High-end video cards that assist compatible 3D applications to render a scene faster are the norm in professional environments, from CAD to animation for movies and games. These cards are also often marketed to video editors as well, with vague claims that they are faster than normal cards. As followers of this site know, one of my favorite rants is to point out that when you actually look at the specifics, $2,000 "professional" video cards have almost no impact on performance over $200 gamer cards in every day editing situations and only help on specific encoding and rendering tasks. In other words, if you are converting your library of MPEG-2 DVD-Video discs to MPEG-4 MP4 files, then sure, you can cut your render times in half. For $2,000. If you are capturing, editing, adding titles, color correcting and mixing audio most of the day, then you will see zero improvement for your $2,000 investment.

Modern GPUs (graphics processing unit) have more transistors in them these days than your CPU (central processing unit) and it certainly could be argued that they are just as complex, if not more complex, than a CPU. The latest graphics cards typically have a billion transistors (or more), while the latest CPUs might "only" have 700 million (for an Intel Core i7). Granted, the GPU's raw processing speed is a lot slower, but still, it seems a waste to have that GPU sitting idly by, redrawing the screen occasionally as you drag clips around on the timeline. The problem is that it is hard to farm individual tasks off to the GPU and keep some on the CPU at the same time. It's an either-or situation.

Since CPUs and GPUs are different, especially in terms of parallel processing, the two brains in your computer are better at different tasks. Even with eight threads, the raw power of your CPU (3.0GHz) is generally better at crunching serial instructions really fast, one after the other. Modern GPUs (800MHz), on the other hand, are better at parallel processing, with some ATI cards sporting "800 stream processors" (see the illustration) and I/O to extremely fast on-card RAM (DDR5). Does that mean 800 simultaneous threads and 20x rendering speeds? Probably not, especially considering the divvying up of the render tasks by the dispatch processor and the gluing back together tasks by the reassembly processor (which I just made up), but I won't quibble with the more general "massive parallel processing" assertion or independent benchmarks. Rendering and encoding happen to be processes that can be efficiently attacked in parallel. Thus the current State of the Art is to have your CPU handle all of your editing tasks and let the GPU take over rendering at the end of the day. Better than nothing, I suppose, but pretty primitive and even then, rendering is not an entirely parallel process, so there's some crosstalk.

Nvidia makes GPUs (with CUDA hardware-assist technology) and Intel makes CPUs. The two companies cooperate as much as two separate companies can, I guess. Nvidia's neck-n-neck competitor is ATI, while Intel's main competition is AMD (which, honestly, lags behind right now). ATI and AMD are now the same company, however, so assuming the corporate cultural merging is going smoothly, one might expect more cooperation between the CPU and the GPU, especially if "one" were a PR flack at AMD/ATI.

It is with that observation (and a healthy grain of skeptical salt) that we can see that AMD/ATI has an interesting point to make when it suggests that GPU-assisted computation could be a lot more efficient if serial tasks were handled by the raw speed of the latest CPU (from Intel or AMD, say 8 threads running at 3.0GHz) and parallel processing tasks were shunted to the GPU (hypothetically 800 threads running at 800MHz), by whatever miracle drivers/software the company that makes both could come up with. In principal, the new Stream processing "both CPU and GPU" (as touted by the company's PR) shows some significant performance advantages (275% faster) over its rival's "either CPU or GPU" strategy, at least in terms of rendering in supported applications, which means Cyberlink Espresso today. Again, neat, but not revolutionary, since it is still only about renders. If you spend most of your day rendering, horray! If you spend most of your day editing, compositing and mixing... meh.

Still, there was one aspect of this announcement that did make me cheer: the new Stream drivers, which are available already as a part of the standard Catalyst package, work on all of the newish AMD/ATI cards all the way back/down to the ATI Radeon HD 4350, which can be had for $25 after rebate on newegg.com at the extreme low end. $25. That's a far cry from the latest "professional" cards that start at about $800 and run all the way up to $2,500 (and higher). So yea, keep that salt handy while you are reading AMD/ATI's press release and yea, keep in mind that we are only talking about select applications and only rendering, but it might not be long before a $100 for a new video card might actually be a reasonable upgrade for video editors.  

Now if we could just get Adobe, Apple or Sony to write a background rendering engine that hit the GPU while we were plugging away on our timeline...

References:
- AMD's Video Acceleration Press Kit - with video, benchmarks and other fun stuff
- ATI ups the GPGPU ante, the INQ
- Debunking Hardware-Assisted Rendering