Nestled in the heart of the Presidio, it’s fitting that Industrial Light & Magic’s offices are in one of the most picturesque locations you will find in San Francisco. The Academy Award–winning visual effects company has been delighting audiences since 1975 with its motion-picture computer-generated “mojo”—so it is appropriate that the ILM campus is surrounded by stunning vistas, which include office views of the Golden Gate Bridge, The Palace of Fine Arts, and the Presidio’s lush gardens.

Even more appropriate, perhaps, is the statue of Yoda that holds silent vigil at the main entrance to ILM's offices in the 23-acre Letterman Digital Arts Center. When it comes to using cutting-edge technology in innovative ways to produce rich special effects that continue to evolve and push the medium in terms of realism, ILM is indeed a Jedi master when it comes to its craft.

To be fair, ILM—founded by George Lucas and owned by Lucasfilm—has a bit of an advantage over its competitors. Deeper-than-average pockets mean better-than-average access to bleeding-edge technology—the best “toys” of the trade, if you will. And this access translates into the ability for the company to continue to attract some of the brightest minds.

And you can't argue with the results: Since its inception, ILM has worked on more than 300 films, including a laundry list of some of the biggest visual effects-enabled blockbusters—Star Wars, Indiana Jones, Jurassic Park (and pretty much the rest of Spielberg’s resume of films), Ghostbusters, Pirates of the Caribbean, and, most recently, Harry Potter and the Half-Blood Prince and Transformers 2: Revenge of the Fallen.

ILM’s biggest “toy” by far is its self-titled “Death Star” computing cluster. Originally built by SGI in 2004 and since converted to a Linux system using AMD processors, the Death Star now includes some 5,000-plus multi-core CPUs. A massive CPU “render farm,” the Death Star has been tapped to render the bulk of the CG effects you’ve enjoyed on the big screen during the last six years. At night, the Death Star becomes even more powerful, reaching out to all the unused nodes on the network to boost its core CPU “cluster power” to more than 7,000.

That’s a whole lot of crunching going on, folks—but it’s still not enough, so ILM has starting dipping its toe into the “GPU pool” as a means of boosting its visual effects Force powers to the next level.

For example, many of the fire effects you witnessed in The Half-Blood Prince—including the eye-popping climatic end scene you see pictured in the opening to this story—were rendered with NVIDIA’s Quadro series GPUs running a custom app developed in-house by Digital Artist Christopher Horvath on what was pretty much a whim.

Prior to the last year and a half or so, ILM used the GPU for displaying visuals and Open GL–powered 3D previews, not as a computing device, says Horvath.  A broader acceptance of standards relating to the Open GL language, coupled with the appeal of the GPU’s massive floating-point processing gains for rendering work, now has ILM looking hard at the GPU for all projects.

NVIDIA’s Quadro 4600 and 5600 series cards were used for its work on Harry Potter and 4800 and 5800 series cards for Transformers 2. Due to incomplete support for Open GL standards, Horvath says the Quadros were “still too slow” for the work done on Harry Potter, but they saw a significant acceleration gain when work began on Transformers 2 in November 2007.

At this time, there were still some significant growing pains when it came to making the shift to use the GPU in tandem with the CPU rendering farms—cards were overheating, for example, and Horvath did a lot of his early experimental work on an advanced laptop that was able to run applications the workstations still could not.

“I started playing around, basically, and stumbled by accident into [creating] a little test app that allowed us to draw with the mouse cursor, and wherever the cursor went it would burn,” says Horvath. “And the effect could be stacked to look 3D. Once we had that working it was just a matter of seven or eight months of crossing the T’s and dotting the I’s to get it working right,” he adds with a chuckle.

“The actual GPU part of it was working inside a month and everything else—getting the cameras in, getting the motion blur in, adding deep shadow, and all the rest of the little nuts and bolts, which is 90% of the work, took the rest of the time. We went from having a working little toy—Fire Toy (as it is called in-house) in early February (2008) to our first final in September,” he adds.

Chris’s Fire Toy proof-of-concept application evolved into a much more polished production tool called Verté, which has made it possible to render realistic flame effects in 3D at the desktop level in what is essentially real time. This effectively sidesteps the need for artists to wait overnight (or longer) for the Death Star to render their work,  view it on-screen to tweak, and then rinse and repeat the process. (Verté was also harnessed in a more limited way on Harry Potter to add more detail to water surfaces and for some additional simulation work.)

“The overnight time required to render something equivalent to fire is about 30 hours a frame,” says Horvath. “Spread out to our CPU render farms that’s about a day and a half to see the results [for a typical scene], as opposed to our fire render with Potter, which took about nine seconds a frame” using the GPU-based Verté app.

Not surprisingly, these kinds of results—with a software tool developed on a lark with minimal resources—has ILM taking the GPU much more seriously as a rendering asset going forward.

“Traditionally, ILM has not been a GPU house per se, as it’s fairly new technology to begin with,” says Greg Grusby, Technical Publicist at ILM. “With our Death Star rendering farm downstairs we typically have, just by brute force, the rendering power to get whatever we need done, done.”

“However, now with the GPU, we don’t necessarily have to throw the big iron at every problem…we have lots of different types of problems and maybe now we can solve them in different ways that are a much more efficient use of that processing power,” he adds.

And Harry Potter is a prime example of this principle in action: ILM developed a GPU-based tool that renders fire, and a little bit of smoke, which was then also used on Transformers 2, for the same effects. Since then, the Verté tool has been enhanced significantly and ILM is now “using it like crazy” says Chris—even for “little things” like bits of fire on torches and camp fires in the distance in current projects.

“The fire tool works great because all it does is fire, it’s this very, very single-purpose thing that does its job really, really well,” says Chris.

ILM continues to evolve Verté, and develop other tools that perform similar functions—such as model clouds, for example. Plus, in addition to Verté, Horvath designed a GPU-based rigid body system that was used on Transformers 2 for the massively destructive scene in which Devastator smashes into the top of a great pyramid. Using the GPU, ILM was able to render 120,000 rigid bodies (or fragments in this case) on a single machine in three hours. In comparison, it took a week to render 14,000 rigid bodies (albeit at a more sophisticated level) with the CPU, using multiple machines, for the final scene in Indiana Jones and the Kingdom of the Crystal Skull.

Horvath and Grusby believe the day will come when the GPU can do just about everything the CPU is used for now.

“What’s awesome about the GPU is you have these massive gains in speed—when else are we going to see, in our lifetime, a 2x order of magnitude in speed in six months?” says Horvath. “It blows Moore’s Law out of the water!”

This massive gain in raw GPU horsepower has already allowed ILM to render an effect that used to take 30 hours using the Death Star CPU render farm in less than 10 seconds at the workstation level—all at an equivalent level of image complexity and artist difficulty.

“That kind of speed increase is just a whole new world in terms of us being able to iterate [a scene] over and over again and really finesse it as opposed to just having one or two shots at it, which is exciting,” says Horvath.

He envisions that ILM will continue to develop multiple small tools that use the GPU to solve a particular rendering problem, and that all of these tools will communicate with each other and feed into the production pipeline, rather than all being part of one massive “uber-tool.” The end result for the movie audience is that future films should include computer-generated effects that are even more realistic.

Judging from what ILM has harnessed the GPU for so far in Half-Blood Prince—to surround Dumbledore with a 100-foot wall of flames—and Tranformers 2 (to simulate a massive aircraft carrier getting destroyed by a meteorite, which is shown above) the future of visual effects in the movies is an eye-watering proposition indeed.

Already, with a relatively simple fire tool, ILM is able to create CG scenes at a believable level of photo-realism using GPU rendering techniques that would have been too expensive or too dangerous to film with traditional methods—in far less time.

“And [we’re] just starting to scratch the surface of what you can do when you start thinking in highly, highly parallel ways,” says Horvath.

Presently, ILM does not have a dedicated GPU-based render farm, though Grusby says you can expect that to change soon based on the successes the company has had so far with its initial GPU-based rendering efforts.

Shall we start the name contest now for this GPU farm—Yavin IV perhaps?