Interview: NVIDIA CUDA, what applications for the general public?

1 – The ABC: CUDA a non-technical summary.

Gaming and image software (photos and video) is increasingly compatible with CUDA, a technology introduced by NVIDIA in 2006. Basically it means that games can become even more realistic and processing time in large applications can be reduced.

2 – CUDA in detail: the interview and examples…

Following on from our news on DirectX Compute in NVIDIA graphics drivers, we decided to carry out a short interview with Stephane Quentin, an NVIDIA spokesman. In particular we wanted to ask him what CUDA is and what its use will mean in terms of impact on the general user.

Digital Versus: Hi Stephane, can you describe your role at NVIDIA?

Stéphane Quentin: Hello! I’m part of the European press relations team at NVIDIA.

Digital Versus: Can you describe CUDA technology simply and rapidly?

Stéphane Quentin : CUDA (Compute Unified Device Architecture) is the name of the processing architecture used in our GPUs. This architecture was introduced in November 2006 with the GeForce 8 generation and has been used in all our GPUs (graphics chips) since then. When we launched CUDA, we also launched C for CUDA, the only high level language capable of exploiting this processing architecture at first.

Today, almost 4 years later, there are other languages and APIs than can be used to access this GPU computing architecture: Fortran for CUDA, Java and Python for CUDA, OpenCL and DXCompute. From a technical point of view CUDA uses the GPU execution cores (16 to 240) as well as the very high GPU memory bandwidth to carry out massively parallel processing. NVIDIA GPU’s therefore have two modes, a graphics mode (video games) and a processing mode (CUDA). Apart from the fact that each of these modes uses the execution cores, each also has specific optimisations.

DV: You hear a lot about CUDA in relation to video encoding or for very specialised applications but what abou the general user? Will CUDA affect them directly and how?

SQ: The CUDA revolution has already arrived! Anyone who has a GeForce 8 and upwards with recent drivers is using CUDA. There are something like 100 million GPUs using CUDA already installed. However, applications need to be able to exloit the wonderful power of parallel processing. GPU processing is only advantageous if an application can expoit massively parallel algorithms and not sequential ones (processor type architecture).

In fact, since the introduction of CUDA, over 60,000 developers have looked at the application of GPU Computing and how it can be used to produce performance gains. Over 500 applications have already been developed to work with CUDA architecture. You can find them at the CUDA Zone. The performance gains vary from several percent to factors of over 100, depending on the application.

For the general user, the first apps they come across using massively parallel algorithms are indeed in the domain of photo/video and audio. There are now over 10 general public video applications that use CUDA (Badaboom, Nero Move it, Cyberlink PowerDirector 7, CoreAvc, Arcsoft TotalMedia Theatre and SimHD, MotionDSP vReveal,  Super Loiloscope, TMPGenc 4.0 Express, Cyberlink Expresso). Folding@Home and SETI@Home also use CUDA.

Last but not least, PhysX uses CUDA in gaming. Physx GPU acceleration is now (or will be very shortly) in over 13 games (Unreal Tournament, Ghost Recon, Mirror’s Edge, CryoStasis, Terminator Salvation, Darkest of Days, Warmonger, Crazy Machines 2, Sacred 2 , U-Wars, MK2, Star Tales, Dark Void) and Batman: Arkham Asylum which will take PhysX implementation to new heights with, for the first time, the use of GPUs for the destruction of objects.

Of course the GPU processing revolution will spread with the arrival of operating systems that support GPUs natively as a processing ressource (Snow Leopard and Windows 7). We have just brought out the first drivers that support DX Compute.

DV: CUDA is an NVIDIA proprietary technology. Are you working towards homogenisation of technologies? This would accelerate the development of applications that use this aspect of graphics cards and make the technology available to all computer owners, whatever their hardware.

SQ : As we have said CUDA describes the processing architecture of NVIDIA GPUs and is intrinsically linked to the hardware itself.

The arrival of DX Compuate and OpenCL APIs will certainly homogenise the GPU Computing software offer, as we have seen for graphics APIs. However, as with any standardisation, often the lowest common denominator is used and will doubtless reduce functionality and performance (something already experienced with graphics APIs). The development of new GPU architecture has also made necessary the rapid development of APIs that use it. You can now download the CUDA 2.3 SDK, an SDK that best follows the architecture of the available NVIDIA GPUs. When new architectures are launched, CUDA will still be the best solution for exploiting their power and functionality.

In fact, the acceleration of GPU Computing development is essentially focussed around SDKs in all languages. This is what we have been doing since launching CUDA as part of the training in hundreds of universities and schools around the world. It is key to leave the choice of language, API or programming interface up to the developer so that they can go with what they prefer or are used to. This is why we're increasing the languages and environments that can use CUDA.

Remember that GPU Computing offers unimaginable performance gains in the world of PCs! It revolutionises the use of PCs and their architecture. GPUs are a must when it comes to 3D and they are more than a judicious choice for massively parallel applications. We are on the brink of an era when PCs will use two processors, a sequential processor (the CPU) and a parallel processor (the GPU). Applications will use one or the other (or both) according to their nature.

So there you have it! Now you know a little more about this new use for graphics cards. Thanks to Stephane for replying ot our quesitons

Product survey: graphics cards 2009