First GTX 670 review(s) up (tt & oc.net) * TT OC review added*

[blastingcap]

these cards cost like $25 to make. Sure they have room.

They cost more than that to make, come on now. The real problem is that R&D doesn't come for free. Gotta spread that huge cost among as many units as possible.

Edit: I see you edited to say $50 instead of $25. But that is still too low considering what TSMC charges (TSMC has to spread its huge fixed fab costs among units, too), the PCB and packaging costs, and shipment costs.

 

[tviceman]

No but their CEO has stated they expect revenues to be down, you can take his word for it can't you?

As per gk104 gtx660ti vs 78xx, if NV price them to pressure, AMD can simply lower the price. 78xx is a much smaller die.

I see where JHH expected revenue to be down because of the continued hard drive shortage and availability of 28nm wafers. We shall see how it all translates. Again, AMD had higher selling prices and less revenue. Their graphics cards were not constantly and completely sold out everywhere, sans the first week or two of the hd7970's release.

And as for your continued argument that AMD can simply lower prices, you failed to read everything I wrote. AMD can't wave a magic wand and simply lower prices on one or two products. If they drop their price on the hd7870 by a meaningful amount, it will have an affect on the hd7950 and hd7850, which will then have an affect on the hd7970 and hd7770... and so on, and so fourth.

 

[Jaydip]

Which just proves what I said earlier.

Pro tip: if you buy like that you're not getting something that's actually cheaper.

I agree but waiting to get it cheaper is very boring

 

[Zebo]

itx mobo, ivy, with that short card and a waterblock..think of the possibilities. I've always wanted a little cube on my desk, like apples cube, but not worthless

 

[Zebo]

They cost more than that to make, come on now. The real problem is that R&D doesn't come for free. Gotta spread that huge cost among as many units as possible.

Edit: I see you edited to say $50 instead of $25. But that is still too low considering what TSMC charges (TSMC has to spread its huge fixed fab costs among units, too), the PCB and packaging costs, and shipment costs.

They are selling video cards for $20 and not losing money or they would not sell them.

 

[Silverforce11]

I see where JHH expected revenue to be down because of the continued hard drive shortage and availability of 28nm wafers. We shall see how it all translates. Again, AMD had higher selling prices and less revenue. Their graphics cards were not constantly and completely sold out everywhere, sans the first week or two of the hd7970's release.

And as for your continued argument that AMD can simply lower prices, you failed to read everything I wrote. AMD can't wave a magic wand and simply lower prices on one or two products. If they drop their price on the hd7870 by a meaningful amount, it will have an affect on the hd7950 and hd7850, which will then have an affect on the hd7970 and hd7770... and so on, and so fourth.

Which is fine given that the 79xx is a mid-range chip being sold at high-end prices for all these months already. You think AMD doesn't know this when they set the price so high? They are raping consumers as long as they can get away with it. Likewise, given the die sizes on the 78xx, it could be $200 and still be profitable. Not as high as current rape/pillage prices, but certainly not doom and gloom and nowhere as bad as being forced to sell >500mm2 dies at mid-range prices.

As for revenue: What would you do: your product is selling fine, its being matched by limited supply. Would you sell it for less unless you had to? ie. When its no longer selling fine and supply > demand. Currently TSMC is restricting supply to both vendors. You figure it out.

 

[tviceman]

I think GK110 will be exclusively dedicated to compute and will go in Teslas and Quadros. GK104 is the "one size fits all" for gaming.

There is no way Nvidia is going to rely exclusively on the HPC and Pro market to make up for all the R&D it will have cost to develop Kepler, and then build Kepler. No way at all. And remember the supposed performance rumors of GK104 before it came out? Oh wait, there weren't any. Not until reviewers got their hands on cards. So any performance rumors regarding big Kepler are all BS. But if you want to speculate, and speaking in percentages, the expected Big kepler specs (2304 shaders, 512-but bus) is comparatively more beefed up (50% more shaders and potentially double the mem bandwidth) over GK104 than GF110 was over GF114 (GF110 had 33% more shaders and 50% bigger mem bus, a little less on the bandwidth). GF110 was about 35-40% faster than GF114.

So extrapolate from there.

 

[blastingcap]

They are selling video cards for $20 and not losing money or they would not sell them.

Clearances don't count, as they may be taking a loss, but better some revenue than none at all. Also, after-rebate counts only partially. They get money from reselling your address to marketers, and many people forget to mail in rebates anyway.

The avg selling price of even the crappiest discrete cards at launch is over $20. That's with dirt cheap GPUs with minimal pcbs and packaging, not high-end GPUs and the big pcbs, boxes, and accessories.

 

[SirPauly]

I think GK110 will be exclusively dedicated to compute and will go in Teslas and Quadros. GK104 is the "one size fits all" for gaming.

Possibly, but considering the lower than expectation on yields, constraints - doesn't make much sense to bring the monolith to the GeForce market -- especially in the context of the GK-104 commanding a 499 MSRP at this time.

The margins of the professional brands makes a lot of sense to be introduced first for the monolith and when 28nm matures and constraints much less -- why not introduce a refined, polished seamless monolith to GeForce -- named the GTX 780? For the price-point that nVidia may like 649 dollars? And to keep that 499 dollar price point strong with the second tier chip.

I just can't see all the R&D and other resources on just the compute market or professional families when it can bring additional revenues and margins to GeForce as well.

It may be for just the professional markets but time will tell.

 

[Crap Daddy]

There is no way Nvidia is going to rely exclusively on the HPC and Pro market to make up for all the R&D it will have cost to develop Kepler, and then build Kepler. No way at all. And remember the supposed performance rumors of GK104 before it came out? Oh wait, there weren't any. Not until reviewers got their hands on cards. Speaking in percentages, the expected Big kepler specs (2304 shaders, 512-but bus) is comparatively more beefed up over GK104 than GF110 was over GF114. And GF110 was about 35-40% faster than GF114.

I wasn't saying that Kepler was designed for HPC exclusively but that GK110 will go exclusively into Tesla. We might see the big K for gaming when the 8000 series from AMD is out.

 

[Jaydip]

Can we stop financial discussions in this thread please,all of this revenue forecasts are too damn boring

 

[Zebo]

Okay - all I'll say is there is a lot of room to work - you'll see all those lies about process as to why AMD was raping were just lies as 7xxx crash in price and sell merrily.

 

[tviceman]

Which is fine given that the 79xx is a mid-range chip being sold at high-end prices for all these months already. You think AMD doesn't know this when they set the price so high? They are raping consumers as long as they can get away with it. Likewise, given the die sizes on the 78xx, it could be $200 and still be profitable. Not as high as current rape/pillage prices, but certainly not doom and gloom and nowhere as bad as being forced to sell >500mm2 dies at mid-range prices.

As for revenue: What would you do: your product is selling fine, its being matched by limited supply. Would you sell it for less unless you had to? ie. When its no longer selling fine and supply > demand. Currently TSMC is restricting supply to both vendors. You figure it out.

I look on newegg and I see double digit selections for all of AMD's cards, sitting there building up eDust. I don't think AMD is any longer supply constrained. They have been meeting / exceeding demand for sometime now. If I were AMD, I would not have intro'd the hd7970 at such a high price, nor would I have taken as long as AMD has taken to react to Nvidia's products (and upcoming products). But then again, unlike AMD I think I would like selling my stuff and beating my competition not only to the market, but into paying customer's hands as well.

There is nothing quite like having a 1 mile head start in a 5 mile race, but then literally shooting yourself in the foot at the second mile.

 

[Silverforce11]

There is no way Nvidia is going to rely exclusively on the HPC and Pro market to make up for all the R&D it will have cost to develop Kepler, and then build Kepler. No way at all. And remember the supposed performance rumors of GK104 before it came out? Oh wait, there weren't any. Not until reviewers got their hands on cards. So any performance rumors regarding big Kepler are all BS. But if you want to speculate, and speaking in percentages, the expected Big kepler specs (2304 shaders, 512-but bus) is comparatively more beefed up (50% more shaders and potentially double the mem bandwidth) over GK104 than GF110 was over GF114 (GF110 had 33% more shaders and 50% bigger mem bus, a little less on the bandwidth). GF110 was about 35-40% faster than GF114.

So extrapolate from there.

Big K has 50% more shaders. Assume identical IPC (clock-for-clock) for gaming.

How fast do you think they will run the Big K core at? Because 1058mhz sounds a bit high for such a freaken huge chip (same rumors ~600mm2).

Let assume NV decides to be conservative and stick to the 300W TDP. Then assume the following: It's got ~50% more TDP to play with, so in a perfect world it should be 50% faster. But how does focussing on compute/HPC affect efficiency? Thats the key point here. If by going down that road reducs their efficiency, 50% TDP may only lead to ...% more performance. That's the unknown currently.

 

[LOL_Wut_Axel]

I agree but waiting to get it cheaper is very boring

Yeah, but wasn't the GTX 480 launched more than two years ago? Pretty long wait to get it at $200, LOL. That, and the power consumption is atrocious. OC vs OC. it's similar to the HD 7850 with the difference being power consumption, $50, and that you're getting a current card. For $50 more the HD 7850 seems like the way to go if you're gonna OC.

Now, great things come to those who wait. Hopefully nearing the end of this year I'll be able to get a GTX 670 for $350 or under and if not an HD 7950 for $300-320. If I get the 670 I'll probably leave it stock because of NVIDIA's shenanigans of automatic voltage control and the fact scaling isn't all that great. 980MHz stock is still good. If I don't get it my way but find 7950s cheap, I'll overclock to 975MHz for stock HD 7970/GTX 670 performance and try for an undervolt. If not, I'll keep lowering the clock speed until I can undervolt. The max goal for me with all my hardware is high performance with high efficiency, so that's what I want to aim at. :thumbsup:

 

[tviceman]

I wasn't saying that Kepler was designed for HPC exclusively but that GK110 will go exclusively into Tesla. We might see the big K for gaming when the 8000 series from AMD is out.

Okay I gotcha. Yes I agree, I think Nvidia will release big K to the pro market first, and then as a gtx780 this fall or late summer.

 

[blastingcap]

Okay - all I'll say is there is a lot of room to work - you'll see all those lies about process as to why AMD was raping were just lies as 7xxx crash in price and sell merrily.

If you think AMD has room, they have nothing on NV. That midrange GTX680 is selling for $500 on a voltage-locked midrange-looking PCB, and the GTX670 PCB costs even less to make! And of course Intel has even thicker margins.

 

[SirPauly]

Can we stop financial discussions in this thread please,all of this revenue forecasts are too damn boring

No!

 

[tviceman]

Big K has 50% more shaders. Assume identical IPC (clock-for-clock) for gaming.

How fast do you think they will run the Big K core at? Because 1058mhz sounds a bit high for such a freaken huge chip (same rumors ~600mm2).

Let assume NV decides to be conservative and stick to the 300W TDP. Then assume the following: It's got ~50% more TDP to play with, so in a perfect world it should be 50% faster. But how does focussing on compute/HPC affect efficiency? Thats the key point here. If by going down that road reducing their efficiency, 50% TDP may only lead to ...% more performance.

GTX580 had 33% more shaders, and lower clock speed than gtx560ti, yet was still 40% faster, and TDP was 250 watts.

I think Nvidia has way more potential with Big K over GK104 as GF110 had, and I also think they will have more wriggle room with power consumption - either through their efficiency improvements or by aiming for 275-300 watts. Either way, Big K should be at the very least as much of a jump over GK104 as GF110 was over GF104.

 

[tviceman]

If you think AMD has room, they have nothing on NV. That midrange GTX680 is selling for $500 on a voltage-locked midrange-looking PCB, and the GTX670 PCB costs even less to make! And of course Intel has even thicker margins.

I honestly don't think PCB's cost that much (relative to the MSRP). I am guessing a smaller PCB might save around $5 or so, but nothing that is substantial.

 

[Jaydip]

No!

Come on,my head hurts from all this revenue stuff :'(

 

[Silverforce11]

gtx580 vs gtx560ti, looks like 25-30% difference.

the distinction here is that the gtx560ti while compute crippled, it was done so via bios and drivers as the cuda cores are essentially the same.

gk104 vs big K is totally different, in that gk104 is already peak gaming performance per W and per mm2, while big K is going to take a hit to both. Thus, putting it at 25% above gk104 is not unrealistic.

In total, gk104 is ~35% faster than a gtx580. If big K is another 25% above that, it puts it into the 60-70% improvement range. That is SPOT ON for a full node shrink with similar die size (with both architectures compute focused).

 

[blastingcap]

I honestly don't think PCB's cost that much (relative to the MSRP). I am guessing a smaller PCB might save around $5 or so, but nothing that is substantial.

You are right to some degree (though 384-bit bus is more expensive to make than 256-bit even if they were the same size PCB--more layers) but it all adds up. Smaller packaging and lighter weight means smaller costs for packaging and shipping.

1GB less VRAM.

Smaller die.

Sales to HPC and pro graphics (eventually).

Probably higher volume, if history is any indication.

Yet NV can charge AMD-esque prices because NV parts perform as good or better for the price. (NV is also sticking to the mature 40nm process longer, as it's cheaper and they are apparently still selling 40nm cards quite well, also 40nm HPC/pro graphics cards of course.)

It all adds up to how NV should have significantly better margins on their 28nm chips than AMD. NV also has no meaningful corporate debt, compared to AMD's big debt and getting bled to death by Intel on the CPU front. NV can withstand a protracted price war in gaming-grade GPUs far better than AMD can. But both companies seem content to keeping prices high, profits up, and investing the proceeds for future battles. That's probably as it should be, given the tough fights they have ahead, in CPU/APUs and in mobile.

 

[Jaydip]

gtx580 vs gtx560ti, looks like 25-30% difference.

the distinction here is that the gtx560ti while compute crippled, it was done so via bios and drivers as the cuda cores are essentially the same.

gk104 vs big K is totally different, in that gk104 is already peak gaming performance per W and per mm2, while big K is going to take a hit to both. Thus, putting it at 25% above gk104 is not unrealistic.

In total, gk104 is ~35% faster than a gtx580. If big K is another 25% above that, it puts it into the 60-70% improvement range. That is SPOT ON for a full node shrink with similar die size.

I don't think BigK is aimed for gaming at all.Its primary focus is compute performance.

 

[Silverforce11]

But both companies seem content to keeping prices high, profits up, and investing the proceeds for future battles. That's probably as it should be, given the tough fights they have ahead, in CPU/APUs and in mobile.

This is the correct analysis. They are there to make as much profit from consumers as they can and given TSMC's poor production, they have to be insane to price-war to the ground.