NVIDIA GF100 & Friends speculation

[Richthofen]

It's certainly going to affect the price they can afford to sell it for.

Nope not really as proven with the GTX460 and 470. All Nvidia will care about is to regain market share. They are in a very good position to do so. So if the GTX560 does perform where everybody is expecting it at the moment you will se a wide variety of cards directly battling with Barts based products.
DIE size definatly is not the thing they will bother much around with at the moment. They want to regain lost ground and financially they are in a much better position than AMD as the last quater has already shown.

 

[Alexko]

Nope not really as proven with the GTX460 and 470. All Nvidia will care about is to regain market share. They are in a very good position to do so. So if the GTX560 does perform where everybody is expecting it at the moment you will se a wide variety of cards directly battling with Barts based products.
DIE size definatly is not the thing they will bother much around with at the moment. They want to regain lost ground and financially they are in a much better position than AMD as the last quater has already shown.

If push came to shove, AMD could sell the HD 6870/6850 for about $140 and $90 respectively, pretty much what the 4870 and 4850 were selling for towards the end. Do you really think NVIDIA could do the same with GF114?

According to Charlie, GF104 is already selling at a loss.

 

[Florin]

If push came to shove, AMD could sell the HD 6870/6850 for about $140 and $90 respectively, pretty much what the 4870 and 4850 were selling for towards the end. Do you really think NVIDIA could do the same with GF114?

Sure.

According to Charlie, GF104 is already selling at a loss.

Ah so he thinks the same then.

But then what does he know?

 

[Alexko]

Sure.

While making money?

Ah so he thinks the same then.

But then what does he know?

The same as what?

 

[A1xLLcqAgt0qc2RyMz0y]

If push came to shove, AMD could sell the HD 6870/6850 for about $140 and $90 respectively

And AMD's graphics division which made only $1 million last quarter would be posting losses and the stock would take a huge hit.

According to Charlie, GF104 is already selling at a loss.

Charlie is wrong. According to Charlie everything nVidia sells is for a loss which can't be true since nVidia made $89 million last quarter.

 

[Mintmaster]

48 more SPs and a 100Mhz boost. Could add up to around 30% above 460 which would put it directly below 6950 but at $229 if Nvidia maintain the price point of the 460.

That's a bit optimistic, don't you think? The 6870 has 17% more shaders and 16% higher clock speed than the 6850, yet it is only 16% faster on that chart.

Don't expect the 560 to gain on the 460 much more than that for this suite of tests. It'll catch up to the 6870.

 

[Alexko]

And AMD's graphics division which made only $1 million last quarter would be posting losses and the stock would take a huge hit.

Come on, it's not that simple and you know it. First, AMD wasn't selling Barts at all in Q3. Second, they attributed the poor results to loss of business in the mobile market due to "real or perceived" shortages.

Not to mention that AMD's graphics division's results don't necessarily directly reflect the graphics division's performance, because we don't know what expenses it covers and how its earnings are recorded. Have development costs for Llano been shouldered by the CPU division? The GPU division? Both? 50/50? 80/20? What about Ontario?

What happens when AMD sells a CPU and an IGP to an OEM? Could the IGP be thrown in at cost to make the CPU more attractive? What if there's a discrete Radeon as well? Right now AMD's CPU division isn't doing terribly well, but it's AMD's core business so it must look good for investors. I wouldn't be surprised if it were leaning on the graphics division financially.

Charlie is wrong. According to Charlie everything nVidia sells is for a loss which can't be true since nVidia made $89 million last quarter.

I don't think he's ever said that. For one thing, he's certainly not denying NVIDIA's huge margins for Quadros and Teslas.

 

[Alexko]

That's a bit optimistic, don't you think? The 6870 has 17% more shaders and 16% higher clock speed than the 6850, yet it is only 16% faster on that chart.

Don't expect the 560 to gain on the 460 much more than that for this suite of tests. It'll catch up to the 6870.

There's no reason GF104 should exhibit the same kind of scaling. On that same chart, we see the GTX 470 and 480, with the latter having a 23.5% theoretical advantage: (480/448) × (1401/1215).

In practice, it's 21% faster (108/89), so that's about 98% scaling… for a high-end card that is more likely to run into system limitations than the GTX 460/560 would be.

 

[mczak]

There's no reason GF104 should exhibit the same kind of scaling. On that same chart, we see the GTX 470 and 480, with the latter having a 23.5% theoretical advantage: (480/448) × (1401/1215).

In practice, it's 21% faster (108/89), so that's about 98% scaling… for a high-end card that is more likely to run into system limitations than the GTX 460/560 would be.

You conveniently left out the +30% memory bandwidth the GTX 480 has over the GTX 470, which the GTX 560 will not have (and the +20% more rops, not that I think that makes much of a difference).
You're too optimistic. Yes in synthetics (like perlin noise - not that it will be fast enough to catch Barts there...) it could be that much faster. There are already GTX 460 with pretty much exactly the clocks like the GTX 560 on the market, just look at their results. They are still ~10% slower than HD 6870. So, with another SM (+16%) that should be enough to catch up, but there's zero reason to assume you get more than linear scaling with SMs, which would be needed to catch HD 6950. At best, with good SM scaling it could be very slightly faster than HD 6870, but still it will be closer to HD 6870 than HD 6950 likely. At worst, it might not scale well with SMs (though I expect it does) in which case it would have trouble catching even the HD 6870.

 

[A1xLLcqAgt0qc2RyMz0y]

I don't think he's ever said that. For one thing, he's certainly not denying NVIDIA's huge margins for Quadros and Teslas.

nVidia has positive gross margins on all product lines including the GeForce line. That means that they are profitable on the Gaming GeForce line. Something that Charlie and his followers deny. So there you have it just one of the many lies from Charlie that his devoted followers quote daily.

 

[CRoland]

nVidia has positive gross margins on all product lines including the GeForce line.

Do you have a source for that?

(Not really doubting it's true. I'd just like to see more details.)

 

[silent_guy]

According to Charlie, GF104 is already selling at a loss.

No sane business is going to be selling their sweet spot product and probably highest runner at a loss. No fabless company is going to sell any silicon at a loss. Charlies's calculations may be a wet dream for some ATI fans but they belong in fantasy land. (You should go to his forums once in a while: highly entertaining. Some commenters suggesting that even Tegra chips are being sold at a loss.)

By now, 40nm is relatively mature (it's 2 year old, after all). The financials should be very decent, even for the highest die sizes.

What possible upside could there be to gain market share while making a loss?

 

[A1xLLcqAgt0qc2RyMz0y]

Do you have a source for that?

(Not really doubting it's true. I'd just like to see more details.)

It has been stated that the GeForce margins are positive but lower than the Tesla in various earnings conference calls over the last year.

You can read them here: http://seekingalpha.com/transcripts/for/NVDA

Also nVidia has stated that they have to be in the GeForce line to be able to afford to produce the Tesla line. Thus indirectly stating they must make a profit on the GeForce line to help offset the huge costs of developing GPUs on a yearly basis. Thus they are not selling GeForce GPUs at a loss.

http://www.dvhardware.net/article46630.html

"I do not see the economic model [with MIC]. We are able to produce those [Tesla] GPGPUs because the ultimately there is one GPU for GeForce consumer graphics, Quadro professional business. It costs $500 million to $1 billion to develop those new products every year, it is a huge investment. Unless you have that [consumer and professional] economic engine in the background, I cannot imagine how one could make a GPU without having a graphics business," said Mr. Gupta.

 

[A1xLLcqAgt0qc2RyMz0y]

Please Lets get back to Speculations on hardware

Sorry but I had to respond to counter the financial FUD that was being posted here about nVidia.

Now please lets get back to the hardware speculation on the Fermi family.

I have seen rumors that the GTX560 will be released on Jan 20, 2011. Has this date been confirmed?

Will the dual GPU card from nVidia contain GTX560s or GTX570s?

When will the dual cards from AMD & nVidia be released?
And who will release theirs first?

 

[Alexko]

You conveniently left out the +30% memory bandwidth the GTX 480 has over the GTX 470, which the GTX 560 will not have (and the +20% more rops, not that I think that makes much of a difference).
You're too optimistic. Yes in synthetics (like perlin noise - not that it will be fast enough to catch Barts there...) it could be that much faster. There are already GTX 460 with pretty much exactly the clocks like the GTX 560 on the market, just look at their results. They are still ~10% slower than HD 6870. So, with another SM (+16%) that should be enough to catch up, but there's zero reason to assume you get more than linear scaling with SMs, which would be needed to catch HD 6950. At best, with good SM scaling it could be very slightly faster than HD 6870, but still it will be closer to HD 6870 than HD 6950 likely. At worst, it might not scale well with SMs (though I expect it does) in which case it would have trouble catching even the HD 6870.

Fair point about bandwidth. Let's look at it differently. Compared to the GTX 470 (which is a tad faster than the 6870) and assuming that the leaked specs are true (384 SPs, 820MHz) and that my math is correct, then the GTX 560 would have:

• 15% higher shader throughput,
• 4.4% lower memory bandwidth,
• 8% more fillrate,
• over 50% more texturing throughput,
• 20% less memory,
• 33% less geometry throughput (assuming no driver weirdness).

Shouldn't that be faster than the GTX 470? Therefore faster than the 6870 and close to the 6950?

No sane business is going to be selling their sweet spot product and probably highest runner at a loss. No fabless company is going to sell any silicon at a loss. Charlies's calculations may be a wet dream for some ATI fans but they belong in fantasy land. (You should go to his forums once in a while: highly entertaining. Some commenters suggesting that even Tegra chips are being sold at a loss.)

By now, 40nm is relatively mature (it's 2 year old, after all). The financials should be very decent, even for the highest die sizes.

What possible upside could there be to gain market share while making a loss?

The GTX 460 was most likely introduced as a profitable product, but right now you can get one for $120 after MIR. Considering the kind of product we're talking about (367mm² die, 150W) can NVIDIA really be making money on this? Sure, the intention here would be to get rid of inventory to make room for GF114, but still.

It has been stated that the GeForce margins are positive but lower than the Tesla in various earnings conference calls over the last year.

You can read them here: http://seekingalpha.com/transcripts/for/NVDA

Also nVidia has stated that they have to be in the GeForce line to be able to afford to produce the Tesla line. Thus indirectly stating they must make a profit on the GeForce line to help offset the huge costs of developing GPUs on a yearly basis. Thus they are not selling GeForce GPUs at a loss.

http://www.dvhardware.net/article46630.html

"I do not see the economic model [with MIC]. We are able to produce those [Tesla] GPGPUs because the ultimately there is one GPU for GeForce consumer graphics, Quadro professional business. It costs $500 million to $1 billion to develop those new products every year, it is a huge investment. Unless you have that [consumer and professional] economic engine in the background, I cannot imagine how one could make a GPU without having a graphics business," said Mr. Gupta.

You're grasping at straws, here. Just because the GeForce line has positive margins overall and JHH has made a few vague statements about needing the graphics business for Tesla doesn't mean the GTX 460 is making money.

 

[mczak]

Fair point about bandwidth. Let's look at it differently. Compared to the GTX 470 (which is a tad faster than the 6870) and assuming that the leaked specs are true (384 SPs, 820MHz) and that my math is correct, then the GTX 560 would have:

• 15% higher shader throughput,
• 4.4% lower memory bandwidth,
• 8% more fillrate,
• over 50% more texturing throughput,
• 20% less memory,
• 33% less geometry throughput (assuming no driver weirdness).

Shouldn't that be faster than the GTX 470? Therefore faster than the 6870 and close to the 6950?

That's hard to tell as the SMs are too different. So for instance the 15% higher shader throughput doesn't take into account that utilization should be a bit less on average - OTOH it has 50% more SFU throughput.
Also, the 8% more fillrate is only true for z fill. Color fillrate deduced from the rops is meaningless (for almost all formats), the GTX 470 still has a quite large advantage there (14x64bit export from SMs vs. 8x64bit per clock).
IMHO could easily end up about as fast as GTX 470.
I think looking at the OC GTX 460 makes just more sense - you have none of the architectural differences to deal with. FWIW, some considered the evga gtx 460 ftw to be as fast as the HD 6870 - the GTX 560 should indeed be a bit faster (same memory clock, 4% less core clock but one SM more). It depends on the apps and settings used - in particular it looked to me like with resolutions below 1920x1200 indeed the HD 6870 wasn't faster any more overall.

 

[Bouncing Zabaglione Bros.]

It has been stated that the GeForce margins are positive but lower than the Tesla in various earnings conference calls over the last year.

You can read them here: http://seekingalpha.com/transcripts/for/NVDA

Also nVidia has stated that they have to be in the GeForce line to be able to afford to produce the Tesla line. Thus indirectly stating they must make a profit on the GeForce line to help offset the huge costs of developing GPUs on a yearly basis. Thus they are not selling GeForce GPUs at a loss.

So you're sure that the Geforce line is in positive profit because you're lumping in the high profits from the professional line based on the same family of chips? That's changing the goalposts but I see where you are coming from.

It seems to me that Nvidia would rather like to make some/more profit from the consumer cards too, instead of burning cash on them to keep market share, and not have to use the professional products to subsidise the consumer cards.

 

[Alexko]

That's hard to tell as the SMs are too different. So for instance the 15% higher shader throughput doesn't take into account that utilization should be a bit less on average - OTOH it has 50% more SFU throughput.
Also, the 8% more fillrate is only true for z fill. Color fillrate deduced from the rops is meaningless (for almost all formats), the GTX 470 still has a quite large advantage there (14x64bit export from SMs vs. 8x64bit per clock).
IMHO could easily end up about as fast as GTX 470.
I think looking at the OC GTX 460 makes just more sense - you have none of the architectural differences to deal with. FWIW, some considered the evga gtx 460 ftw to be as fast as the HD 6870 - the GTX 560 should indeed be a bit faster (same memory clock, 4% less core clock but one SM more). It depends on the apps and settings used - in particular it looked to me like with resolutions below 1920x1200 indeed the HD 6870 wasn't faster any more overall.

Here are the results from Damien's Barts review: http://www.hardware.fr/medias/photos_news/00/29/IMG0029787.gif

At 800/1000, the GTX 460 is just 6% slower than the 470 (1920×1200 AA4X) and 1% slower than the 6870.

The GTX 560 could be up to 17% faster than that: (384/336)×(820/800) = 1.17. Let's call it 10% faster.

And here, the 6950 is just 7% faster than the 470.

So I'm expecting HD 6870 < GTX 470 < GTX 560 < HD 6950, all within a ~12% performance range, going by that last graph. I guess drivers could help Cayman, though.

 

[Jawed]

Sorry but I had to respond to counter the financial FUD that was being posted here about nVidia.

If you knew what you were talking about or could interpret the available facts correctly, it would have been worth reading. /offtopic

 

[trinibwoy]

Out of curiosity, what are the available facts that show nVidia is losing money on the 460? All I see is people disagreeing based only on their own presumed omniscience.