GT 4xx Fermi cards = Compute Capability 1.0 ?! This is what the nVidia page says. Mistake?

Hi,

I am suddenly confused about which Compute Capability is supported by a given card.
The following link

http://developer.nvidia.com/cuda-gpus

says:

GeForce GT 430: Compute Capability 1.0
GeForce GT 520: Compute Capability 1.2

whereas the (presumably less accurate) Wikipedia page ( http://en.wikipedia.org/wiki/CUDA#Supported_GPUs ) says:

GT 430: Compute Capability 2.1

I’m confused. Could somebody here clear this up for me?

Thanks,
Fred

The one on Nvidia’s site is the correct one. Both GT 430 and GT 520 are low-end cards, and as such they cannot support the more advanced features of CUDA available in higher-end and more expensive cards. The Compute Capability version can be checked using a software utility called CUDA-Z if you wish to test your own card.

Now, all this means that if a CUDA application requires 1.3 compute capability, then it cannot run on cards with compute capability lower than 1.3. Backwards compatibility is retained though, so a CUDA app made for 1.0 compute capability cards will work just fine on cards with higher compute capability.

Basically what you’re saying is that this is a matter of power (GT 430 vs GTS 450), and not generation (GT 240 vs 440).

You’re saying a GT/GTS/GTX 4n0 card’s compute capability basically increases as ‘n’ increases.

But look: the GT 520’s compute capability is 1.2, and this card is less powerful and expensive than the GT 430.

Even more surprising, the 9500 GT’s CC is 1.1, but it is 1.0 for the 9600 and 9800 GT.

And the 9800 GT’s CC is 1.0 vs the 8400 GS’s is 1.1.

Not to mention the GT 430’s CC is kind of abnormally low (1.0) for a Fermi-based card.

That doesn’t sound right. Does it?

What makes a card capable of any particular level of CUDA support other than mere core counts, eh?..

I’d really have thought Nvidia would have fixed the webpage since this thread. Nvidia doesn’t manufacture any compute capability 1.x chips since a long time ago, and I’d doubt they pile that much inventory.

A lot. Compute capability 1.x and 2.x devices have different, binary incompatible architectures.

I was under the impression that Nvidia purposely disabled advanced CUDA features on their low end cards by default, just like they disabled hardware PhysX support on GF 210 even though the card is capable of it after enabling it through a registry tweak. My apologies if I’m mistaken and that it’s due to Nvidia’s incompetence in maintaining information on their website. :tongue:

I’m under the impression that Nvidia purposely misleads people looking for cheap CUDA-capable cards. :biggrin:

If that’s not the case I’m sure they’ll fix the website soon after finding this thread.

Hehe, I certainly wouldn’t be surprised if they pulled something like this on purpose. After all, they’ve given us the GT230, GTS240, GTS250, GeForce 100 series, and GeForce 300 series to make uninformed customers believe that they’re getting state-of-the-art cards when in fact they’re getting re-branded previous generation cards. :thumbsdown:

And thus the focus behind the question, to which the OP might’ve given some consideration…

1.1 supported concurrent{kernel execution and memcopy}, Integer Atomics
1.2 supported shared mem atomics, floating atomics
1.3 supported double precision
2.0 supported {concurrent kernel execution}, dual PCI DMA (read & write simultaneously),
more shared memory, L1 cache etc…

So, the compute capbility is just not about the cores…