nVidia GT200 Series Thread

[Aberforth]

excelltron, the boundaries you talk is only a delusive assumption of human mind not being able to think good enough, such assumptions existed several years ago and would continue to exist.

 

[Keysplayr]

Originally posted by: Extelleron


They wouldn't be equal to the GT200 at all. I'm saying if you could split the GT200 into four chips, and together you would have 240SP / 80 TMU / 32 ROP, then they would (in theory) be 144mm^2.

Die size doesn't matter much at all for consumers, but for nVidia it does. Die size is the most important thing, along with power/heat, that gets in the way of advancing performance. The faster you want a chip to be, the bigger you need to make it.... and obviously there are limits on that.

Here are supposed overclocking results/performance for the GTX 280:

http://digidownload.libero.it/hackab321/gtx280.bmp

It seems to scale very well with clocks, but the overclocking results are not very good. I was hoping GT200 would be capable of 700MHz+, and heat was restricting the clocks. Could be that they just got a bad chip.

144mm^2 per chip you mean? But you would still need four of them to equal a single GTX280, and now, you're using even more wafer real estate. 144x4 = 576mm^2, PLUS all the transistors to connect them all together (hypertransport like or whatever).

I don't get your line of thinking here. You end up with using more wafer. Each 144^2 chip would have to have additional transistors to be able to connect with others.

At any rate, that is neither here nor there. It is what it is.

 

[Extelleron]

Originally posted by: keysplayr2003

Originally posted by: Extelleron


They wouldn't be equal to the GT200 at all. I'm saying if you could split the GT200 into four chips, and together you would have 240SP / 80 TMU / 32 ROP, then they would (in theory) be 144mm^2.

Die size doesn't matter much at all for consumers, but for nVidia it does. Die size is the most important thing, along with power/heat, that gets in the way of advancing performance. The faster you want a chip to be, the bigger you need to make it.... and obviously there are limits on that.

Here are supposed overclocking results/performance for the GTX 280:

http://digidownload.libero.it/hackab321/gtx280.bmp

It seems to scale very well with clocks, but the overclocking results are not very good. I was hoping GT200 would be capable of 700MHz+, and heat was restricting the clocks. Could be that they just got a bad chip.

144mm^2 per chip you mean? But you would still need four of them to equal a single GTX280, and now, you're using even more wafer real estate. 144x4 = 576mm^2, PLUS all the transistors to connect them all together (hypertransport like or whatever).

I don't get your line of thinking here. You end up with using more wafer. Each 144^2 chip would have to have additional transistors to be able to connect with others.

At any rate, that is neither here nor there. It is what it is.

You're using the exact same amount of area on the wafer, and a few more transistors to connect them together, yes.

And you do that at the rate of increasing yield exponentially. The yield on a 576mm^2 chip at TSMC is probably <50%, but the yield of a 144mm^2 would be much, much higher. And as I said, if GT200 were really made out of 4-die, it could have been produced on a 55nm process, so the die size could have been lower. I'm throwing numbers out here, but if I can get 75% yield out of a 144mm^2 chip, and say 40% yield out of a 576mm^2 chip, then I am saving a HECK of a lot of money moving to a 4-die GPU, even though I am using the same amount of wafer space.

 

[Extelleron]

Originally posted by: Aberforth
excelltron, the boundaries you talk is only a delusive assumption of human mind not being able to think good enough, such assumptions existed several years ago and would continue to exist.

I don't know what you are getting at here. There are boundaries in the tech world, as I said those boundaries can be pushed back with newer fabrication technology. But you can't just produce whatever you want at any point in time. Intel couldn't produce an 820M transistor quad-core CPU back in 2000 when we were on 180nm, just like you can't produce a 4 Billion transistor CPU w/ 16 cores right now on 45nm. I shouldn't say you couldn't, because you could.... but you couldn't build it to sell it.

 

[Aberforth]

Originally posted by: Extelleron

Originally posted by: Aberforth
excelltron, the boundaries you talk is only a delusive assumption of human mind not being able to think good enough, such assumptions existed several years ago and would continue to exist.

I don't know what you are getting at here. There are boundaries in the tech world, as I said those boundaries can be pushed back with newer fabrication technology. But you can't just produce whatever you want at any point in time. Intel couldn't produce an 820M transistor quad-core CPU back in 2000 when we were on 180nm, just like you can't produce a 4 Billion transistor CPU w/ 16 cores right now on 45nm. I shouldn't say you couldn't, because you could.... but you couldn't build it to sell it.

I am talking about architecture limitation than the die itself, when you come up with a good architecture you wouldn't need a 16 core fiasco (*especially* when they dunno how to make use of 2 cores), one core will be equal to the speed of 16 cores. Such things rarely happen as you know. Multi-core itself is a big compromise. The architecture you see these days aren't innovations, they are forced and backed by commercial gains.

 

[Extelleron]

Originally posted by: Aberforth

Originally posted by: Extelleron

Originally posted by: Aberforth
excelltron, the boundaries you talk is only a delusive assumption of human mind not being able to think good enough, such assumptions existed several years ago and would continue to exist.

I don't know what you are getting at here. There are boundaries in the tech world, as I said those boundaries can be pushed back with newer fabrication technology. But you can't just produce whatever you want at any point in time. Intel couldn't produce an 820M transistor quad-core CPU back in 2000 when we were on 180nm, just like you can't produce a 4 Billion transistor CPU w/ 16 cores right now on 45nm. I shouldn't say you couldn't, because you could.... but you couldn't build it to sell it.

I am talking about architecture limitation than the die itself, when you come up with a good architecture you wouldn't need a 16 core fiasco (*especially* when they dunno how to make use of 2 cores), one core will be equal to the speed of 16 cores. Such things rarely happen as you know. Multi-core itself is a big compromise. The architecture you see these days aren't innovations, they are forced and backed by commercial gains.

That strategy worked back in the 1990's and early 2000's, but it isn't going to work anymore. There will be IPC increases, but not enough to fuel the CPU industry all by itself.

The reason we have multi-core is improving IPC by such a margin requires exponential increases in transistor count (thus die size), and heat/power consumption. Multi-core ends up being more practical in terms of the number of transistors/amount of power/heat utilized in return for the added performance.

But that's getting a bit off topic here, as this is supposed to be a thread for discussing GT200.

 

[BFG10K]

What I am talking about, and I've said this before, is that the die size of GPUs keeps rising despite the move to smaller processes at a rapid rate. I've supported this with more than enough evidence. So tell me why this won't continue?

It will continue. Now tell me why you don?t think this doesn?t affect multi-GPU given single GPUs are their building blocks? If die size is a problem for single core why don?t you think it won?t eventually become a problem for four such cores slapped onto one die?

Our four chips would have a die size of around ~144mm^2. That is a very acceptable die size for a chip and the yields would be excellent. But it gets even better.... we don't need to use a 65nm process anymore, we can go to 55nm. This reduces heat/power consumption, allows us to increase clocks, and allows for lower die sizes. Given a 100% shrink, the die size of our chips would actually be 103mm^2 (obviously this doesn't take into account that very few die shrinks will be 100%).

This is all well and good until it?s time to move forward again and you already have four cores on a GPU. What then? You?ll either have to make each of those four GPUs faster and as you start doing this their die sizes get bigger and eventually you?ll hit die size limits again like you did with single core.

Your other option is to keep adding more and more cores and expect the driver to deliver n-way multi-GPU scaling which of course will never happen. Two-way scaling isn?t robust enough let alone n-way.

Our hypothetical GT200 will have significantly better yield, lower power consumption, and higher performance than the single-GPU GT200 nVidia will put out.

The only way it?ll have higher performance is if four-way scaling works substantially in every game which of course will never happen. Four-way scaling in cherry picked benchmarks from reviews doesn?t count.

And this is ignoring other things such as driver complexity, input lag, micro-stutter and driver problems not present on single GPU solutions.

And creating an additional SKU is extremely easy; we just put 3 die instead of 4. There we have our GTX 260, but we aren't wasting any die space.

But we?ll get absolutely no performance gain unless the driver can scale from 3-way to 4-way in all applications. That will never happen.

Which GT200 do you think is better? The single-GPU one, or one made out of 4 die? I think it is pretty obvious.

It is pretty obvious ? the GT200 by far. To think otherwise demonstrates a fundamental lack of understanding of the limitations of multi-GPU systems. Slapping X cores on either a CPU or GPU doesn?t guarantee X times speed-up, In fact it doesn?t guarantee any kind of speedup.

There is a reason why Intel does the exact thing I am saying; the yield of two 143mm^2 chips makes it much cheaper to produce than a single 286mm^2 chip. Intel doesn't suffer any noticeable performance loss from having two die, so nVidia shouldn't have a problem either.

You?re assuming Intel chose between a bigger single core and several smaller cores when that necessarily isn?t the case. It could be they designed a single core to be as fast as possible in line with current market competition and then found they could add other cores for ?free? because the manufacturing process was advanced enough.

Look at Prescott; it had multiple cores but it was a furnace, it was slow, and they couldn?t increase clock speeds anymore because of thermal limits. Clearly multi-core isn?t the holy grail you think it is.

 

[bryanW1995]

Originally posted by: HOOfan 1

Originally posted by: Aberforth

I've never seen such an incompetent company really...

more like I've never seen such an incompetent "tech news" site

+1


these scores don't even matter when you think about. As my boss has told me on several occasions, all things are relative. If the 280 is 20% faster than 9800gx2, while the 4870 is 40% slower, then we'll all be very excited about the 280. If the 280 is 65% faster than 9800gx2 while the 4870 is 50% faster than 9800gx2, the 280 will not sell worth a crap at $649 and will probably be considered a "failure".

 

[Nemesis 1]

BfG10K , Your surprizing me here. I have read your post for years. They always been pretty good .
I think maybe that multicore gpus are here to stay and infact are the near future for all gpu makers. Its the best easiest way to go forward.

For me I don't really care which of these GPU's has best performance. ATI is at least 1 generation away from what really has my interest . NV maybe 1 generation maybe 2 if they stay with 1 core . We already know intel will be many cores scaling perfectly with more cards.

Thats maybe why ATI isn't moving to far away from whats they already have . Intels first card will be a DX11 part. Thats known fact . Ati may be looking past this round to the next. ATI is going to get ready for Intel and not NV . Why should AMD worry about NV I see zero reason for ATI to even consider nv . The battle will be platforms and Intel /AMD Will both have great platforms . Idon't think Intel or AMD are real worried about NV platform on a Via cpu.

Its going to be intersting to many. Just how the XT platiform from intel.The top nehalem system that uses qplink to the gpu actually works with NV products . It is possiable not even a single card NV unit will work on this platform. Intel may be cutting them out of the high end all together. Which makes sense considering how NV is with sli locking everyone else out. Look for AMD to do the same thing. NV needs to jump right by DX10.1 and go DX11 . On the next generation it is the 6 month of 08 now. 4 months till nehalem . Larrabee won't trail to far behind. Altho I have heard all sorts of dates for that .

I haven't every been A AMD fan but I must admit. they got my attention . K10 I LOL at . Whats going on with ATI is interesting and good for us all . Hector does have a plain .

I thought all the hoopla over the r600 was amazing . The lack of info on were gpus are heading is kinda eerie ALL 3 companies getting ready for DX11 and the big show is amazingly low info available.

I been readying all the forums and the stuff is great . Being on the outside looking in this round is hard its taken its toll on many posters. The guys on the inside looking out have done a great job of keeping the lid on things . Congrats guys.


One other thing that I have now been considering is . How AMD feels About THE VIA/NV relationship . NV isn't really a company you can ignore here . Ihope ATI.AMD /Intel understand this.

Because this changes alot of things for both AMD and intel . What if NV /Via actually come up with something really good?
Via starts taken market share. This is a double wammie for AMD . Because along with Intel they start loosing market share. First wammie . The second wammie is this would kill the idea of Intel being the only player if AMD goes down . It may also hurt AMD lawsuite against intel . I love tech and its never been more exciting than right now. EVEN pong can't touch this.

 

[Martimus]

edit: I am removing this because it doesn't really have to do with the original post. We should take the multip-GPU discussion to a different thread. I appologize, I didn't realize what thread I was posting to when I responded to BFG.

 

[cm123]

EVGA Corporation GEFORCE GTX 280 PCIE 1GB GDDR3 2PORT DVI-I HDTV
MFG#: 896-P3-1280-AR

its listed in disty (wholesale) in the $680's

one online company selling it now for $733.80

http://www.shopblt.com/partnum/u086_896p31280arh.html

Seems like way too much money - and I nearly always buy the lastest and greatest - maybe its just me though...

guess all the talk of it coming so cheap since the 4870 is so fast just not looking like that's going to happen - lets hope though the 4870 is fast as GTX 280 though, in hopes help drive pricing just a little.

As fyi... the disty lists shipping date of 06-28-2008 ( I can order it actually right now) and the company above lists 06-26-2008 - shows only 5 units coming into each warehouse which for this disty is sign not much coming 1st shipment on 06-28 (20 units total)

Working on getting more info - will post it if I can (and can release that info) - NDA should lift on Monday from what understand now.

sorry on the edit - got GTX 260 (very little for now) info too...

EVGA Corporation GEFORCE GTX 260 PCIE 896MB GDDR3 2PORT DVI-I HDTV
MFG#: 896-P3-1260-AR

no disty price yet - same online company selling it for $511.53

http://www.shopblt.com/partnum/u086_896p31260arh.html


$656 (from another disty) ASUS GEFORCE GTX280 1G DDR3 2*DVI-I 512BIT HDTV HDCP PCIE2.0
part# ENGTX280/HTDP/1G

$654 - PALIT VCX GTX280 1024MB GDDR3 DUAL-DVI HDCP HDMI & CRT TV-OUT VIDEO CARD
part # XNE/TX280+T305

$411 - PALIT VCX GTX260 896MB GDDR3 DUAL-DVI HDCP HDMI & CRT PCI-E
XNE/TX260+T394

remember all but the online companies pricing is from disty which is wholesale - normally that means our price about 5% higher than that (not always) at places like NE / ZZ and so on...

also, drop part number into google - of course will find few more things about for each card...



Merged into main thread.

Video Mod BFG10K.

----

Today Disty updated couple areas - one stock (in-stock) dates all changed to the 19th and the GTX 280 wholesale price changed to $648 for EVGA card -

The EVGA GTX 260 is not $477 and has stock date of 06-28-2008

Its looking as we'll see many more GTX 280's right off than 260's - pretty interesting to see that Nvidia feels they can or should be able to command such large price -

 

[BassBomb]

$411 - PALIT VCX GTX280 896MB GDDR3 DUAL-DVI HDCP HDMI & CRT PCI-E
XNE/TX260+T394

interesting

 

[cm123]

Originally posted by: BassBomb
$411 - PALIT VCX GTX280 896MB GDDR3 DUAL-DVI HDCP HDMI & CRT PCI-E
XNE/TX260+T394

interesting

that was my bad - when cut and paste thought I messed up (and did) - just paste again - its the 260.

 

[Extelleron]

Way overpriced at this point.

The latest rumors put GTX 260 at $399 and the GTX 280 at $499, and IMO even that might be overpriced.

HD 4870 will be $299, and HD 4850 CF for $400 should equal the GTX 280 in most cases (in Vantage Extreme, it will beat the 280).

See here: http://forums.vr-zone.com/showthread.php?t=287874

 

[taltamir]

amd doesn't have a great platform, AMD has a CHEAP but balanced platform... and intel doesn't have a great platform, intel has an awesome CPU with craptacular video.

If intel puts the hurt on nvidia by refusing to license nehalem chipset making ability to them, and nvidia gives in and trades SLI for it, it would be very bad for AMD.

Although nVidia could always go with via and their x86 license.

 

[cm123]

Originally posted by: Extelleron
Way overpriced at this point.

The latest rumors put GTX 260 at $399 and the GTX 280 at $499, and IMO even that might be overpriced.

HD 4870 will be $299, and HD 4850 CF for $400 should equal the GTX 280 in most cases (in Vantage Extreme, it will beat the 280).

See here: http://forums.vr-zone.com/showthread.php?t=287874

got to wonder though with the rumors now seeing the GTX prices is the AMD/ATI will stay that low? would GTX 280 being twice the costs of the HD4870 stand much of chance unless of course its wild performance over HD4870 - then have to ask, would Nvidia really bring to market the GTX 280 and 260 like this so much higher than ATI? or do they know more than rumors and hence the higher price?

 

[tuteja1986]

rape fest begins :! the 512MB 7800GTX all over again :!

 

[nRollo]

Originally posted by: tuteja1986
rape fest begins :! the 512MB 7800GTX all over again :!

It may be a little premature to forecast pricing and availability on the GTX260/280 as they haven't launched yet.

Often parts listed as "for sale" pre- launch command a higher price, because supply is short as most vendors honor NDA, demand is high.

 

[OCGuy]

If I see one more Inquirer link I might puke. The same people that said 8800GTX would need an extra external power brick before launch.......:disgust:

 

[Piuc2020]

I think ATI is going to win this generation, those are outrageous prices, crossfire HD4870s would cost less and (according to rumours) might end up being faster than a GTX 280.

This is nice because ATI will finally get back in the game and the fierce competition will make it even better for us customers.

 

[cm123]

Originally posted by: nRollo

Originally posted by: tuteja1986
rape fest begins :! the 512MB 7800GTX all over again :!

It may be a little premature to forecast pricing and availability on the GTX260/280 as they haven't launched yet.

Often parts listed as "for sale" pre- launch command a higher price, because supply is short as most vendors honor NDA, demand is high.

---

guessing you are nvidia staff or was or have buddy that is - hence the comment (I do not mean this bad in anyway, its totally fine) - Ingram Micro, ASI, MaLabs, D&H all listing these though at these prices, since they as guessing you already know have to ramp up, items are being posted right now for wholesale.

I can't say I've seen disty pre-sale item at higher costs (not to say it never happened) - of course the companies listed above being wholesale are bound normally by Nvidia launch partners to NOT over price the product - and sure NDA, however again, maybe EVGA, Asus and Pal. not too good about getting that signed by IGM and ASI (hence having the most info posted).

are you nvidia rep by chance? I have no nda (signed) as fyi... just in case you are

 

[Foxery]

Extelleron: I see where you're coming from, and in fact we'll see this concept in action eventually in the form of Intel Larabee. However, 4 small cores still cumulatively require the same power and produce the same heat as 1 large one. Manufacturing yield rises slightly, but at what cost to R&D and additional circuitry for inter-core communication? GPU architecture (and the nature of its work) may simply not lend itself well to being broken apart in this manner, either.

The advantages probably don't outweigh the difficulty in implementation. Otherwise, they'd be doing it.

GT200 power/heat concerns just boil down to nVidia's need to pack more transistors into this product while remaining stuck on a manufacturing process where this is unwise. A 55nm die shink will do amazing things for it. As for the next generation, well, we'll see...

 

[tuteja1986]

Originally posted by: Piuc2020
I think ATI is going to win this generation, those are outrageous prices, crossfire HD4870s would cost less and (according to rumours) might end up being faster than a GTX 280.

This is nice because ATI will finally get back in the game and the fierce competition will make it even better for us customers.

well looking at the early leaked benchmark , i ain't impressed with either of them (R4XX or GT2XX).


CRYSIS 1920x1200 AA OFF & High Quality settings :
GT260 : 29.75
GT280 : 35.75

$449 for the 260, $649 for the 280 is nvidia recommended price i think.

now to wait for detail article on 16th.

Rumors of yield problem maybe the reason for the high price as they can't pump out enough cards to meet the demand.

 

[Piuc2020]

Crysis just doesn't play well with the current architecture of cards, something is limiting cards in Crysis heavily and just incrementing shaders, rops, etc, linearly is obviously not going to fix the performance until the bottleneck is discovered and fixed.

 

[allies]

Originally posted by: Piuc2020
Crysis just doesn't play well with the current architecture of cards, something is limiting cards in Crysis heavily and just incrementing shaders, rops, etc, linearly is obviously not going to fix the performance until the bottleneck is discovered and fixed.

Poor coding is the problem IMO.