NVIDIA GF100 & Friends speculation

[aaronspink]

Well according to Nvidia, they're a "software company" as they have more software engineers than hardware engineers

That would explain a lot

 

[DavidGraham]

No disabled GPCs in GTX 465 ?:

I think that way they kept all of their raster engines intact , will that make a difference ? is that arrangement even balanced architecturally ?

http://www.brightsideofnews.com/news/2010/5/19/final-nvidia-geforce-gtx-465-specs-are-out.aspx

 

[Jawed]

End of the road for GTX470?

http://www.kitguru.net/components/graphic-cards/faith/end-of-the-road-for-gtx470/

On Wednesday, KitGuru heard that, for now, no more orders are being taken for the nVidia GeForce GTX470.

[...]

KitGuru UPDATE: Word has now reached us of a 375w, dual-GPU, GTX490. If true, then that could explain any reluctance to take new orders for the GTX470. nVidia could have choosen to move its GTX470 cores to the new GTX490 card, while it positions the 200w GTX465 chip against AMD’s Radeon HD 5850.

This website seems to be fishing for hits though:

http://www.kitguru.net/components/graphic-cards/zardon/nvidia-to-drop-cuda-on-gtx465/

All things being equal, how do you make a product run faster? Easy. You make it simpler. Right now, there is a rumbling in Asia that nVidia will drop CUDA support.

 

[Dade]

http://www.kitguru.net/components/graphic-cards/faith/end-of-the-road-for-gtx470/
This website seems to be fishing for hits though:

http://www.kitguru.net/components/graphic-cards/zardon/nvidia-to-drop-cuda-on-gtx465/

In my opinion, it doesn't make any sense, if the 465 can support OpenCL, it can support CUDA too.

May be they could try to differentiate the products ? (i.e. do you want to run a CUDA application ? You have to buy one of the expansive cards).

 

[psolord]

They could always disable CUDA, but I don't see the reason why. I mean not a hardware reason at least. These chips are still GF100 based no?

 

[3dilettante]

Yields have to be pretty horrid if the second-tier salvage part doesn't have the volume to satisfy more than the niche of a niche of a niche for a 375W board.

I don't get the CUDA thing at all. There's nothing physical to be pared away that suddenly makes CUDA non-functional.

 

[Jawed]

I tend towards the view these are bullshit news style articles. I'm not sure, is this website new and trying to get itself on the map?

 

[neliz]

I tend towards the view these are bullshit news style articles. I'm not sure, is this website new and trying to get itself on the map?

sounds like they're rehashing old pcpop/chiphell articles or something, according to them, the "news" is a 375W GTX490..

Heck, they could've taken the rumors from everywhere.

 

[Blazkowicz]

No disabled GPCs in GTX 465 ?:

I think that way they kept all of their raster engines intact , will that make a difference ? is that arrangement even balanced architecturally ?
http://www.brightsideofnews.com/news/2010/5/19/final-nvidia-geforce-gtx-465-specs-are-out.aspx

here's how I interpret it : in the event a raster engine is hit, then you'd lose all four multiprocessor from one quarter of the chip. after that you can only afford losing one multiprocessor on the rest of the chip.

So (one could calculate the probabilities) there wouldn't be that many qualified chips with a missing GPC (it's good for many defects in one quarter, less for more spread out defects), and/or it wouldn't be worth the trouble of releasing 3-GPC boards (a significant difference that firmware and drivers would have to care about). two different boards under the same name.

3-GPC and 4-GPC cards wouldn't be equal in performance either (meaningful or not, I don't know)

 

[Colourless]

Well according to Nvidia, they're a "software company" as they have more software engineers than hardware engineers

Yeah well, what software are the working on... consumer drivers, or HPC software. Of course the QA for these drivers would be a nightmare. Eyefinity is not a sure thing to work with all titles and Nvidia may want to be sure "it just works" with all titles.

 

[rpg.314]

http://www.kitguru.net/components/graphic-cards/faith/end-of-the-road-for-gtx470/

Dunno about this.

This website seems to be fishing for hits though:

http://www.kitguru.net/components/graphic-cards/zardon/nvidia-to-drop-cuda-on-gtx465/

This is fudzilla quality reporting.

 

[Lonbjerg]

Yeah well, what software are the working on... consumer drivers, or HPC software. Of course the QA for these drivers would be a nightmare. Eyefinity is not a sure thing to work with all titles and Nvidia may want to be sure "it just works" with all titles.

AFAIK most are involved in TWIMTBP team.

 

[Jawed]

Nvidia GeForce GTX 465 Specs and Launch Date

http://www.bit-tech.net/news/hardware/2010/05/20/nvidia-geforce-gtx-465-specs-and-launch-dat/1

I'm not convinced Bit-tech is doing much better The artist's impression isn't even post-ironic.

 

[Kaotik]

Yeah well, what software are the working on... consumer drivers, or HPC software. Of course the QA for these drivers would be a nightmare. Eyefinity is not a sure thing to work with all titles and Nvidia may want to be sure "it just works" with all titles.

Making sure "it just works" is just impossible due various engine limitations in game engines, which is why Eyefinity doesn't "just work" in all titles either

 

[Rys]

here's how I interpret it : in the event a raster engine is hit, then you'd lose all four multiprocessor from one quarter of the chip. after that you can only afford losing one multiprocessor on the rest of the chip.

I'm of the opinion that the ROPs aren't coupled to a cluster like that anyway. I bet you could lose all the ROPs and keep all of the ALU intact in GF100.

 

[KimB]

In my opinion, it doesn't make any sense, if the 465 can support OpenCL, it can support CUDA too.

May be they could try to differentiate the products ? (i.e. do you want to run a CUDA application ? You have to buy one of the expansive cards).

Well, I wouldn't be horribly surprised to see nVidia drop CUDA at some point, considering that from what I understand there is very little reason to program for CUDA instead of OpenCL on nVidia hardware.

 

[trinibwoy]

I'm of the opinion that the ROPs aren't coupled to a cluster like that anyway. I bet you could lose all the ROPs and keep all of the ALU intact in GF100.

I think he was referring to the rasterizers not the ROPs.

 

[Sontin]

Well, I wouldn't be horribly surprised to see nVidia drop CUDA at some point, considering that from what I understand there is very little reason to program for CUDA instead of OpenCL on nVidia hardware.

CUDA ist their interface to the machine code/language. There is no "CUDA programming". You can use C, C++, Fortran, OpenCL, DC, PhysX...

 

[rpg.314]

CUDA ist their interface to the machine code/language. There is no "CUDA programming". You can use C, C++, Fortran, OpenCL, DC, PhysX...

I would guess that you are not a programmer either. Right?

 

[btdvox]

CUDA ist their interface to the machine code/language. There is no "CUDA programming". You can use C, C++, Fortran, OpenCL, DC, PhysX...

CUDA is an API, API's are usually a Library to Program with. You mentioned PhysX which is also an API.

C, C++ are Programming Languages.

As the person above stated, I'm assuming your not a programmer. Because CUDA way, way higher than machine level.