Of Parallel Processing and CUDA

While we could make this a huge section, we will only go over the very tip of the iceberg that is CUDA and how it relates to the G200 architecture. It should be mentioned up front that due to the unified architecture of both the G80 and the G90 cores, CUDA applications will work on them as well, just not with the speed they will on the G200 series cards. Unfortunately, due to the time constraints of this review and the very beta nature of some of the applications, we were not able to benchmark the parallel processing features of the GTX 280. In the weeks to come, expect a full article dealing with CUDA, the applications which benefit from it and benchmarks. Until then, here is a little primer.


What is CUDA?

Nvidia has this to say about their CUDA architecture:

CUDA is a software and GPU architecture that makes it possible to use the many processor cores (and eventually thousands of cores) in a GPU to perform general-purpose mathematical calculations. CUDA is accessible to all programmers through an extension to the C and C++ programming languages for parallel computing.

To put that into layman’s terms it means that we will now be able to take advantage of the massive potential offered by current GPU architectures in order to speed up certain tasks. In essence, CUDA should be able to take a task like video transcoding which takes hours on a quad core CPU and perform that same operation in a matter of minutes on a GPU. Not all applications can be transferred to the GPU but those that do will supposedly see an amazing jump in performance.

We could go on and on about CUDA but before we go into some of the applications it can be used in, we invite you to visit Nvidia’s CUDA site: CUDA Zone - resource for C developers of applications that solve computing problems


Folding @ Home


By now, many of you know what Stanford University’s Folding @ Home is since it is the most widely used distributed computing program around right now. While in the past it was only ATI graphics cards that were able to fold, Nvidia has taken up the flag as well and will be using the CUDA architecture to make this application available to their customers. From the information we have from Nvidia, a single GTX 280 graphics card could potentially take the place of an entire folding farm of CPUs in terms of folding capabilities.


Video Transcoding


In today’s high tech world mobile devices have given users the capability to bring their movie collections with them on the go. To this end, consumers need to have a quick and efficient way of transferring their movies from one device to another. From my experience, this can be a pain in the butt since it seems like every device from a Cowon D2 to an iPod needs a different resolution, bitrate and compression to look the best possible. Even a quad core processor can take hours to transcode a movie and that just isn’t an option for many of us who are on the go.

To streamline this process for us, Nvidia has teamed up with Elemental Technologies to offer a video transcoding solution which harnesses the power available from the GTX’s 240 processors. The BadaBOOM Media Converter they will be releasing can take a transcoding process which took up to six hours on a quad core CPU and streamline it into a sub-40 minute timeframe. This also frees up your CPU to work on other tasks.

If these promises are kept, this may be one of the most-used CUDA applications even though it will need to be purchased (pricing is not determined at this point).


PhysX Technology


About two years ago there were many industry insiders who predicted that physics implementation would be the next Big Thing when it came to new games. With the release of their PhysX PPU, Ageia brought to the market a stand-alone physics processor which had the potential to redefine gaming. However, the idea of buying a $200 physics card never appealed to many people and the unit never became very popular with either consumers or game developers. Fast forward to the present time and Nvidia now has control over Ageia’s PhysX technology and will be putting it to good use in their all their cards featuring a unified architecture. This means that PhysX suddenly has an installed base numbering in the tens of millions instead of the tiny portion who bought the original PPU. Usually, a larger number of potential customers means that developers will use a technology more often which will lead to more titles being developed for PhysX.

Since physics calculations are inherently parallel, the thread dispatcher in the unified shader architecture is able to shunt these calculations to the appropriate texture processing cluster. This means a fine balancing act must be done since in theory running physics calculations can degrease rendering performance of the GPU. However, it seems like Nvidia is working long and hard to get things balanced out properly so turning up in game physics will have a minimal affect on overall graphics performance.


Our Initial Thoughts Regarding CUDA

We have quickly run through three of the emerging uses for the CUDA technology on the parallel processing architecture of modern Nvidia GPUs and we must say it looks promising. Even though the technology may be new by a consumer’s standpoint, the potential of CUDA is virtually limitless. What needs to be done is to clearly define support for this architecture through the use of both pay-for-use and FREE applications. Without the use of home-brew programs for CUDA to speed up everyday tasks, we can see this technology flying us all by without any widespread adoption.

At this point, all we have are Nvidia’s claims regarding performance and ease of use with CUDA and today’s tasks but only time will tell if these claims can become a reality. What is needed is long-term support from Nvidia for this architecture and from what we have seen; the boys in green are ready.