CUDA vs. SLI Any performance difference?

Accelerated Computing CUDA CUDA Programming and Performance
1

I understand CUDA is meant to take advantage of multiple GPUs for parallel processing, but let’s say you have one graphics card,
with one GPU, or even a SLI configuration with two cards in one machine, that does not use CUDA.

Would the performance to be the same in these cases?

2

CUDA is meant to take advantage of any number of GPU’s in your system (1 or more), and cannot make use of multiple GPU’s if they are connected in SLI. The parallel processing part comes from the nature of the GPU. It is much less flexible than a standard CPU, but as a result, needs fewer transistors to work (meaning that you can fit many more ‘cores’ on a single die).

3

I understand that, but what I’m trying to say, is that, if you consider a case where parallelism is not a factor, is the performance

the same with or without CUDA? I am just curious.

4

What is the difference between an apple???<<

You can’t compare CUDA with SLI.

I don’t understand your question. In the domain of (3D) graphics (Direct X or OpenGL) parallelism is always a factor. So an SLI-bundle of GPUs – in comparison to a single GPU – may accelerate (e.g. frames per second) your graphics application.

In the domain of GPGPU one’s goal is to write (general purose) parallel algorithms to be executed on the GPU. CUDA is the programming framework for developers to write such algorithms. BTW you mustn’t bundle your GPUs with SLI if you want to harness the power of any GPU with CUDA-programms. If parallelism is not a factor a single CPU thread will be faster than a single GPU thread. ROUGHLY speaking, this is 'cause a single scalar processor and the memory management of a GPU is much more primitive than the CPU of your host computer.

Regards

Navier

5

What performance are you taking about? Graphics performance (framerates or polygons per seconds or something)? Because CUDA (or the lack of it) effectively has nothing to do with the graphics performance of a given card - except in the sense that the last two generations of GPUs from NVIDIA support CUDA and older ones do not.

CUDA is a completely separate feature which allows programmers to write and run computationally intensive non-graphics functions on the GPU. PhysX is probably the highest profile example of this, but people around here are doing stuff like computational physics, chemistry, biology, and real time image analysis on their CUDA GPUs.

It is all about using the GPU for something other than graphics.

6

Thanks for the replies. That is essentially what I was trying to understand.