Maxwell 4 GPU: What You Need to Know

October 21, 2016 News, Technology 32 Comments

What is Maxwell GPU?

The recently released Maxwell 4 includes a brand new version of our rendering technology that runs on nVidia GPU devices. This new GPU version is not a new feature but rather a replication of the functionality of the Maxwell production engine from CPU to GPU.

The new GPU version can be used not only for interactive rendering but also for producing final images.

It is available as a rendering option in the Maxwell 4 render settings panel, so in V4 you can choose whether to render in CPU or GPU depending on your scene specifics.

How does the GPU compare to the CPU in V4?

Maxwell GPU supports most of the key features of the CPU version. It is unbiased, physically-based and spectral. It uses the same models as the CPU version for everything: materials, sky, film, lenses, tonemapping, etc., so you don’t need to apply special shaders or settings when using GPU, just change from one render engine to the other and click “render”.

For a detailed comparison table please check our web here.

Is GPU faster than CPU?

In general terms we can say that the GPU option will speed up the rendering process greatly, it can be up to 10 times faster than its CPU equivalent. For instance, it will be very fast in scenes with simpler lighting setups and perhaps slower in complex ones where the light rays have to bounce a lot. Notice that GPUs have thousands of computing cores, but these are not very fast, while CPUs have tens of cores but faster, so with the GPU Maxwell will be computing thousands of rays at the same time but slower and with CPU you just compute tens of cores but faster. We will soon provide some comparison scenes to illustrate the differences.

The GPU option is highly suitable for product design and architecture exteriors.

EXAMPLE: a typical product design scene with one object lit with IBL textures, or an architecture exterior with diffuse/opaque materials will render very fast. However, an interior with strong indirect lighting, subsurface scattering or any other phenomena that involves complex light transport behaviors will render more slowly.

N.B. It is important to know that the speed of the GPU rendering will largely depend on the quality of the graphics card used (number of CUDA cores, their speed and memory interface speed).

Soon we will have a renewed Benchwell available, where users will be able compare their Maxwell performance on different devices and graphic cards.

What are the current limitations of the Maxwell GPU option in V4?

Even though the GPU version aims to mirror Maxwell on CPU with the respective improvements of rendering using a graphic card, at this moment there are still some features missing and some limitations we have faced due to the nature of GPU rendering.

  • MEMORY: One of the main limitations of GPUs in comparison with CPUs is the amount of memory available. Massive scenes (typically VFX ones) are difficult to fit in the current generation of GPUs. However, most of the scenes will fit in just fine!
  • STABILITY: In general GPUs are more unstable than CPUs so there might be some issues related to drivers and particular system setups. We will deal with these constantly, by releasing new builds to the Early Builds area to make the GPU version more and more stable.
  • CUDA ONLY: Maxwell GPU only runs on nVIDIA devices that have CUDA support. 
  • OS COMPATIBILITY: Mac machines by default use AMD cards so even when the GPU option in V4 is available for OSX in the interface, you won’t be able to run the render using the GPU as AMD cards are not supported. Nevertheless, will be able to send the render job through the network to other machines that mount nVidia cards.
  • MULTILIGHT: Multilight™ is not available in the GPU version of Maxwell 4 (it is included in CPU rendering).
  • MULTI-GPU: Currently the GPU version is available on one unit only, but we are certainly planning to support multi-GPU in the near future. Then you will be able to render with several GPUs in the same system. We intend to have this feature available by the end of the year.
  • MAXWELL GRASS: Maxwell Grass is not yet available in Maxwell GPU as it requires a lot of memory.
  • RENDER BOOLEANS: In the GPU version Render Booleans are not working
  • RENDERED OUTPUT: For the moment in GPU we are supporting the following channels: Alpha, Z-buffer, Material ID, Object ID, Normals, Position and UV, but we will be try to add the rest in the upcoming updates.

Which graphic cards are supported?

Maxwell GPU is only compatible with nVidia CUDA, which means that it only runs on nVIDIA devices with CUDA capabilities.

We are supporting Maxwell and Pascal architectures (CUDA 5.0 and up).

It could also work with Kepler cards but the experience won’t be so good. Older cards such as Fermi or Tesla probably won’t work at all. Check your graphics card specifications to see if it will work. In this Wiki page you can find a table with the different cards and architectures.

Currently, we are not able to support AMD cards but if the situation changes in the future and we are able to do so, we will consider it.

Is Multilight™ available for GPU rendering?

Multilight™ is available in Maxwell 4 but only when you render in CPU. It is not available for GPU rendering because in the GPU cards there is not enough memory to keep separate image buffers per light. Even though Multilight™ for GPU is not on the short-term roadmap yet, we will try to break this limitation in 2017 as the amount of memory included in the graphics cards increases.

Please note we have also released Standalone Multilight™, which is a tool that can be used regardless of whether you have Maxwell. You can try it out here! On the web there are demo scenes for Multilight™ available.

Is there a way to use Maxwell GPU on Mac OS?

Maxwell GPU rendering is not currently supported on Mac OS. For GPU rendering, nVIDIA graphic cards with CUDA support are required, and for the time being Apple does not provide machines that include these.

If you do have an older Mac model with an nVIDIA card, we cannot guarantee you a great experience, because those cards are obsolete and have not been tested. In addition, while GPU connections are possible through third-party Thunderbolt devices, these have not been thoroughly tested and we cannot guarantee stability.

Should Apple include graphics cards appropriate for GPU rendering in the future, we will likely support these.

What’s coming next?

Apart from general stability and better memory management, there are many features we would like to add and support issues to resolve. We are constantly working to make this version more complete and suitable for our users’ needs. Please do check the roadmap on our web, we will be updating this according to current developments.

Some of the features & fixes on the roadmap include:

  • General stability and drivers compatibility for as many devices as possible
  • Better memory management of large resolutions
  • Complex IORS
  • Coatings
  • Dispersion
  • Missing visibility flags
  • IES, Spots and Projectors
  • Subsurface Scattering
  • Thin subsurface scattering
  • Procedural textures
  • Multi GPU support
  • Custom lenses
  • Procedural geometries (hair, particles, grass, etc)
  • Support for CPU and GPU running at the same time
  • Additive layers
  • Missing minor features (blocked emitters, etc)
  • Render region