NVIDIA to Counter Radeon HD 6970 ''Cayman'' with GeForce GTX 580

[safan80]

From TPU

I hope this is true I want to see cards with 512 cuda cores and for nvidia to give a damn about performance. The GT430 is such a joke. Nvidia this year has let everyone down. I was expecting a card faster than the 5970 from nvidia, boy was I wrong.

AMD is undertaking its product development cycle at a breakneck pace, NVIDIA trailed it in the DirectX 11 and performance leadership race by months. This November, AMD will release the "Cayman" GPU, its newest high end GPU, the expectations are that it will outperform the NVIDIA GF100, that is a serious cause for concern, for the green team. It's back to its old tactics of talking about GPUs that haven't even taken shape, to try and water down AMD's launch. Enter, the GF110, NVIDIA's new high-end GPU under design, on which is based the GeForce GTX 580.

The new GPU is speculated to have 512 CUDA cores, 128 TMUs, and a 512-bit wide GDDR5 memory interface holding 2 GB of memory, with a TDP of close to that of the GeForce GTX 480. In the immediate future, there are prospects of a more realistic-sounding GF100b, which is basically GF100 with all its 512 CUDA cores enabled, while retaining its 384-bit GDDR5 memory interface, 64 TMUs, and slightly higher TDP than that of the GTX 480.

[Chrono Detector]

Specs sound nice, wonder if its true, though I suspect that this card might even be more power hungry than the GTX 480.

[h0bbes]

Probably released at least a few months after cayman..

Edit: google translation was a bit confusing but a December launch would be nice. Hopefully this won't be like the original Fermi launch

[safan80]

The German sources they grab the intel from are here:

http://www.3dcenter.org/news/2010-10-13

http://www.pcgameshardware.de/aid,79...fikkarte/News/

[flopper]

So its out when amd using 28nm?
Right...

[Pantsu]

Calling a 512sp Fermi a GTX 580 would be in line with the AMD renaming.

Seriously, this rumor has been thrown around since the release of GTX 480. I still don't see anything concrete though, just guesswork. I suppose they will do a kicker at some point, but I doubt it's just going to be a 512sp fermi gf100b. That should have come out a long time ago, and in any case it would only be like 5% faster.

[Olivon]

So its out when amd using 28nm?
Right...

When AMD will use 28nm ? With Turks, that's it ?

[Svnth]

Calling a 512sp Fermi a GTX 580 would be in line with the AMD renaming.

Seriously, this rumor has been thrown around since the release of GTX 480. I still don't see anything concrete though, just guesswork. I suppose they will do a kicker at some point, but I doubt it's just going to be a 512sp fermi gf100b. That should have come out a long time ago, and in any case it would only be like 5% faster.

Not really. If it's a respin of the chip and can get higher performance than GF100 ever could than it's as "new" as a slightly tweaked architecture on the AMD side.

This isn't a feature battle given that it's all still DX11 so new architecture isn't really meaningful if it's not significantly faster. Also there's no standard for labeling something a "new architecture". Tiny changes could justify a respin which could arguably called a new architecture. I'm willing to bet the new ATI products aren't drastically changed. So just because they're "different", is it really a "new architecture"?

G92 wasn't that much different than G80, architecturally, but it was a justified "generation increase" because it brought something new to the table.. value, and it was a full top to bottom stack. Hence it was justified in being differentiated.

[Dimitriman]

Those specs are impossible at 40nm.

fermi 100b based on gf104 with 512 cores is about the only thing nvidia can do right now till 28nm and by that time amd will also be on time with a fully loaded northern islands chip as well at 28nm

512 cuda 128tmu and 512 bit @ 40nm would be bigger than a belgium waffle.

[snoid_zero]

Calling a 512sp Fermi a GTX 580 would be in line with the AMD renaming.

Nope. Based on the article if true, 512sp GTX 580 will have 128 TMUs, and a 512-bit wide GDDR5.
That's 64 TMUs more and 128-bit wider than GTX 480.
This must be something else.

[zalbard]

So, when is it coming? In 6 months?

[xBanzai89]

That 512-bit sounds like total overkill pairing it with GDDR5. If it actually comes out with that and increases performance what does it matter I guess then.

[Dami3n]

512 bits + GDDR5 doesn´t sound reasonable at this time. A waste of money IMHO

[SKYMTL]

Why would NVIDIA need to complicate their lives and add sufficient memory controllers to allow for 512 bit?

Remember, GF100 in its FULLEST form has a 6x64-bit layout which allows for a maximum of a 384-bit memory interface. If they wanted to add another TWO 64-bit controllers they would not only need to find the space for them on an already massive chip but they would also have to realign the ROP / Cache distribution since both of these items are closely associated with the memory controllers.

[Zloyd]

So if they manged higher clocks with the full 512 shaders etc this would be like what 10-15% faster than a 480 , basically more or less equal to Cayman(going by speculation that cayman is 40% faster than a 5870) but probably hotter and still a powerhog. Hope they call it 485 and not 580.

[-Boris-]

A GTX 580 with those specs wouldn't compete with cayman. And it would be much harder to build than GTX 480, which they still haven't been able to release a full version of.

[Olivon]

512-bit memory controller is way too high.

576SP/96TMU/384-bit would be nice

[Carfax]

So if they manged higher clocks with the full 512 shaders etc this would be like what 10-15% faster than a 480 , basically more or less equal to Cayman(going by speculation that cayman is 40% faster than a 5870) but probably hotter and still a powerhog. Hope they call it 485 and not 580.

Do we even know how many SPs Cayman will have?

[iMacmatician]

512 CUDA cores would mean 32 CC/SM (or a multiple of 32). So no GF104 type SM, or is NVIDIA changing it again for 64 CC/SM?

So I'm assuming the memory controller still can't get significantly above 3.6 GHz, since the memory bus has been widened? 256-bit at 6 GHz would be 83% of 512-bit at 3.6 GHz, while allowing for a smaller die (and 384-bit at 6 GHz = 125%).

[kgk]

The German sources they grab the intel from are here:

http://www.3dcenter.org/news/2010-10-13

http://www.pcgameshardware.de/aid,79...fikkarte/News/

GF110 mit 768SE neuer Chip mit 768 Shader-Einheiten und 128 TMUs an einem 512 Bit DDR Speicherinterface, ohne den GF100-Ballast nach GF104-Vorbild, Die-Fläche mit 550-650mm² größer als GF100 und damit auch höhere Verlustleistung, Zielmarkt dieses Grafikchips wären dann eher die DualChip-Lösungen von AMD
Potential gegenüber GeForce GTX 480: ca. +50%



Chances of that happening on 40nm though?

But holy hell, 768 shader, 128 TMU, 512bit with 2GB? Would love something with that kind of performance in single GPU.

Find it a little annoying that articles like the OP's include Dual GPU cards when talking about "performance king". That kind of stuff needs an asterisk. Single GPU power should be where the real credit is given, not that I'm a fan of NVidia.

[PatRaceTin]

Those thing will be overheat

[Manicdan]

lol at 512bit

sounds like its gonna be super hot

[ElSel10]

fermi 100b based on gf104 with 512 cores is about the only thing nvidia can do right now till 28nm and by that time amd will also be on time with a fully loaded northern islands chip as well at 28nm

So basically, at this point there is absolutely no way for Nvidia to catch up with AMD, and within a few generations of this lagging behind, they will likely go bankrupt?

[Dami3n]

GF110 mit 768SE neuer Chip mit 768 Shader-Einheiten und 128 TMUs an einem 512 Bit DDR Speicherinterface, ohne den GF100-Ballast nach GF104-Vorbild, Die-Fläche mit 550-650mm² größer als GF100 und damit auch höhere Verlustleistung, Zielmarkt dieses Grafikchips wären dann eher die DualChip-Lösungen von AMD
Potential gegenüber GeForce GTX 480: ca. +50%



Chances of that happening on 40nm though?

But holy hell, 768 shader, 128 TMU, 512bit with 2GB? Would love something with that kind of performance in single GPU.

Find it a little annoying that articles like the OP's include Dual GPU cards when talking about "performance king". That kind of stuff needs an asterisk. Single GPU power should be where the real credit is given, not that I'm a fan of NVidia.

That card could be a full GF104 dual. 384 CUDA Cores x2, 64 TMUs x2, 256 bit x2 (marketing), 1GB GDDR5 x2.

[zalbard]

That card could be a full GF104 dual. 384 CUDA Cores x2, 64 TMUs x2, 256 bit x2 (marketing), 1GB GDDR5 x2.

Okay, now that starts making more sense. Sounds likely.