I'm beginning to realize all the grumbling about "slightly increased performance for same price, shitty generation" really could be related to the alleged fact process shrinks are no longer reducing transistor prices that much.
Although, 580's are now as low as 359 on newegg, and that was a $550 card not long ago, so thats at least a little impressive for the SI/Kepler gen.
Although, 580's are now as low as 359 on newegg, and that was a $550 card not long ago, so thats at least a little impressive for the SI/Kepler gen.
I for once would like AMD and nVidia to continue push towards GPGPU acros their GPU ranges. Especially with AMD focus on APU's I can't see them going back to graphics oriented cores.
Imagine if in few years with new consoles build around GPGPU cores we will get game engines properly utilizing new architectures. Then going graphics only will look silly and I think performance gap to older, less efficient compute cores will be massive. Just check Civ5 scores for HD5xxx/6xxx and new HD77xx+ or nVidia as an example.
We can't be short sighted and only look at current games compelty forgetting about evolution and our targets for future.
Imagine if in few years with new consoles build around GPGPU cores we will get game engines properly utilizing new architectures. Then going graphics only will look silly and I think performance gap to older, less efficient compute cores will be massive. Just check Civ5 scores for HD5xxx/6xxx and new HD77xx+ or nVidia as an example.
We can't be short sighted and only look at current games compelty forgetting about evolution and our targets for future.
That's better than nothing, as far as news go. Thanks.
I wonder if the 670 Ti will come with 1 SMx disabled or 1 GPC disabled.
According to the previous gen policy, it should be the first and I certainly hope this will be the case.
Computerbase concluded that the GTX 680 is 55% faster than my 570 at 1920X1080+4XAA which are the settings I game at, so a supposed -1 SMx GTX 670 Ti, should be around 40% faster at GTX 680 clocks, which is the absolute minimum I can tolerate in order to upgrade. Ah what the hell, throw in another 10% overclock and you are around 50% more performance than a 570.
I also wonder how the rest of the bunch will be shaped. Maybe 1 GPC disabled for the 660 Ti and 192 bits bus or would it be something better than that?
I'd like things to be like that
GTX 670 Ti = GK104 - 1SMx at 900Mhz / 256 bits bus (with the ability to be stable at 1.1Ghz, with good radial HSF, not being noisy as a jet at that frequency)
GTX 670 = GK104 - 1 GPC at 850Mhz / 256 bits bus
GTX 660 Ti = GK104 - 3 SMxs at 850Mhz / 192 bits bus
GTX 660 = GK104 - 2GPCs at 900Mhz / 192bits bus
Although I am very skeptic about the last one. Is there a point selling a chip, which is half disabled?
NathansFortune
Regular
How big is the 106 die? Is it comparable to 7870?
That's kind of inevitable when you ship the elephant's trunk and don't talk about the elephant, isn't it?
The real fear has always been pricing -- nvidia (then) will have no choice but to charge a lot of money for their larger dies, or (now) they'll have no choice but to charge a lot of money for compute-capable dies because the market will be smaller.
These tests don't make use of shared memory, maybe LuxMark does.
Also in these demos there is little thread divergence, something that happens a lot with path tracing, I think, as it does like random incoherent raytracing per pixel, afaik.
iMacmatician
Regular
There was a 256 CC GF100 Quadro, but maybe it makes less sense to sell such a chip for the consumer space.
Also, regarding 192-bit GK104s, isn't the GK104 (at least the 680) somewhat bandwidth limited?
Nobody (who can say) has any hard information as far as I know, but if I use the 294 mm^2 die size of GK104, my estimate for GK107's die size (≈121 mm^2), and if the scaling from GK104 to GK106 is the same ratio as from GK106 to GK107, then that gives an estimated ~189 mm^2 for GK106.
Would that make sense if GK106 has specs in the middle of GK104 and GK107, that is, 768 CCs and 192-bit bus?
That would be my guess as well. And I'd expect a June release. GK110 should come much later.
Furthermore Mandelbrot/Julia 2D and Julia 4D,don't make use of memory bandwidth or even L1/L2 cache at all, so not hindered by some apparent Kepler weaknesses.
The fluid 3D, relies a lot on volume texture sampling, which Fermi was like flawed at. Doubling texture units, higher clock, and probably better trilinear hardware volume texture sampling fixed that.
(BTW 7970 still better at fluid 3D, 60 fps versus 40 fps on Kepler.)
Oh I see.
So GTX 660 Ti=Castrated GK104
GTX 660 = GK106?
I missed that about the Quadro, since they are out of my interest bubble, but it's good to know, thanks.
As for the supposed 192bit GK104 version, well my supposed GTX 660 Ti, would be deprived of 1/4 of its bandwidth, while being even more castrated on the SMx side (3 out of 8) so the resulting bandwidth should be enough, since GTX 680 proved that it can stand its ground against the competion, with a 8SMxs to 256bit bus ratio. 5 SMxs to 192bus ratio would provide even more bandwidth per SMx, depending on the memory clock of course.
Dont forget they offtly just do a compilation of speculation done in their rumor thread.
But how a card with half Cudacores of the 680 will end ? clock will matters but this seems low ( seeing it should goes against 7870, who is at GTX580 level performance wise ). ( well 3Dcenter bring perf at 6950 - 7870 level, and 560TI level ( i will be happy they explain me how they put the 6950, GTX560TI and 7870 on same level of performance. 7870 is more close of 6970-580 ( and again, faster in many case of the 580 )
But how a card with half Cudacores of the 680 will end ? clock will matters but this seems low ( seeing it should goes against 7870, who is at GTX580 level performance wise ). ( well 3Dcenter bring perf at 6950 - 7870 level, and 560TI level ( i will be happy they explain me how they put the 6950, GTX560TI and 7870 on same level of performance. 7870 is more close of 6970-580 ( and again, faster in many case of the 580 )
Last edited by a moderator:
I dont' write newsblurbs at 3DC obviously but I'd think you might read and understand german. If yes and you can't interpret terms like "Prognose" and "Richtung" as it was meant, then you might want to re-read it. It obviously doesn't render in absolutes and gives an estimate for a specific performance ballpark like in between up to A and down to B.
Yes we "speculate" a lot at 3DC (tragic irony to mention something like that in a thread like this) and the site is known of it's really "bad track record" when it comes to such speculations. Considering the typical far worse nonsense that floats around the net you might want to ask yourself how serious 3DC is about such matters.
I dare not say. NV naming schemes are a mystery to me.
this is a rather large performance gap, this is what i should have maybe say in first. At this point i could have use 5850 to 6970 .
( this said this is not an aggression against 3Dcenter, i know some peoples there and respect them, just i dont like the idea to make complilation of speculation on the first page, but that's my personal opinion .)
This said, this let us with something: what can bring on the table a 768Cuda cores chips when we know how act the 1536 CC parts now ? even with outrageous clock? I can easely imagine a GK106 with 30% less CC, but i have some pain to imagine what can be the performance with half of it. Why dont keep the 256bit bus with 2gb. We can imagine this card should goes against 7870 and replace the GTX580, does an half 680 could do it ? Or does Nvidia decided to decline the 680 differently maybe ( 3 cards in the upper line up )
Last edited by a moderator:
Then go back to some Pitcairn reviews and ask yourself why that one with merely 20CUs can come as close as it does to the Tahiti salvage part. There are other examples that would work more or less the same way within the Kepler family of products, but it's way too early for that.
