Rethinking the Pipeline: DreamWorks Animation Advances the Art

Published:03/14/2012   Last Updated:03/13/2012

By Lee Purcell

Infused with an abundance of multi-core processing power and bolstered by new multi-machine server-farm architectures, DreamWorks Animation has embarked on a radical re-engineering of their animation pipeline to better accommodate artistic aspirations and to improve the workflow across an increasingly complex studio environment.

The classical waterfall model that has historically dominated animation work-where work proceeds stage-by-stage within the strict, linear path of the pipeline-is giving way to a more fluid, interactive model. Central storage of data, distributed processing to better handle dynamic workloads, and tight iterative feedback loops to inform artistic design characterize this swiftly evolving pipeline model. Working in concert with Intel, DreamWorks Animation engineers are re-vamping the architecture of their production systems to the ultimate benefit of artists. While still a work in progress, this new model derives its performance from massive parallelization and uses distributed processing intelligently to take advantage of resources across the infrastructure.

Iterative Design
Improved iterative design techniques are one of the biggest creative advances resulting from increased use of digital-media technology. Iterative design plays an important role in many visual arts, including photography, moviemaking, and animation. The traditional photographer spends years learning the characteristics of film, camera, light meter, and how to read a given scene. Nonetheless, each time the shutter opens to allow a fraction of a second’s worth of photons to enter through the aperture, the end result is not known until the film is developed. Similarly, conventional moviemaking is subject to the same interruption between the capture of the subject and the viewing of the results, which may or may not satisfy the director’s intentions. Traditional cel-based animation prior to computer-generated (CG) digital processes had been tremendously laborious, requiring hundreds of images, drawn by hand-from initial sketches to final cels-to create even a few seconds of animated material.

Digital content creation introduces a new paradigm, an environment in which the lapse between capture and viewing is seconds rather than hours or days. Photographers, moviemakers, and animators benefit from the ability to preview and evaluate scenes more quickly as they are captured or created. If the artistic design is not effective or compelling, the immediate feedback informs the artist, and changes can be made. A change of lens, aperture, or point of view may make all the difference between a mediocre end result and a visually stunning work of art.

Probably no single advance in the visual arts has created such a dramatic change in process as CG animation. Enabled by computer processors that dwarf capabilities of processors from a few years back, being able to dynamically manipulate millions of pixels per second has opened up new creative avenues for animators. Applying this processing power to the tasks within the animation pipeline will enable a restructure of the workflow and a more interactive approach to building characters and scenes. The speed at which new design ideas can be tried and viewed will create an environment that favors experimentation, skillful refinement, innovation, and broad artistic vision.

Research and Development Explores New Architectures

DreamWorks Animation’s upcoming animated film- How To Train Your Dragon-takes advantage of the high-performance computing (HPC) capabilities of next-generation Intel® microarchitecture code-named Nehalem. Using architecture that incorporates both multi-core and multi-machine processing principles, this HPC framework is being constructed around a distributed processing model. Development is ongoing as the DreamWorks Animation engineering team explores the best methods to exploit compute resources concurrently for processor-intensive activities. The fundamental direction of this architectural redesign is abundantly clear: to take tasks (whether a multi-hour final rendering of a scene or a real-time visualization of a single frame’s lighting) and selectively assign them-whether to one processing core or to a mini-server farm.

Within this environment, workstations equipped with Intel® Xeon® processors give artists the performance capabilities to obtain rapid feedback as modeling, effects, physics, lighting, and other tasks can be previewed in near real time. Server farms powered by Intel Xeon processors improve the rendering times for the kinds of increasingly complex animated sequences that characterize DreamWorks Animation’s latest films.

Lincoln Wallen, Head of Research and Development at DreamWorks Animation, is leading the re-engineering program on the pipeline. With the additional compute power available from platforms based on Harpertown and Nehalem architectures, the focus has been on finding ways to enable better creative decisions to be made. “Our objective is to allow the artist to apply substantial computing power to the creation of the final images,” Wallen said. “Our ability to do this is constrained by the need to continually revise the images when the results are not to the artists’ liking. By enabling faster feedback on previews of the images, we can support better creative decision making.”

Informing the artist-through richer, more elaborate feedback loops-guides the creative process at a number of stages in the pipeline. “In the old days,” Wallen said, “and even today, the animators animate in an environment that is not lit. The reason that it is not lit is because they need frame rate in order to animate. They need feedback for the behavior, for their control. But, the lighting also needs to be computed. If the lighting is computed dynamically, that computation alone outmatches the desktop.”

“So, our artists typically turn off the lighting in order to get the basic animation working. But then at some stage, you want to do a full render of that animation to see how it feels when lighting and the other elements, such as hair and clothing, are integrated. At some stage, all of the elements have to come together, and the sooner, the better.”

“We typically break the scene down into components in order to gain interactivity,” Wallen continued. “Interactivity is key to the creative work. But, you have got to put it all back together again in order to see that what you have created actually works in the integrated experience. The more you can pull into the creative environment, and sustain at frame rate, the better the creative decisions that the artist is making at that point in time become.”

As part of the Intel and DreamWorks Animation alliance, Intel engineers have worked closely with the DreamWorks Animation engineering team to maximize the performance and efficiency of the pipeline. This work extends from the tool development deep into the computing infrastructure. Derek Chan is the Head of Digital Operations at DreamWorks Animation and is responsible for the compute infrastructure. “One of the key advantages of Nehalem architecture for us,” Chan said, “is the power-to-compute ratio. We are power-limited and data-center space is a precious commodity, so Nehalem architecture gives us benefits in terms of compute power per watt.”

“The memory system also helps us significantly for certain types of tasks,” he added, “by increasing throughput. We can potentially use fewer of the Nehalem blades for our rendering farms than we currently need for other solutions. Again, density of compute is very important to us.”

In summary, Chan commented, “Nehalem is-by a substantial margin-the fastest architecture we have ever used in the render farm. It performs well on a per-core basis and the wide-open I/O bandwidth lets us keep all cores busy, much more efficiently than previous architectures.”

DreamWorks Animation Film Director Jennifer Yuh Nelson works on
character development using an interactive tablet display.

The Large-Scale
Architecture Challenge

“The re-architecture that we are undertaking,” Wallen said, “is fundamentally about distributing compute and allocating hardware-whether that is in process or in complexity.”

“I do think the theme of single-processor to multi-processor-distributed processing-is one of the biggest challenges that we are grappling with. It is a big step. We are in a better situation in that we have been utilizing an HPC-like environment for many years, so we have gotten used to taking a problem and distributing it across a large compute farm. But, now with multi-core technology, and as we look ahead to the capabilities that Larrabee offers, it is boundaries between things like processors and memory, the fact that you want to run tasks in different modes on the desktop and then on the farm interactively and then overnight in batch mode.”

“This transforms the problem to: how do I want to run the program right now? It needs to be far more flexible than just writing code and executing it. Our means of describing the problem becomes more data driven, more abstract. We have more places where we are transforming that problem into a different form or into a distributed form. Then, we have a greater need for on-demand management of the compute results, whether that is memory on a single platform or machines in a farm or networks in a particular area or whatever. All of this is pretty cutting-edge; it is almost like a massive operating system.”

Looking Ahead to Larrabee
The unique architecture of Larrabee provides some key advantages in terms of character animation for DreamWorks Animation, because of the tight coupling between the processor and the graphics processing unit (GPU). Typical video games, which rely heavily on the performance of the GPU, don’t use a lot of mesh-based surfaces or deformable geometry. This is a very different kind of animation from what is used in animated films.

“Our characters are deeply and extensively deformable,” Wallen said. “When we animate these characters, we are changing their geometry all the time-frame by frame. The relationship between the rendering, shading, and deformation is far more tightly coupled than would be the case with rigid geometry. Our ability to get some aspects of rendering into our animation environment, some lighting and so on, really comes down to our ability to sustain both the graphical compute and the deformational compute at the same time. Putting that image on the screen is not just about computing it graphically. It is about deforming the geometry, as well. That deformation is typically a CPU activity.”

“It could be clothing,” he continued, “there could be simulation involved, could be hair, it could be the skin, or appendages, or whatever. What gets broken into the two parts of the machine (CPU and GPU) in a traditional architecture-physics simulation and graphics-is far more intimately entwined in our pipeline. Traditionally, we have been on a standard CPU. We haven’t exploited GPUs very much-precisely for that reason. For displaying windows and fixed geometry on the screen, of course we use a GPU, but in terms of doing our core character deformation and visibility, we use the CPU.”

“When you start to bring those two characteristics, CPU and GPU, closer together (in particular, allowing algorithms to run cooperatively across those compute resources), then you start to give us the ability to use the hardware to accelerate our core character deformation systems. And, that is one of the promises of having such a large number of cores on a single machine on an animation desktop.”

Parallelism as a Performance Enabler
Clearly, the nature of graphics software programming is being transformed dramatically by architectures, such as Larrabee, that favor concurrent operations and shared data stores.

“The way compute power is scaling,” Wallen said, “brings it far closer to the way we have to work anyway. We have always done multi-machines and used large HPC infrastructure. Now we are seeing the microarchitecture for individual machines going large scale: it is going multi-core, highly multi-threaded.”

“The gap between a farm and an individual machine-conceptually, at a software level-is actually diminishing. Differences remain: in a multi-machine configuration, you’ve got a network environment, and, obviously in a single-machine configuration, you are in the data paths and dealing with caches. And, we are dealing with network-attached storage in the farm. But, architecturally, these things are starting to look very similar.”

“Obviously,” Wallen said, “there is a lot of work in making a single machine load utilize all the cores. But, we recently demonstrated the new animation execution engine that is getting a fantastic speedup from our existing solution on an HP xw8600 workstation.”

Some of this gain, Wallen noted, is because the software has been revamped to use all four processors. But a major part of the improvement is based on another factor. “We are architecting the software to accommodate many more cores per CPU,” Wallen said. “So the bottlenecks in execution are minimized-it has become a far more parallelized system.”

Targeting the Needs of the Artists
The fundamental motivation for the ongoing pipeline re-architecture effort is to better support the artists and filmmakers at DreamWorks Animation.

“We want artists to do art,” Wallen said, “and the CG pipeline has traditionally been intensely technical. In being so technical, it has limited the range of people who can actually express themselves. Part of the goal of what we are doing with tools right now is to make the creation of the art itself far more natural. More artistic interfaces are one goal. Another is providing feedback to the artist that is much more immediate and much more complete-complete in terms of what the artist is seeing.”

Shrek’s Law Meets Moore’s Law
A well-known axiom of the animated film world is that in each successive animated work the total number of rendering hours grows, regardless of the fact that processing power increases substantially each generation. Animated moviemakers quickly take advantage of that extra processing power to build more elaborate character interactions, richer textures, better special effects, more realistic lighting, and other improvements into the frames.

DreamWorks Animation has its own perspective on this phenomenon, as coined by Chief Technology Officer Ed Leonard. According to “Shrek’s Law,” for every Shrek film made the amount of compute power devoted to rendering has doubled. This doubling is necessary to accommodate added realism and scene complexity as audience expectations in entertainment continually rise. With the move to 3D, the doubling also extends to the additional frames required for stereoscopic 3D, which of course requires twice the number of frames as a non-3D movie that is the same length. The computational power of each successive processor advance gets consumed as quickly as it is provided-creating a never-ending dependency that counterbalances Shrek’s Law with Moore’s Law.

Moore’s Law predicts the transistor density of processors doubling approximately every two years. This exponential growth, originally forecast by Intel co-founder Gordon E. Moore in 1965, has proven a reliable guide to processor advances and is expected to continue for at least several more years. Given the nature of Shrek’s Law, it will have to-just to keep pace with the creative ambition at DreamWorks Animation.

The close relationship between Intel and DreamWorks Animation has been strengthened by the ongoing testing of the latest technologies from Intel and the feedback from DreamWorks Animation on what works and what doesn’t. Technology advances-from Nehalem to Larrabee-undergo intensive evaluation under simulated production conditions to hone performance and eliminate problems.

The Next Act
Intricately rendered scenes, expressive characters, and stellar special effects signify the arrival of next-generation 3D CG feature-film animation from DreamWorks Animation-enabled by the latest platform technologies based on Intel® Core™ processor architecture. We’re collectively looking forward to the next act, as CG animation becomes even richer and deeper, enabled by a new generation of hardware and software tools. Given the nature of Shrek’s Law and Moore’s Law, the experiences of movie viewers are just going to keep getting better and better.

Building Avenues for Creativity
Match artistic skills and a flair for storytelling with a production pipeline based on a well-engineered, world-class hardware platform, and you have the core ingredients behind the DreamWorks Animation and Intel alliance. As Intel has embedded the concepts of parallelism more deeply into processor architectures, delivering higher and higher levels of processing power, DreamWorks Animation development teams and Intel application engineers have been working together to develop software and techniques to exploit the additional performance capabilities.

DreamWorks Animation Editor Joyce Arrastia evaluates an animated sequence.


Animation work processes, once prolonged and tedious, have become more interactive with added opportunities for creativity. David Burgess, Head of Character Animation on Monsters vs. Aliens, said, “One of the nice things about the CG paradigm is that you are already working with a model in real time. All you have to do is set a pose, and you can let it play with the audio and you can watch that. You can sculpt your performance out of watching it play. The whole CG process is very interesting.”

“In the computer-generated realm,” he continued, “the more powerful the processor, the more information you can get into your workspace and into your different windows that you are working with. So, if you have a really, really powerful computer, you can turn on your background, you can turn on all of the characters in the shot, you can bump the rez up so you can actually see facial expressions and stuff like that. Whereas, if you don’t, if you have something that can’t handle all that data, you end up stripping all that away.”

With a powerful set of tools and technologies at their disposal, DreamWorks Animation artists, animators, and technical staff members are resolutely finding new ways to innovate, communicate, and entertain. Intel supports these efforts, helping enable artistic achievement through technological expertise.
For more perspective on the work of David Burgess, view the video.

Powering the Animation Pipeline
The box-office success of Monsters vs. Aliens (MvA) and critical acclaim that followed mark a milestone in the film industry. No longer pitched as a gimmick or novelty, 3D has become an integral element of creative filmmaking, important for heightening the overall audience experience and providing a fresh set of tools for storytelling. New platforms from Intel enhanced the animation pipeline for the production of MvA, as revealed through discussions with DreamWorks Animation technical staff members.

Not comfortable to simply rest on its laurels, DreamWorks Animation is fully engaged in its next work, How to Train Your Dragon, which will capitalize on the hardware advances and performance capabilities of Intel® Core™ processor power. The abundant processor power is being transformed into new methods for pre-visualizing scenes, constructing intricate character interactions, modeling complex environments, and delivering richer special effects. The thoughts and aspirations of the innovators at DreamWorks Animation are the driving factor: the collective work of a creative team determined to topple conventional expectations with each successive animated work.

To gain a look behind the scenes at the Intel® technology used in the creation of MvA, view the video titled Tech of MvA .

“Nehalem is-by a substantial margin-the fastest architecture we have ever used in the render farm. It performs well on a per-core basis and the wide-open I/O bandwidth lets us keep all cores busy, much more efficiently than previous architectures.”
Derek Chan, Head of Digital Operations, DreamWorks Animation

“One of the nice things about the CG paradigm is that you are already working with a model in real time. All you have to do is set a pose, and you can let it play with the audio and you can watch that.”
David Burgess, Head of Character Animation,
Monsters vs. Aliens

“One of our goals is to let artists do art and hide from the artist the technical complexity of compute resource and how that is applied. They are not programming. They do not have to engage with that level of detail. But, it requires enormous compute power just to separate them from the complexities of processing.”
Lincoln Wallen, Head of Research and Development, DreamWorks Animation

“We want artists to do their art and the CG pipeline has traditionally been intensely technical. In being so technical, it has limited the range of people who can actually express themselves. Part of the goal of what we are doing with tools right now is to make the creation of the art itself far more natural. More artistic interfaces are one goal. Another is providing feedback to the artist that is much more immediate and much more complete-complete in
terms of what the artist is seeing.”
Lincoln Wallen, Head of Research and Development, DreamWorks Animation


Capture the buzz. Subscribe to Intel® Software Dispatch for Visual Adrenaline. (Did we mention it's fun, informative, visually stimulating, free, and you can unsubscribe at any time?)

Additional Resources

Optimizing Without Breaking a Sweat

Product and Performance Information


Intel's compilers may or may not optimize to the same degree for non-Intel microprocessors for optimizations that are not unique to Intel microprocessors. These optimizations include SSE2, SSE3, and SSSE3 instruction sets and other optimizations. Intel does not guarantee the availability, functionality, or effectiveness of any optimization on microprocessors not manufactured by Intel. Microprocessor-dependent optimizations in this product are intended for use with Intel microprocessors. Certain optimizations not specific to Intel microarchitecture are reserved for Intel microprocessors. Please refer to the applicable product User and Reference Guides for more information regarding the specific instruction sets covered by this notice.

Notice revision #20110804