• Blogs
• Forums
• Tradeshows
• Partners
• White Papers
• Events

 

Nvidia Quadro Family

Mar 1, 2006 12:00 PM, By Joshua Kolden, Crack Creative

Realtime, High-Quality Previz

Crack Creative provides visualization and production services to the feature film, commercial, and architectural markets. It's our previsualization services, however, that have changed our production methods most dramatically over the past few years. For example, in a typical workflow, one could expect perhaps four previz shots per artist per day.

But Crack has created a completely different paradigm for visualization with its interactive, virtual environment for directors and designers to work within. The immediacy of this process yields about 150 previz takes a day, with roughly half being printed. In a typical week, we might generate anywhere from 10 to 30 minutes of raw material. To compound the challenge, we usually deliver at HD resolutions. This throughput of CG material is staggering, requiring us to investigate new pipelines for our production process.

Part of the solution: We rely heavily on high-performance graphics cards to provide an immersive environment that offers enough quality and detail to make most of the key creative decisions for a shot. Lighting, shading, and movement must be top quality so the directors and designers can make informed creative decisions.

We selected Nvidia Quadro cards for their high baseline performance, solid software support, and scalability using SLI technology. We built our own “whitebox” systems, fitting them with dual Xeon 3.2GHz CPUs installed with dual Nvidia Quadro FX 4400 cards running in SLI mode. Rather than building a traditional, dedicated render farm, we currently use the artist's workstations to render the images overnight, leveraging the use of these same systems.The realtime quality that these cards provide is a blessing and a curse. After our production sessions with the client, we might end up with a list of about 75 shots, which average out to six minutes of HD CG — that's around 8,632 HD frames to be rendered per day.

To deal with this amount of material, we created a “Virtual Lab.” It's like the overnight telecine that's done for a film shoot, but in this case we needed to use hardware and software to render out many shots overnight so that we would be ready for the next day. While there are many software renderers that fit our basic image quality requirements, Nvidia Gelato is the only one we've found that also provides a useful solution to the massive production pipeline concerns.

Those who have heard of Nvidia Gelato may know that it's a GPU-accelerated software renderer that runs on Nvidia Quadro cards. We have a lot of GPUs to throw at tasks, so that was reason enough for us to consider Gelato. We've also found Gelato's other features are equally, if not more, important.

Software rendering doesn't just concern the time required to render out a frame. With typical production tools, there are a number of steps required before a scene can be sent to the queue for rendering. In testing our production setup, we found that this time was prohibitively expensive.

In normal postproduction, the amount of time it takes to load a scene in your package of choice is trivial compared to the amount of time you spend working on that scene. For example, a typical artist may only load one to five shots a day to his workstation. If it takes five or 10 minutes each time, it's not that big of an impact on their productivity.

In our Virtual Lab however, we found that loading scenes and generating render scripts (such as RIB files) consumed the majority of the time lab techs were spending. We have so many scenes to open that we found we had more then 10 hours of scene load time alone in a given week.

This is where Gelato has made all the difference. While Gelato comes with a Python binding as an example interface, its real power is in its interface abstraction. At its most basic, this means Gelato can take input from RenderMan or Mental Ray files, or any other render file types. All you must do is define the binding. You can also adapt Gelato to read Alias Maya files, FBX, OBJ, etc. This is just the feature we needed to avoid the massive amount of render prep that most solutions require.

Crack uses Ruby as the standard scripting interface language for all of its tools. We selected it because of its flexibility; it allows us to generate web pages, interact with databases, create shell tools, provide GUI tools, and manage render queues — all with the same scripting libraries. Additionally, it is trivial to imbed Ruby into other applications such as Maya. This allows us to write Ruby scripts that can control applications.

As part of our Virtual Lab, we have written a Ruby binding to Gelato that allows us to procedurally interact with the renderer. Just like the Python binding that comes with Gelato, it's possible to have small files that produce incredibly detailed scenes. Need a thousand trees to populate a forest? No problem, just stick a few lines of code in a loop. RIB and most other scene description formats don't have these features.

With Gelato integrated into our Ruby pipeline, it is possible to make almost every Virtual Lab task completely automated. We take our scene files, interpret them directly to Gelato, render on the queue, generate Avid files, and update the web pages for the shot with a small set of easily maintained scripts. The entire process can be managed from our internal web pages. On top of this, we get the massive performance gain of Gelato's GPU assisted rendering, leveraging every ounce of our hardware investment.

The facility throughput and overall performance gain give render boxes with Nvidia Quadro graphics cards installed a huge price/performance advantage. For example, after we've finished our production day, we put our graphics hardware to work overnight, something that all other solutions leave idle. It also means that when we staff up for particular jobs we are able to give artists “render” boxes for their workstations, allowing us to be much lighter on our feet to address the specific demands of a project.

Another little-recognized feature of Gelato that we find extraordinarily valuable is its low-overhead resolution independence. In most renderers, you can expect a proportional increase in time as image resolution increases. Ray tracers are particularly prone to this; REYES is better, but it is still tightly coupled to resolution. Gelato, however, is affected very little by resolution changes. When we generate HD-resolution images, this feature makes it possible to do HD output at very little additional time over SD renders.

Taking a look at production times again, the best software render solutions average about 10 to 20 minutes per HD frame, which includes loading times. Gelato, meanwhile, is not only fast, but allows us to cut out most of the prep overhead. The end result: An impressive 30 seconds to 120 seconds per HD frame with all the same quality and features. Even if we had no other use for the graphics hardware at all, we are achieving the throughput we need at a fraction of the cost of a typical render farm.

Other features such as render quality, geometry support, caustics, ray tracing, and ambient occlusion are, of course, very important to us as well. Despite the productivity wins, we would not be able to use Gelato if it didn't provide the latest rendering tools.

At Crack Creative, we're focused on shifting the paradigm of CG production. While CG is still mostly done offline, we're moving it to an online, interactive process that delivers final-quality work at an exponentially faster rate. Nvidia Gelato, powered by Quadro graphics cards, is the only renderer that fits this new paradigm, delivering both the quality as well as a new tier of performance.

---

Joshua Kolden is a virtual production supervisor at Crack Creative. The Los Angeles-based visualization company creates effects work for clients that include James Cameron, Roland Emmerich, Warner Bros., Fox, and Chiat/Day. For more information, please visit www.crackcreative.com.

© 2008 Penton Media, Inc.