"Even if I set out to make a film about a fillet of sole, it would be about me."
- Federico Fellini -
        Syndicate

* Atom 1.0
* RSS 2.0

Videopia the Book for only $9.99! 236 pages, all original content on EVERY topic from light to copyright.

"In my opinion, the content is worth much more than you are charging for it..."
                            - David B.

"This booklet is a great resource, and worth many times the price. Heck, it's worth four bucks just for the entertaining read alone..."
                            - Shane G.

"This is the best explanation of
Internet video I have ever read."
                            - Mark H.

Debunking Hardware Assisted Rendering PDF Print E-mail
Written by D. Eric Franks   
Monday, 29 September 2008 06:19

There's been a long history of "professional" video cards, from ATI to Nvidia to Canopus, Matrox and Pinnacle. The cards are supposed to boost 3D and video rendering performance, although I've never seen benchmarks or comparisons proving anything. Of course, they also cost at least twice as much as a standard video card. Recently, Nvidia announced the latest technology upgrade for hardware-assisted renders, dubbed "CUDA," that will - theoretically - improve the performance in the Adobe CS4 suite and Cyberlink video editing. We'll see. Or, more likely, maybe we won't.

nvidia_quadro.jpg

 

Limited Improvements

It's not that I think anyone is lying here or being deceptive, it's just that there are so many caveats and asterix that it is almost impossible to demonstrate real-world performance gains. Early hardware assisted rendering cards were almost exclusively for 3D rendering, which is not surprising, considering video cards for a decade now have been primarily designed and manufactured for gaming. So if a software developer could tap into a gaming card's 3D hardware, more power to them. The problem is that the software needs to support the card and not the other way around. If you work in Nvidia marketing, this can make selling hardware rendering challenging, since you want to say "Improved render times!" and not "Improved render times in applications that support the Open GL 2.0 spec and then only when performing certain specific types of renders, but not others."

NOTE: I am not saying Hardware Assisted Rendering is completely worthless. I'm just saying its benefits are overstated by marketing and I've never seen any numbers objectively demonstrating increased performance.These cards probably do improve performance for a very limited set of professionals (e.g., AutoCAD) under very specific circumstances and are probably worth their weight in gold in those situations. If you are unsure if these cards will help you or not, they probably won't.

The situation is even more narrowly definied with Canopus, Matrox and Pinnalce, where a hardware card was tied to not just a specific piece of software, but typically limited to specific video formats and special effects. As long as you were working within these limitations, the hardware works great and, as many people can attest, when the card matches your workflow, the real-time performance is amazing.

Limited Upgrades

There is another downside to tying your software to your hardware: you can't upgrade. The special, expensive cards are tied to your computer's hardware and your version of software, so you can't upgrade to a faster machine when the time comes. And that time seems to be about three years. So while a new hardware assisted render card will give you improved performance - at a high price - for a year, the latest plain old vanilla computers without fancy graphics cards will start outperforming your system in short order. And they'll do it without any of the limitations of a hardware card tied to a specific software release.

 Again, before you become too angry with me, let me restate: Hardware assisted rendering does work provided you accept its limitations. What you need to ask yourself is: Is it worth $1,000 to get improved performance in some situations or should I save that money and buy a rockin' fast gaming card for $250 today and buy another even faster one in a year and get the latest software updates too?

This also raises another controversial issue: What, exactly, is different about the "professional" cards from Nvidia and ATI? My informed hunch is that there is no difference in the hardware (re-tooling a manufactuing line is not likely) and cheaper cards are intentionally firmware-crippled.

Nvidia CUDA

Nvidia's new Cuda technology promises to be both different and the same. First, it's more than just 3D and video, although it is certain that the first developers to take advantage of GPU assisted rendering will be companies like Adobe and Cyberlink. This makes sense, since video rendering is amazingly computationally intensive and offloading video rendering to an otherwise idle GPU is a great idea. Considering modern GPUs have transistor counts that rival or even beat CPU cores, the potential gains seem obvious. Personally, I'd like to see the numbers in both artificial benchmarks and real-world render speed tests.

And I'd like to know what is different about "professional" cards, again, show me the numbers. Will a one-year old $1,000 "professional" card outperform the latest $250 gamer's card? Will it outperform the latest CPU with a crappy $50 video card?

Why CUDA? Why tie your software to specific hardware when you could support an industry standard like Open GL or even Direct X? These questions demand answers!

Personally, I'm excited about the potential here and initial demos of CUDA-accelerated Photoshop have been pretty cool, but I'm also skeptical. The long history of hardware assisted rendering has been marked by unnecesarily expensive graphics cards and limited scalability. I hope I'm entirely wrong this time!

References:

- Nvidia's David Kirk on CUDA, CPUs and GPUs
- Nvidia's CUDA Zone

< Prev   Next >