX launched?

Perhaps, early on NV knew that they couldn't get GK110 out on time and last fall decided to use GTX670/670Ti to compete with HD7900 series. When HD7900 series launched, they were relieved since they realized the card they intended to sell at $349-399 could be pushed for $499. They re-badged GTX680 last minute. Since GTX680 was meant to be upper mid-range, it's no wonder there was no GTX660Ti/670 to launch alongside 680. It appears now NV is going to create 670 SKU much like GTX460 1GB (GTX680) --> GTX460 768mb (GTX670). However, this becomes problematic for 660Ti.

Based on that, where are they going to get bazillion faulty GK104 GTX660Ti chips? They'd have to manually laser cut working $399-499 670/680s and sell them for $249.

 

[blastingcap]

The GTX 660 Ti is rumored to be featuring 6 SMX clusters — thats 1152 CUDA cores.

Detailed specification is as follows:

Transistors Count: 3.5 billions (GK104 not GK106!)
Process: 28 nm
SMXs: 6
CUDA Cores: 1152
TMUs: 96
ROPs: 24
Base Clock: 1006 MHz
Memory Clock: 1502 MHz
Memory: 1536 MB GDDR5 192-bit
Bandwidth: 144 GB/s
Single-Precision Computing power: 2.35 TFLOPS
TDP: 150 W
Estimated Launch Price: $249



GTX670 = 1 cluster off. GTX680 = GTX670Ti. But you knew that!

MSI GeForce GTX 670 OC 2GB comes with 2 GB of GDDR5 memory. It features GK104 GPU with 1344 CUDA cores, 112 texture and 32 raster operating units. 670 OCs will prob. sell for $429, leaving stock 670s at $399.



JHH never mentioned yields being a problem with 28nm manufacturing. In the last conference call he said 28nm yields are better than there were for Fermi. I am not sure why you keep talking about this when it's contrary to facts. He specifically mentioned ramping up capacity at TSMC, not yields.



Ya, NV is unlikely to gimp GK104 into a GTX660Ti. They might release some GTX660 model but I doubt it will be GK104 based. For the foreseeable future, it seems they'll use GTX560 Ti / 448 core / GTX570 to fill $170-260 price bracket, leaving a gap in the $260-$400 space. Odd.

NV did say that they have gone from pay-per-working-GPU to pay-per-wafer so they are more affected by yields now even if yields are better. And it seems NV has been negatively affected by the change, else why cry about it per the extremetech article?

If a mature 40nm is cheap enough, it makes more sense to simply rely on those parts for longer rather than transition to the supply-constrained 28nm space.

http://www.extremetech.com/computin...y-with-tsmc-claims-22nm-essentially-worthless

 

[sontin]

You do know that yields aren't dependant only on the foundry side, right? AMD has claimed that 28nm chips are yielding as expected, Nvidia has said that they are below expectations. The process is the same in both cases, so it's pretty easy to see where the problem lies.

And both companies had the same expectations? :hmm:
nVidia's gross margin was 51% last quarter, AMD's 46%. It's easy to see that nVidia has much higher expectations from a gross margin standpoint than AMD.

Please point me to where Qualcomm has said that their 28nm chip yields are worse than expected. Otherwise you're arguing a strawman, because I never claimed that they aren't supply constrained. Which makes sense, because they're manufacturing SoCs that have thin margins and are completely dependant on volume, but makes much less sense for Nvidia because GPUs are on the opposite end of the spectrum.

Qualcomm doen't need to talk about yields. At the moment there business not totally relates to TSMC's 28nm process. It's different to nVidia.

 

[blastingcap]

And both companies had the same expectations? :hmm:
nVidia's gross margin was 51% last quarter, AMD's 46%. It's easy to see that nVidia has much higher expectations from a gross margin standpoint than AMD.



Qualcomm doen't need to talk about yields. At the moment there business not totally relates to TSMC's 28nm process. It's different to nVidia.

Was that the margin for both CPU and GPU or just GPU? AMD's CPU division is a perennial whipping boy and I wouldn't be surprised if that division dragged down the company average.

 

[RussianSensation]

NV did say that they have gone from pay-per-working-GPU to pay-per-wafer so they are more affected by yields now even if yields are better.

Yes, but that's not the same as stating that yields are below their expectations as claimed in this thread. NV has stated many times the yields are fine. Their main issues are 28nm wafer costs and capacity constraints. Those 2 factors have been highlighted multiple times. I can't find any evidence that Kepler has yield issues.

If a mature 40nm is cheap enough, it makes more sense to simply rely on those parts for longer rather than transition to the crowded 28nm space.

http://www.extremetech.com/computin...y-with-tsmc-claims-22nm-essentially-worthless

OK but in that's not related to yields but cost per wafer and capacity constraints which raise 28nm wafer prices even more given multiple customers at TSMC.

What you are linking has been discussed in reference to cost structure inherently becoming unfavourable as AMD/NV have to subsidize more expensive node transitions, not an issue with yields per say:

"In fact, Nvidia does not expect the cost per transistor to fall more than marginally below previous generations at 28nm, essentially losing out on potential profits. In other words, when Nvidia invests in future generations, the return on investment diminishes significantly."
http://beta.fool.com/topdowntrends/...law/4018/?ticker=QCOM&source=eogyholnk0000001

What we end up with is a 294mm^2 chip for $499...NV is paying higher costs to TSMC, but they are passing these costs to the consumers. It's no surprise then that GTX690 is $999. I hope not, but this generation may signal a new trend for GPUs - $400-500 may become the standard for good performance. The days of $269 HD5850 or $299 HD6950 might be gone for good unless TSMC and GloFo figure out a way to cut costs of future node transitions.

Considering this generation people are paying $500 for HD7970/GTX680, they have just sent the signal to both AMD and NV that they are perfectly fine with paying $500 for what essentially are upper mid-range chips. There you go. AMD and NV just got the green flag that consumers don't expect 50-75% generational increases anymore from the previous fastest high-end card (GTX580). That sets up the stage for GK110 (or w/e chip it will be called) to become GTX780 next generation when it should have been GTX685 all along.

 

[Silverforce11]

@ Russian, im talking about gk106 not gk104 derivatives.

The reason we aren't getting gk106 products soon is due to supply constraints, ie. they can't fill current low volume demands on 28nm, they simply cannot release a different chip competing for the same 28nm production. Thus any other gk cards will be gk104 derivatives, using best what they have.

 

[Arzachel]

The GTX 660 Ti is rumored to be featuring 6 SMX clusters — thats 1152 CUDA cores.

Detailed specification is as follows:

Transistors Count: 3.5 billions (GK104 not GK106!)
Process: 28 nm
SMXs: 6
CUDA Cores: 1152
TMUs: 96
ROPs: 24
Base Clock: 1006 MHz
Memory Clock: 1502 MHz
Memory: 1536 MB GDDR5 192-bit
Bandwidth: 144 GB/s
Single-Precision Computing power: 2.35 TFLOPS
TDP: 150 W
Estimated Launch Price: $249

The 660/660ti is rumoured to be a salvage part, be based on a totally different chip and only come out in six months. I guess we'll get some concrete information in the coming weeks.

JHH never mentioned yields being a problem with 28nm manufacturing. In the last conference call he said 28nm yields are better than there were for Fermi. I am not sure why you keep talking about this when it's contrary to facts. He specifically mentioned ramping up capacity at TSMC, not yields.

Wrong, I've quoted him in the last page talking explicitly about yields.

Will see when market share numbers are released and financial numbers.

I guess this is why I disagree with you. I'm an enthusiast, not a shareholder. I'm sure they will earn quite a bit on all those mobile GPU contracts they've been burning the vast majority of 28nm wafers on but I could care less about that.

 

[RussianSensation]

Wrong, I've quoted him in the last page talking explicitly about yields.

"Huang calmly explained that TSCM is "doing fabulously" with the 28 nanometer ramp, particularly compared to the 40 nanometer processes that were used on its prior "Fermi" generation of GPUs. The yields on 28 nanometer wafers are actually higher at the same point in the cycle than they were at TSMC for the 40 nanometer node."

He did note that yields were worse than their expectation was but they aren't the main problem here. He said: "Our expectation is that yields will improve, and as output increases, our costs will go down."

That's always the case with all nodes. No news here.

"We're ramping our Kepler GPUs very hard, but we could use more wafers," Huang said on the call.

The main problem with Kepler availability is not yields, but wafer capacity/availability. Yields are better than 40nm generation and NV was selling 520mm^2 GTX470/480 Fermi chips for $350-499. It's impossible to make an implication from this that yields are primarily responsible for $499 price of 294mm^2 GTX680 or for its limited availability since yields are higher on Kepler than they were on Fermi and the chip is far smaller and uses a less expensive 256-bit bus. Since availability of GTX470/480 Fermi cards was higher and yields were lower than they are on 28nm node, yield is not the main issue that's responsible for poor availability or the high price.

 

[blastingcap]

Financial calls are a joke sometimes, people know it too. Don't rely too heavily on comments made there when everyone knows SEC hardly ever does a damned thing and certainly won't prosecute on the usual puffery. Have you ever heard JHH on those calls? I want to slap him every time he repeats himself on those calls like when he gushes over Apple products, for instance. At least use different adjectives, man. "Exquisite." "Delightful." Puke.

We are straying a bit far from the thread topic, and also from known facts. RS has pointed out some facts (customer willingness to pay, etc.) but the yield/wafer thing we only know in the aggregate and from circumstantial data. And JHH's comment about comparing it to 40nm Fermi at the same point "in the cycle" is unclear--when does the cycle begin? When TSMC started accepting orders? Or NV tried fabricating their first Fermi? Or a respin? In any case 40nm Fermi was initially a disaster, so saying something is better than a disaster isn't saying much--it could be anywhere from a less-severe-disaster to a miracle.

 

[BallaTheFeared]

Nv is coasting at $500 thanks to no real threat, now again $1000 because AMD has no answer for Nv current perf/w dominance.

 

[Arzachel]

"The gross margin decline is contributed almost entirely to the yields of 28nm being lower than expected. That is, I guess, unsurprising at this point,” said Jen-Hsun Huang

"Fabulous" yields are still below their expectations. But I totally agree that yields aren't the only reason that GTX 680s are very scarce. They've been too optimistic about 28nm and taken on too many contracts from Laptop OEMs, poor yields are just the icing.

And both companies had the same expectations? :hmm:
nVidia's gross margin was 51% last quarter, AMD's 46%. It's easy to see that nVidia has much higher expectations from a gross margin standpoint than AMD.

You know what overpromising and underdelivering is called? Mismanagement

 

[blastingcap]

Nv is coasting at $500 thanks to no real threat, now again $1000 because AMD has no answer for Nv current perf/w dominance.

Nv never coasts. Why are you denigrating, indeed, LIBELING, Nv in such a way? Nv has never rested on its laurels. It keeps pushing the envelope, trying to deliver smoother frametimes and framerates for its customers, including its innovative Adaptive VSYNC technology, hardware controls to try to blunt the impact of microstuttering, up to four displays on one card without resorting to mini/Displayport adapters, GPU Boost to maximize performance if thermals allow for it, CUDA which is widely used in such things as Adobe products and HPC, PhysX which adds a new dimension of realism to games, superior Tessellation performance for three generations of GPUs in a row, Bezel peeking functions for multimonitor output, and superior value as seen by holding prices constant and delivering 35% more performance, comparing gtx 680 to gtx 580 launch prices? Not to mention Nv's continued push for more efficiency which leads to better thermals and noise, and continued partnership with gamedevs to produce well-functioning games, in addition to the symbiotic TWIMTBP program?

HOW DARE YOU libel Nv by claiming they are coasting when they continually innovate. Stop being an AMD fanboy.

 

[railven]

Before I get pulled into the CEO Forum chair wars, I just wanted to point something out I inferred from the quotes/comments.

Isn't it a given that "yields" would be better from a wafer when the chip itself is smaller? GK104 is what, almost 3/5 the size of GF100 was?

Also, isn't yields the total result NOT per? IE, using the wafer availability defense wouldn't refute the low yields comment if anything it would prove it:

Yields == total chips harvested

Bad yield from low wafer availabilty, even if more chips are made per wafer.

Trying to say they don't have a yield issue but then say they don't have enough wafers, well that makes me look at the statement a little cross eye'd. I'm sure you met they aren't having bad yield per wafer, just not having enough wafers, but in the end it still translate to bad yields.

 

[Jaydip]

Guys so have we reached a consensus about yields yet ?:biggrin:
U know in case of a doubt ask the CEO,call JHH ()

 

[sontin]

Before I get pulled into the CEO Forum chair wars, I just wanted to point something out I inferred from the quotes/comments.

Isn't it a given that "yields" would be better from a wafer when the chip itself is smaller? GK104 is what, almost 3/5 the size of GF100 was?

Also, isn't yields the total result NOT per? IE, using the wafer availability defense wouldn't refute the low yields comment if anything it would prove it:

Yields == total chips harvested

Bad yield from low wafer availabilty, even if more chips are made per wafer.

Trying to say they don't have a yield issue but then say they don't have enough wafers, well that makes me look at the statement a little cross eye'd. I'm sure you met they aren't having bad yield per wafer, just not having enough wafers, but in the end it still translate to bad yields.

Yields is usefull to determine the gross margin. But in the end you need the supply for the revenue. Even with better yields TSMC can not supply enough wafers.

You know what overpromising and underdelivering is called? Mismanagement

Mismanagement? Their Q1 margin guidence is much higher than there Q3 2011 (November 2010) margin. That was their first real Fermi quarter.

 

[Arzachel]

Before I get pulled into the CEO Forum chair wars, I just wanted to point something out I inferred from the quotes/comments.

Isn't it a given that "yields" would be better from a wafer when the chip itself is smaller? GK104 is what, almost 3/5 the size of GF100 was?

Also, isn't yields the total result NOT per? IE, using the wafer availability defense wouldn't refute the low yields comment if anything it would prove it:

Yields == total chips harvested

Bad yield from low wafer availabilty, even if more chips are made per wafer.

Trying to say they don't have a yield issue but then say they don't have enough wafers, well that makes me look at the statement a little cross eye'd. I'm sure you met they aren't having bad yield per wafer, just not having enough wafers, but in the end it still translate to bad yields.

This is where IDC swoops in and enlightens us all, but in essence there are two kinds of yields, parametric and functional.

Parametric yields mean how close to estimations do the resulting chips perform (leakage, clock speeds etc.)

Functional yields, the ones that are usually refered to when using the term, mean how many usable dies can you get from a wafer. Basically good dies/dies on the wafer(*100%). The defect rate per cm^2 is more or less constant for the same design on the same process, so smaller chips are affected less.

As to your other questions, no, it's not a given that a smaller chip yields necessarily better than a bigger one, because the process and design can vary. Yes, it's a given that GK 104 yields better than GF 100, because iirc the yields for GTX 480 were in single digits for quite a while.

 

[railven]

This is where IDC swoops in and enlightens us all, but in essence there are two kinds of yields, parametric and functional.

Parametric yields mean how close to estimations do the resulting chips perform (leakage, clock speeds etc.)

Functional yields, the ones that are usually refered to when using the term, mean how many usable dies can you get from a wafer. Basically good dies/dies on the wafer(*100%). The defect rate per cm^2 is more or less constant for the same design on the same process, so smaller chips are affected less.

I do recall that post of his...where is IDC - this seems like the kind of thread he'd come in and pwn somebody (possibly me ) haha.

Anyways, ducking out now - I just like to use the products, I care not who lines their pockets best.

 

[SirPauly]

I guess this is why I disagree with you. I'm an enthusiast, not a shareholder.

I'm not a shareholder of AMD or nVidia and been a PC gamer for many, many years.

 

[Arzachel]

I do recall that post of his...where is IDC - this seems like the kind of thread he'd come in and pwn somebody (possibly me ) haha.

Anyways, ducking out now - I just like to use the products, I care not who lines their pockets best.

IDC posts are the best part of these threads, rare as they might be. I went away from the three huge Bulldozer threads much more knowledgable about the underpinings of the tech which is now my favorite part of hardware releases. Well to be fair, those Bulldozer threads where 99% baseless speculation with small bits of knowledge sprinkled in, but it was fun I guess

I'm not a shareholder of AMD or nVidia and been a PC gamer for many, many years.

Sorry, didn't mean to imply that you're a shareholder, just that you're looking at this from a different viewpoint. Even if Nvidia's current strategy works from a financial standpoint, I'd prefer that they would focus on their desktop products.

 

[SirPauly]

Guys so have we reached a consensus about yields yet ?:biggrin:
U know in case of a doubt ask the CEO,call JHH ()

And I think he will be there in nVidia's financial conference call answering actual questions! hehe!

 

[SirPauly]

On the GTX 690:

Improved Frame rate Metering

Kepler introduces hardware based frame rate metering, a technology that helps to minimize stuttering. In SLI mode, two GPUs share the workload by operating on successive frames; one GPU works on the current frame while the other GPU works on the next frame. But because the workload of each frame is different, the two GPUs will complete their frames at different times. Sending the frames to the monitor at varying intervals can result in perceived stuttering.

The GeForce GTX 690 features a metering mechanism (similar to a traffic meter for a freeway entrance) to regulate the flow of frames. By monitoring and smoothing out any disparities in how frames are issued to the monitor, frame rates feel smoother and more consistent.

Curious about the bold part.

 

[SirPauly]

IT was kinda odd to see nVidia reference micro-stutter, hehe! Curious to know how effective the GTX 690 may be in curtailing this.

 

[BallaTheFeared]

Pretty sure it's just marketing. If it's anything it's a further improvement on what they already have.

Perhaps they're going for hearts and minds, letting those with 7970CF out there know they can get a 690 for $100+ less than their 7970s cost them and they won't have to jitter through gameplay at high fps.

Nv has been attempting to combat ms for quite awhile, through both drivers and hardware.

 

[blastingcap]

Pretty sure it's just marketing.

More libel against Nv by Balla in this thread?

 

[Arzachel]

On the GTX 690:

Curious about the bold part.

Been around in the drivers since G80 I believe. It tries to match the frames rendered to make the display smoother, but causes input lag. Bit more info on it in the bottom of this thread.