DEVELOP3D - Technology for the product lifecycle

Nvidia and mental images are reaching for the Cloud to offer ray-traced rendering over the web using stacks of GPUs (Graphics Processing Units) instead of CPUs. Set for official launch at the end of November Nvidia’s RealityServer 3.0 platform will enable architects, automotive engineers and product designers to send 3D scenes up into the cloud with the rendered results streamed back over the web. The major sell for this is significantly reduced rendering times, but the tech will also be able to stream interactive 3D to any web connected device including mobile devices - though of course bandwidth will be an issue.

The platform is highly scalable, and more users can be serviced simply by adding more GPUs. Nvidia is already talking to a number cloud computing providers and expects to announce partnerships with several of them later this year, one of them being Amazon EC2 (Elastic Compute Cloud). The cost of cloud-based deployment is expected to be less than one EURO per hour.

While the Cloud computing aspect of the technology is sure to dominate the headlines, of equal interest is the fact that RealityServer 3.0 can be deployed within the confines of a firewall, not only as a GPU-based ‘render farm’ to serve up rendered scenes in double quick time, but also as a means to distribute interactive 3D graphics throughout the enterprise.

The background to this technology is Nvidia’s CUDA programming architecture that enables Nvidia GPUs to carry out computationally intensive tasks usually reserved for CPUs. CUDA was used to devise a new GPU-based rendering mode called iray, which is based on mental images’ mental ray 3.8 rendering engine. This is different to most rendering technologies which rely on CPUs to do the calculations.

On the hardware side, RealityServer consists of multiple Nvidia’s GPGPU (General Purpose GPU) Tesla cards, which are used to render out the scenes plus a few CPUs, which are really just used for housekeeping, says Nvidia.

The technology is already primed up to be exploited a number of 3D CAD companies. There are over ten major CAD applications that already use mental ray, including Autodesk (3ds Max, Inventor, Revit), SolidWorks, Dassault Systemes (CATIA), and most recently PTC (Pro/Engineer Wildfire).

The critical technology here is mental ray 3.8, which is due for release later this year and will enable GPU-accelerated mental ray rendering for the first time. Once these vendors implement mental ray 3.8 into their core products, they would have all the tools to hook up to RealityServer, says mental images, but for some CAD software, particularly the more mature products that carry a lot of ‘architectural baggage’ the implementation would not be trivial. That said, mental images told DEVELOP3D that development is already underway at many CAD companies and it expects to see applications supporting RealityServer next year.

While mental images was unable to name all the names it did confirm that all of the aforementioned CAD developers are already working on systems that would allow them to virtualise their applications or to at least have server-based collaborative solution directly connected to their applications. As a result the company is confident that this technology is well placed to take a lot of work off the CAD developers’ plate as they are essentially offering them a whole suite of tools to get started faster instead of doing everything themselves. mental images also disclosed that Autodesk showcased the technology at a conference in Munich, Germany only yesterday.

In terms of the actual rendering technology RealityServer is a progressive renderer, so users are able to get a good idea of the final render in seconds or minutes, even though the final rendering may take hours. For comparative render times between CPU and GPU-based solutions it was hard to draw mental images on exact figures. However, the company did provide an example of an architectural scene that took 45 mins to render on a four Tesla cluster system and 8-10 hours on a more traditional four core CPU-based system. That said, it was wary of comparing apples and oranges as the scenes were not identical because the GPU renderer is slightly different from the CPU renderer in terms of shading technology. The company did say that it that would be providing benchmark results from customers next month and the early results are encouraging.

While for most CAD uses the emphasis is likely to be on using Reality Server as a rendering server, mental images was keen to point out that it also provides a platform on which companies to build applications that utilise the technology in different ways. In the automotive sector, for example, it is already working with a number of manufacturers on projects to develop and enhance their in-house design / review pipelines. A dedicated car paint shader is also in development and will be released early next year.

For those that wish to set up their own facility there are three different packages. In true American style there is no small - instead just a M, L and XL. Medium is a 2U rack mounted system with 8 Tesla GPUs and is suitable for smaller architectural offices and product design teams with 10s of concurrent users. Of course, this depends on the intensity of use and some customers may need to dedicate four GPUs for a single task. The ‘Large’ package features 32 Tesla GPUs for 100s of concurrent users, while ‘XL’ features 100 Tesla GPUs for serving 1,000s of users over the web.

Nvidia is still working on overall system costs, but with a single Tesla cards costing in excess of 1,000 EUROS one may speculate that a medium system would cost around 15,000 - 20,000 EUROS just for the hardware. On the software side, however, customers should expect a one-time licensing cost of 2,000 EUROS plus 20% maintenance per Tesla card.

From complex architectural visualisations and 3D city modelling to product design and automotive styling, the CAD-centric target markets for RealityServer are huge. And with mental ray already the rendering engine of choice for most major CAD developers, one may speculate that it’s only a matter of time before RealityServer becomes a widely supported platform for CAD.

What makes this technology particularly interesting is the fact that it is designed to use GPUs in the Cloud and not CPUs, but this is also a current barrier to deployment. None of the large Cloud service providers currently offer GPUs in their facilities, but Nvidia expects this to change early next year. This coupled with the expected release of RealityServer-compatible CAD products should make 2010 a very interesting year for rendering in the Cloud.