Photorealistic Rendering Grows in Importance

Consider how much cheaper, easier and faster it is to make a picture of a car (or a shampoo bottle) than to build a physical prototype. It’s easy to see why the idea of replacing prototypes with renderings would be attractive to automobile (and bottle) manufacturers. But, if a rendering is going to replace a real-world prototype, it has to resemble the real world as closely as possible. It must be reliably predictive.

We’ve always accepted that premise when it comes to, say, weight or material strength. No one wants a CAD system that gets basic measurements wrong. Photorealistic renderings—pretty pictures—seemed less important. Those “pretty pictures,” however, are becoming ever more useful in the design process. Manufacturers want to predict, not only the gross physical attributes of the final product, but also its look and feel. To do this, they need renderings that not only look nice, but also accurately reflect physical reality.

Flipping to Reality

Physics-based techniques to accurately calculate the movement of light in, on and through various materials have been around for years—decades even—but until now, hardly anyone had sufficient hardware horsepower to do much with them. Consequently, rendering has traditionally involved some degree of … well, “cheating.”

CPU or GPU Rendering?

There are two rendering claims that come up again and again in the industry. The first is that CPUs are better for rendering, especially ray tracing, primarily because the amount of available RAM and chip architecture. And, second, that GPUs are better for rendering, especially ray tracing, primarily because of massive parallelism and chip architecture.

“In practice,” says Niles Burbank, senior product manager at AMD, “it boils down to what your preferred software package supports.”

True. But, while some renderers are still CPU oriented, there’s no denying the general industry-wide movement toward leveraging the massive parallelism of GPU cards. GPUs offer attractive amounts of computing power, along with easy scalability. It’s comparatively easy to upgrade your processing power by plugging in an additional GPU card or two.

There are problems, of course. “There is a very well-established infrastructure of tools and applications that take advantage of [x86 CPUs], which is not as mature on the GPU side,” says Burbank.

A CPU-oriented renderer can’t be “ported” to the new GPU environment. The underlying algorithms must be completely re-thought to leverage all the processor cores.

“Applications have been parallel before,” says NVIDIA’s Phillip Miller, “but we’re talking about a different scale. We want hundreds of thousands of processes. In ray tracing, we want a million rays in flight. You have to change the way your application is designed to take advantage of that.”

NVIDIA’s VCA (visual computing appliance) can provide extra GPU computing power for rendering applications. Image courtesy of NVIDIA.

Another oft-cited limitation of GPUs is their comparatively small RAM. Most cards top out at 12GB, which limits the size of scenes you can load and render. “While 12GB may not be enough for some,” says Rave’s Matt Moy, “every time they increase the frame buffer on these cards, they’re opening the door to more and more people.”

Rendering algorithms can be focused obsessively on speed and efficiency, often at the expense of physical verisimilitude. The question was, how few calculations can we perform to produce an acceptably accurate rendering? Can we fool the eye into seeing a reflective object, without following every photon through the scene? Sure we can, and it’s much, much faster than actually performing the computations necessary to calculate physically accurate reflections.

As rendering applications became more sophisticated, and our expectations of the final renders grew ever higher, the tricks got harder and harder to pull off. Finally, a few years ago, things began to flip. “We said, ‘forget about the tricks,’” says Phillip Miller, head of Advanced Rendering Solutions, NVIDIA. “Instead of doing all these CG (computer graphics) approaches, let’s use real physics. You get predictable results and, suddenly, [the program’s] easier to use, because it’s working like the world around you. You put lights where they’re supposed to be and you get the right results.”

The problem has always been one of horsepower, but computing horsepower is cheap and getting cheaper. Consequently, in recent years, renderers have begun to sport physics-based lighting, lenses and materials.

A Workstation is Probably Enough

Extra computing horsepower is perhaps most immediately obvious on your desktop workstation. “In the past,” says Matt Moy, product specialist at Rave Computer, “to work with large sets of data you’d have to log time on a large, expensive cluster. With the advent of GPU (graphics processing unit) computing, we can pack the [same amount] of computing power into a personal workstation.” Rave specializes in building high-performance workstations, which Moy is quick to point out entails more than stuffing a cabinet with the fastest, most over-clocked components. “We’ve very carefully integrated the best of breed components in every category,” says Moy.

Effective cooling is key. The hotter the machine gets, the faster the fans spin and the louder and more annoying it becomes. Worse, heat can slow things down as components react to the high temperatures. Intel’s Turbo Boost feature, for example, allows the CPU to run at faster than its nominal clock speed. This increase, though, is limited by the processor’s power draw and temperature. If it gets too hot, it slows down.

Rave’s cooling design, for example, creates positive pressure inside the chassis. This not only prevents hot spots, but keeps dust from sneaking in through DVD drives, ports and whatever other cracks it can find.

When a Workstation Isn’t Enough

Sometimes, a workstation is inadequate or inappropriate. If you need to plug in some more power, that’s increasingly doable. Specialty computer maker BOXX Technologies, for example, makes the renderPRO, a plug-and-play turbo booster with twin Intel Xeon processors. Plug it into your workstation or laptop and reap instant rewards from its fast RAM and 28 CPU cores.

Taking a different tack, graphics card manufacturer NVIDIA makes the Visual Computing Appliance (VCA), which focuses, unsurprisingly, on GPU power. In addition to its 20 CPU cores, the VCA holds up to eight double-width cards, each with thousands of cores. NVIDIA’s Quadro M6000, for example, packs 3072 CUDA (compute unified device architecture) cores.

“When you’re talking about performance increases of this magnitude,” says Miller, “it completely changes the game. What used to take a coffee break is now interactive. What used to take overnight now takes a coffee break.” Need even more power? You can link multiple VCAs together.

“Honda has 30 machines of this configuration, working together,” says Miller. That’s nearly three-quarters of a million cores. Honda’s VCA cluster produces what Miller characterizes as “interactive reality.”

“They use it for final quality control to examine every piece of the car before it goes to manufacturing,” Miller says.

This kind of setup seems pretty expensive—and it is—until you realize that a physical prototype costs Honda between half a million and a million dollars. “They saved multiple prototypes within the first few months of operation,” says Miller. “That more than paid for their investment.”

Miller points to an example of a taillight. “Honda thought it was designed correctly but, when they previewed it, they realized that the red was bleeding into the white. Had they built it, they would have had a problem, but they corrected the issue before the plastic parts were even made.”

That anecdote hinges on physically realistic, predictive rendering results. A shortcut to fool the eye into seeing realism no longer cuts it. We need results we can trust, even in our “pretty pictures.”