Sign up for the newsletter

Signup for the Newsletter

Email

Can raytracing really kill raster graphics?

Tuesday, February 5th, 2013 | Posted by Jim Thacker

Imagination Technologies’ new Caustic Series2 raytracing boards promise true real-time raytracing for the cost of a professional graphics card. But can they succeed where previous specialist render hardware has failed? And what benefits do they offer over a standard GPU? We spoke to Imagination and its beta testers to find out.

Updated 7 July 2014: Since this story was posted, the R2100 and R2500 cards have been discontinued, alongside Caustic Vizualizer for Maya. Imagination Technologies continues to develop its PowerVR mobile GPUs.

The idea of dedicated hardware to run your 3D software may bring back memories of the 1990s, of SGI workstations and turnkey systems – or more recently, of ArtVPS, and its now-defunct PURE and RenderDrive units. Yet Imagination Technologies, which last week launched its new Caustic Series2 line of raytracing acceleration boards is definitely looking forwards, not backwards.

According to the chip design firm, its technology will make raytrace rendering ubiquitous even on handhelds within a few years – and, in the longer term, render the rasterisation viewports in your 3D software obsolete.

From Apple to OpenRL
Imagination has been making GPUs since the 1990s – its PowerVR graphics cores power many of the world’s mobile devices, including the iPhone, iPad and Kindle Fire – but its raytracing technology is much more recent. Founded in 2006 by three Apple veterans, Caustic Graphics was subsequently acquired by Imagination, where it now forms the firm’s Caustic Professional division.

“In the past, folks had focused on the computation part of the raytracing problem: doing the ray-surface intersections, for example,” says Imagination’s director of business development, Alex Kelley. “That problem was solved with Moore’s Law: [modern CPUs] are now fast enough to deal with ray-surface intersection.

“The Caustic founders looked at it very differently: as a database problem. The problem with raytracing is that it requires the the entire scene be in memory for efficiency and GPU processors just aren’t capable of doing that: typically, you have to go outside of cache. [With our technology], we can make raytracing as fast as rasterisation. Believe it or not, when you start working with very large models, raytracing can be even faster than rasterisation.”

At its lowest level, Caustic’s technology includes OpenRL, a cross-platform raytracing acceleration framework, available on a free perpetual licence. At the mid level, there is the PowerVR Brazil SDK, a commercial rendering tool kit used by developers to add raytraced viewports to their 3D applications. And at the top end, there are Caustic’s own Visualizer plugins. It is this software that the new Series2 cards are designed to accelerate.

Caustic Visualizer for Maya. To date, the only other software to use the PowerVR Brazil SDK is Rhino’s Neon rendering plugin – although Imagination now has two engineers working full-time with third-party developers.

Launched at Siggraph last year, the $299 Caustic Visualizer for Maya adds a fully interactive real-time raytrace mode to Maya’s working viewports, supporting accurate reflections, bounced light and soft shadows. A 3ds Max version is in development, and should be available in beta later this year.

“If you plug the card in, it turns [Visualizer] from semi-interactive into semi-real-time,” says director of product management Michael Kaplan. “You can edit in the viewport, move lights around, change materials, change the geometry – even at the vertex level – and see what it will look like in the final-frame render.”

But unlike, say, iray or V-Ray RT, Visualizer is not intended simply as a final-frame or preview renderer. Plug in one of the new Series2 cards and you also enable its modelling modes.

“They’re raytraced modes but not trying to look like a final-frame render,” says Kaplan. “They add ambient occlusion for shape modelling, or you can do point-sample raytracing if you want to see exactly where shadows fall, or exactly where another object [is reflecting] in a mirror. They’re even faster than final-frame mode to the point where, with the cards, they really are real-time.”


 
Although it provides 16GB of on-board memory, Caustic’s R2500 is a bus-powered single-slot card. Maximum power draw is just 65W: note how small the cooling fans are compared to a comparable professional GPU.

Structurally, the Series2 boards are very simple: a PCI interface, the custom raytracing processors (two in the case of the $1,495 R2500, one in the case of the $795 R2100), and memory – a lot of memory. The R2100 has 4GB; the R2500 has 8GB per raytracing chip, or 16GB in total. For comparison, Nvidia’s current top-of-the-range Quadro and Tesla GPUs each have 6GB of graphics RAM, as does AMD’s current top-of-the-range FirePro.

And only the geometry has to be loaded into graphics memory, since shading is done on the CPU, further increasing the size of the scenes that the cards can process. According to Imagination, the R2100 and R2500 can store up to 60 million and 120 million triangles respectively.

But why not just use the GPU?
According to Kaplan, in Imagination’s own tests, conducted on a top-of-the-range 2.9GHz 16-core HP Z820 workstation, plugging in a Series2 R2500 board sped up Vizualizer’s final-frame mode by an average of 2.6 times. The speed increase was even greater for the modelling modes.

To put that into context, Kaplan claims that when Imagination ran the same test scenes on the same machine in iray, the largest speed increase obtained upon plugging in a Nvidia Tesla C2075 graphics card was 2.0 times. In other words: a $1,495 raytracing board gave a bigger performance boost than a $2,200 GPU.

Cross-renderer comparisons are notoriously controversial – and, it should be noted, these are Imagination’s own benchmark scenes. Nevertheless, these figures are broadly in line with those reported by the beta testers we’ve spoken to.


Beta tester Gavin Greenwalt used the R2500 on this spot for Ace Hardware. “It enabled me to light the shot in the viewport as if I was looking at the final composite,” he says. “Lighting and look dev only took about two hours.”

The fact that larger scenes can be loaded into memory than on an equivalent-price GPU endeared the R2500 to Gavin Greenwalt of Seattle-based animation and VFX house Straightface Studios, who beta-tested the board for eight months prior to its commercial release.

“I’ve tried numerous GPU renderers in the past, but the most frequent problem I run into is memory limits when previewing production scenes,” he says. “For instance, The Ace Hardware spot [shown above] required tens of millions of polygons and multiple 4K textures, since the set extension had to be photorealistic at HD resolutions. Normally, I’m unable to get an accelerated preview [but] the Caustic card is the first hardware solution that we’ve found that can handle real production scenes.”

Greenwalt notes that the increase in Caustic Visualizer’s performance when the R2500 is installed varies from scene to scene, but comments: “The real advantages I’ve found are from scenes that have a lot of complex shaders, reflections or global illumination. Interior scenes are often lit more significantly from indirect sources than direct sources, so traditional rasterised viewports offer pretty much useless lighting previews. Other accelerated raytracing previews [including iray and V-Ray RT] offer similar capabilities, but they often need a second or two for their acceleration structures to update, which … doesn’t come close to the multiple-frames-per-second interactivity of the Caustic card.”

Interactivity was also a key selling point to Greenwalt’s fellow beta tester David S. Perkins of Silent House Productions, which specialises in designing and visualising stage sets for concerts, broadcast events and installations. Perkins used Caustic Visualizer for Maya, running on the R2500 board, to conduct materials and lighting tests on props before the completed 3D assets were assembled in a finished scene.

“To be able to change all the settings interactively in the viewport is invaluable,” says Perkins, who estimates a typical speed boost of “two to three times” over Caustic Visualizer running on the CPU alone. “For interactive visualisation with reflections and materials, I don’t think there’s anything better on the market for the price.”

YouTube Preview Image
Silent House used the R2500 to visualise the set for Pink’s performance at last year’s American Music Awards. Its capacity to display accurate reflections in real time helped avoid problems when filming the two mirrors.

Both Greenwalt and Perkins note some limitations with the current-generation Series2 cards – lack of support for motion blur and photon mapping for Greenwalt; in Perkins’ case, lack of support for Maya’s volumetric light fog. However, neither feels that the idea of purchasing dedicated hardware to speed up a single software package is in itself an obstacle.

“I don’t have a problem with being limited to a single software package: 90% of our production goes through a single renderer already,” says Greenwalt. “We try and update our rendering capabilities and workstations every two to three years. Anything we buy today will be obsolete by the time we refresh our hardware, so we don’t feel terribly locked in.”

Greenwalt also notes that, in practice, the situation is little different to GPU-based rendering. “If you want interactive raytracing [under real production conditions], you need one GPU to handle the viewports and operating system and a second GPU dedicated to rendering for your workstations,” he says. “As far as I’m concerned, GPU rendering already demands a second dedicated card.”

Succeeding where others failed
For Imagination Technologies, it is precisely this specificity that distinguishes the Series2 cards from previous generations of dedicated rendering hardware.

ArtVPS['s hardware products were] general-purpose computers dedicated to raytracing, and it really didn’t turn out to be very cost-effective,” says Michael Kaplan. “They didn’t have a special way to accelerate raytracing: they just had a bunch of compute.”

Kaplan draws a similar distinction between Caustic’s technology and contemporary GPU-accelerated rendering. “You can do [raytracing] on a GPU, but you do it by wasting an awful lot of the processing power … because the memory architecture isn’t suitable. You get pretty good results, but you need large numbers of graphics cards, a huge chassis, a big power supply and an air-conditioned office – you’re certainly not going to put that into a tablet. We can do with 30W what would require between 500 and 1,000W of general compute.”

Coming soon to a tablet near you?
Imagination also notes that, as the main player in the mobile graphics chip market, it has not only more financial muscle than previous rendering hardware developers but the capacity to reach a far wider range of users.

“We’re starting off with separate cards because we want to address the desktop market, but as Imagination combines the raytracing technology with its PowerVR hardware, we’ll start to enter the mobile space,” says Kaplan. “Imagine that [in a couple of years] a tablet has the same raytracing power as we’re seeing here on the desktop. That’s pretty amazing.”

“Ultimately, we see raster graphics viewports being replaced everywhere by raytracing,” concludes Kaplan. “That’s of course a long-term bold statement … but there are just so many advantages to raytracing.”

Update: The article originally stated incorrectly that Caustic was founded in 2009, not 2006.

Visit the Caustic Professional website

Visit Imagination Technologies online

Tags: , , , , , , , , , , , , , , , , , , , , , , , , ,

Comments

You can follow any responses to this entry through the RSS 2.0 feed.

You can leave a response, or trackback from your own site.


  • Dan

    I like their SDK for Open RL. It gets right to the point. It’s a little on the low level side, but they said the same thing in the article and they have mid and higher range products.

  • David

    Thanks for being accurate in pointing out the companies that tried this and crashed and burned. It the one reason I am holding back on this tech. There has to be much, much wider support for it to really succeed. I hope it does, and the card prices come down like gpu’s have done. Otherwise why not just use a GTX 690.

  • mf

    I was thinking is tesla system until this moment. However this seems a reasonable stuff.
    I still have one question, does this boost softs like After Effects as well? Or just for Maya and so for…?

© CG Channel Inc. All Rights. Privacy Policy.