Some selected Q&As from a recent chat with NVIDIA hosted by EVGA:
Q : Do you think that (ugh)ati might become cuda compatible in the future?
A : ATI's compute architecture is not as versatile as the NVIDIA's GPU computing architecture and cannot handle C code and so it will not be able to keep up with CUDA.
Q : When will the other interfaces like OpenCL be available?
A : Open CL was developed on GeForce. GeForce GPUs take advantage of several parallel processing APIs like CUDA, PhysX, Direct3D, etc.
Q : What are the advantages/disadvantages between using CUDA vs. OpenCL vs. DirectX Compute?
A : CUDA is NVIDIA's GPU computing architecture and it runs C with CUDA extensions, OpenCL and DX Compute. The differences between these are stylistic. Today the only one available to develop applications with is C with CUDA extensions. OpenCL and DirectX_Compute are NOT competing against CUDA… they are different ways of accessing CUDA (in addition to C, Fortran, etc.).
Q : Will C++ be available in the future for CUDA?
A : Yes in the works.
Q : When will the other interfaces like OpenCL be available?
A : We already have an OpenCL driver for signed up developers - it's undergoing qualification as we speak.
Q : Which part of the cards architecture is more critical to overall CUDA performance (memory speed, shader speed, core speed or all three)?
A : Good question! The number of GPU cores comes first (by far), then clock speed, then memory speed/bandwidth. For example, the Badaboom and vReveal performance scale almost directly proportionally with the number of GPU cores.
Q : Under Vista/Windows 7 is nVidia working with Microsoft to remove the "must extend the desktop" to get CUDA functioning on multiple GPUs?
A : We already have a solution for VISTA - check out the SLI/PhysX control panel. Things are slightly different for Win7, but we'll have a solution for you.
Q : Is OpenCL going to replace CUDA, or co-exist with CUDA?
A : They will co-exist. GeForce GPUs accelerate PhysX, Direct3D and Open CL applications. We were the first to market with an OpenCL driver earlier this month.
Q : What percentage of the GPU does CUDA use for its application?
A : There are some applications which use 100% of the GPU for CUDA. The tesla cards for example do not have graphics output and are 100% CUDA.
Q : How can ATI cards run OpenCL if they cannot handle C code?
A : In their latest architecture ATI has put in support for those aspects of C that are supported in OpenCL. There is no release of ATI of openCL implementations to date.
Q : Can older cards become CUDA enabled?
A : CUDA is an architecture INSIDE the GPU - with the 8 series and up - you cannot enable it with software if it's not there in the hardware already.
Q : Does CUDA perform differently on professional GPUs (Quadro) when compared to a similarly spec'd GeForce card?
A : Quadro cards have been tested to much higher levels for professional levels of reliable operation and have large framebuffers for CUDA applications.
Q : Are there any current plans to be able to run a virtual machine accelerated(emulating an x86/x86-64) by cuda?
A : No plans I'm aware off.
Q : Will there ever be adjustment in the control panel for CUDA?
A : We may have options to select which GPU to select for CUDA and stuff like that.
Q : How large is the team of developers that worked on CUDA?
A : All the major teams in NVIDIA are working on CUDA including the Architecture team, the software team, the compiler team, the chip team, the application team... It's hard to estimate since it is an inseparable part of the GPU but I would say over 2000 people are working on developing CUDA technology.
Q : Is there any possibility in the future of CUDA be used with XNA?
A : Yes we are planning on adding support for other high level languages including FORTRAN.
Q : Is Cuda a programing platform? Or an architecture hardwired into the GPU?
A : CUDA is a design architecture of GeForce GPUs (8 series and higher); hardwired into the GPU. PhysX is an example of an API that takes advantage of CUDA. Developers can design apps to take advantage of CUDA GPUs using C - the most prevalent language in the world.
Q : Does overclocking a GPU cause instability with CUDA other than the norm that overclocking gives?
A : OCing does not affect CUDA performance....unless of course you fry your GPU to death and it no longer functions
Q : Windows 7 is supposed to sport a feature that allows any GPU, even Intel GMA for parallel processing. How does this compare to CUDA and if CUDA is better, is it because it is built into the GPU architecture?
A : Windows 7 has Direct X compute which will make GPU computing part of the OS. DirectX Compute is supported through NVIDIA's CUDA architecture. Windows 7 will get GPU computing support through the CUDA architecture.
Full chat log in
http://www.evga.com/articles/00486/ .
