• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

NVIDIA Prepping dual-GPU GF104 card

Soldato
Joined
2 Sep 2006
Posts
13,483
Location
Forest Grove, OR, 'Merica
:/

http://www.guru3d.com/news/nvidia-prepping-dualgpu-gf104-card/
ss20100517195705.png
 
Hmm that spec would be a bit unspectacular - sure it would give the 5970 a good run but thats been out ages already... needs to be the equivalent of GTX470 SLI minimum.
 
GTX460 ~ HD5770 no?

So this is looking at what... HD5870 performance?

I just can't really see the point unless it's a pair of GTX470s.
 
most i gonna spend on a gfx card is 130 quid, thing is i got my gtx260 a lot less than 130 last year. and due to prices going up its gonna be hard for the new gen tech to compete with the old gen tech on price vs performance.

those who got the 5870 on launch date must be livid at the price they got it at.
i know it was cheapo but anyone got the exact figure the card went for before the price hikes in recession.?
 
lol, it would give a 5970 a good run for its money, no it ruddy wouldn't. A GF104, normally would be a simple half part of the top end, IF and its a HUGE if, they cut out all the cache and some other minor bits instead of being an aimed 256shader part they might bump it up one shader cluster, slim chance of 2 clusters, but realistically I don't see that happening. THe 450GTX or GTS or even a 460GTX won't be close in performance to the 465GTX, that will be called that for no other reason than it has to be named between the 470gtx and whatever the midrange is called.

I fail to see how a 2x 256shader part at ANY clocks will come close to a 5970. 2x 480sp parts fail to demolish a 5970 and fail to beat it in several games, 2x480gtx's and a 5970 at stock trade blows in some games, the 5970 at full 5870 clocks trade blows in even more games and win several more, I fail to see how 2x 256 or even 320shader parts with far less memory bandwidth(and less memory, meaning no random big wins due to the extra memory the 480gtx has).
 
most i gonna spend on a gfx card is 130 quid, thing is i got my gtx260 a lot less than 130 last year. and due to prices going up its gonna be hard for the new gen tech to compete with the old gen tech on price vs performance.

those who got the 5870 on launch date must be livid at the price they got it at.
i know it was cheapo but anyone got the exact figure the card went for before the price hikes in recession.?

Did you type that backwards or something, why would someone be livid at a 5870 at £300 on launch, did you mean people who bought them a month or two later when they hit £350 should be livid they decided not to buy at launch as that makes a heck of a lot more sense.

I got a 5850 for £193 about a week after launch, most were £200-210 until a month later when most places started pushing prices up by £20-30.

The 5870 was £300 on launch, but not good value compared to the 5850.

THe recession was around about a year BEFORE the 5870 launched though, 2009 was the recovery year, it was Q 3/4 2008 that was the massive crash prices didn't increase due to the recession but due to retailer greed and taking advantage of a stock crisis thanks to TSMC basically binning a 6 week supply of wafers due to basically an unheard of industry first error on a massive scale.

But you're right, economy + pounds strength + TSMC making a truly awful 40nm process + a slightly larger core than last time around all meant that a 4850/4870 situation was simply not possible this generation.

The 4870/4850 was the peak of almost perfect process, the absolute smallest core they are likely to get away with making for years to come, strong pound and very good yields.

A stronger pound, a better yield and process and AMD should be pushing future generations hopefully to pricing inbetween the 4 and 5 series launch prices, thats not going to happen with the refresh though, but the next full new gen if its done at GloFo could be a lot cheaper than the 5 series.
 
seeing as how a single gtx480 beat a 5970 in most cases bar a few benchmarks the new card could be on to something.

Considering thats the biggest pile of rubbish I've read on here in, well, months if not years I really shouldn't respond, but I will, a single 480gtx is not close to beating a 5970, sure there will be a couple situations(probably physx induced) in Nvidia titles it gets close, there are also games where the 5970 stretches its leg aswell and absolutely blows Nvidia away, thats life. On average the 5970 is miles ahead of a 480gtx and not surprisingly, a 5970 is not particularly close to double the speed of a 5870, reduced clocks and not 100% scaling in Xfire. However SLI suffers from nowhere near 100% scaling and it looks set that the GF104, will be two half 480gtx's stuck together, like 5770 xfire, great in some situations, particular lower res, but pushed at high res can't offer the high minimums or same ultimate performance a 5870 can in the toughest of games.

Does anyone expect 2x 5770's to somehow come close to 480gtx SLI, no, because its a truly ridiculous notion, as is 2x 450gts's(assuming a name obviously but Nvidia's 5770 equivilent) being close to 5870 Xfired.
 
lol, it would give a 5970 a good run for its money, no it ruddy wouldn't. A GF104, normally would be a simple half part of the top end, IF and its a HUGE if, they cut out all the cache and some other minor bits instead of being an aimed 256shader part they might bump it up one shader cluster, slim chance of 2 clusters, but realistically I don't see that happening. THe 450GTX or GTS or even a 460GTX won't be close in performance to the 465GTX, that will be called that for no other reason than it has to be named between the 470gtx and whatever the midrange is called.

I fail to see how a 2x 256shader part at ANY clocks will come close to a 5970. 2x 480sp parts fail to demolish a 5970 and fail to beat it in several games, 2x480gtx's and a 5970 at stock trade blows in some games, the 5970 at full 5870 clocks trade blows in even more games and win several more, I fail to see how 2x 256 or even 320shader parts with far less memory bandwidth(and less memory, meaning no random big wins due to the extra memory the 480gtx has).

I was talking about the proposed 2x352(?) SP part... sure 2x256 wouldn't be that great - and mostly pointless.
 
I don't understand the thinking behind two fairly average GPU's on the same card as you'd be better off buying a faster single GPU instead, SLI/Crossfire is only really necessary when there isn't a single GPU solution fast enough.
 
I was talking about the proposed 2x352(?) SP part... sure 2x256 wouldn't be that great - and mostly pointless.

The GF104, as yet, no ones mentioned specs, but normally the 100 will be the full end part, the 104 would be the mid end part(which is almost always half the shader count(5870 1600 sp's, 5770 800sp's etc), the 106 would be the value/performance part and the 108 would be the part that has, well, video out and enough shader power to play tetris in low res :p

Its incredibly unlikely the GF104 would have 352 shaders, thats a 465GTX and its a cut down GF100.

http://www.fudzilla.com/content/view/18822/1/

This doesn't seem wholy unlikely, half the same specs but its 50W lighter in power(and likely a bigger gap than that as Fermi "power ratings" are BS basically. The 465gtx is CLEARLY a cut down Fermi, the 460GTX(450gtx, 450gts, yet to see what they call it) is based on a different core, yet will have the same 256bit bus, the same 1gb gddr5, so the 50W power saving is coming from somewhere and frankly, the only thing left is shaders. Its using 25% less power, 352x0.75 = 264shaders. In reality its incredibly likely to be a 256 shader part possibly with slightly higher clock speeds as it should offer better power usage than a cut down Fermi as something designed to be 256shaders and 256bit bus will use less power than a cut down part at the same specs. Though I doubt it would be any more than 10% higher in clocks because the process itself is still simply bad, and they can't make a £150 card any closer in performance than that to a £220 card, well if they hope to sell any at £200-220 that is.

Which again brings it back to being a dual GF104, which will make it a dual 256shader part, at any clocks, its not going to be in the same ballpark as a 5970, it probably won't be faster/significantly different to a 480GTX, and power wise both are going to be in the 300W ballpark. However due to yields, its very likely this card will be aimed at £300 and that they can produce a card that is £300 ish and offer similar to 480gtx/5870 performance.

It gives them a major boost in performance/pound, purely from the natural increase in yields you get from a much smaller core, on top of that I'd expect everything from GF104 and up to have "40nm process fixes" included so if they could get 2.5-3x the yield, they could actually make a profit on a dual gf104, that offered around 480gtx performance, for cheaper.

Of course, all the people that have for years been complaing AMD could only beat Nvidia's offering with a dual gpu part might be screwed if Nvidia stopped producing 480gtx's and instead sold a cheaper dual smaller core part.

Even more significantly is, what route will Nvidia finally be forced to take in the future.

If it turns out 2xgf104 produces similar speed, same power, but can be sold cheaper AND at a profit and give AMD a real run for its money in terms of pricing and yields......... surely thats proof to Nvidia their large die designs simply aren't working.
 
I don't understand the thinking behind two fairly average GPU's on the same card as you'd be better off buying a faster single GPU instead, SLI/Crossfire is only really necessary when there isn't a single GPU solution fast enough.

Cost, I thought I'd just sum up quickly instead of the big post.

Fermi is supposedly around 100 dies per wafer, and lets call it 20% yield and call it 20cores per wafer, at $5000 per wafer, thats $250 a core, and its not profit at anything above $250, theres R&D, wages, etc to factor in so it won't be in profit till beyond $300-350, when you tack on memory, pcb, cooling, warrantys(they are likely to fail more than any other card for a while due to power and heat) theres next to no profit.

Now you take a core half the size, already you're looking at 200 wafers per die, but with much smaller cores you get exponentially higher yields, nor proportionally higher. So even if it was identical to Fermi but smaller you'd be looking at probably 50% yields plus, so you'd be getting 100 cores per wafer which is $50 a core, take two of them and its $100 for the cores for a dual GF104 card, instead of a $250 core for a 480gtx and in the end you'll end up with similar performance.

That is why AMD's small core strategy is wiping the floor with Nvidia, and no other reason. The reason Nvidia's low end is selling as well or better than AMD's, AMD's small core and Nvidia's are SO small with such high yields and where performance is secondary to simply being able to output video at all, theres little difference between them.

Smaller cores = great profit and lower cost, its that simple and this will likely be the generation where Nvidia finally notices that and adjusts, big time. The even bigger problem though is that GPU's start the design process years in advance, so even if Nvidia decide to go for a small core strategy, it could be 2 years before we see it, or they could have adjusted 2 years ago and Fermi 2 next year is already going to be way smaller.

But notice I also said thats if its identical to Fermi, just smaller, if Nvidia has included the 40nm process fixes, more via's, better alignment of transistor to allow for the variable transistor size on the process(that is a HUGE problem), you might add 10% to the size, but gain 20% in yields.

Meaning they could push it up to a 70% yield part at 180 cores per wafer, which brings you to 126working cores per wafer instead of 100, again pushing the price down as the wafer always costs the same, which saves Nvidia another $10 per core.

But theres ANOTHER upside, at 126 working cores, or 100 core per wafer, even if they use it in a dual card, thats 50-63 dual cards made out of every wafer, instead of only 20. It also massively increases the amount they can produce, because 40nm production is maxed out, there are only 9k wafers for both companies a month at the moment, so if you can get more out of the wafers you can make, thats more cards, more sales and more profit.

A Dual GF104 might not increase performance for any of us end users, but it could bring 480gtx performance, at £300, with 3 times the availability and MOST importantly, meaning every card makes Nvidia a profit and a pretty big profit at that.
 
drunkenmaster...please learn to cliff notes your posts :) for the benefit of us people who b 2 laz3 to reead
 
Sort of off topic but am curious about something. A couple of people on this thread have mentioned 2 gtx480's trade blows with a single 5970 at stock speeds. However the latest reviews from people like Guru3d have shown excellent scaling in SLI and the gtx480 sli even the gtx 470 sli outperforms 2 5870's in crossfire in a lot of games. And 2 5870s in crossfire outperforms a 5970 obviously. So how are they trading blow for blow ? Just curious about peoples views.
 
Last edited:
Back
Top Bottom