Equalizer 0.5.4 – OpenGL Parallel Rendering
The new release of Equalizer is available. Equalizer is an open source project providing a minimally invasive programming interface and resource management system for parallel, scalable OpenGL applications. It allows an application to run unmodified on any visualization system, from a simple workstation to large scale graphics clusters and multi-GPU workstations.
Links:
- Equalizer 0.5.4 release notes
- DB (sort-last) load-balancing
The Equalizer example application running on the same display wall.
The following video shows volume rendering using three GPUs and dynamic 2D load-balancing. The tiling is adapted each frame automatically and transparently to the application by Equalizer. The data set has 256^3 voxels, and is rendered using 2560 slices using the eVolve scalable volume renderer.
Related posts:
256^3 volume rendered by 3-GPUs, sorry guys it is not impressive at all; CPUs based volumetric ray-caster can render the same dataset on MacPro with sampling density 16 samples per cell with similar performance.
Stefan,
The examples shipping with Equalizer focus on showing how to use the API, not on the rendering algorithm. I am sure that another volume rendering algorithm is faster than the one used by eVolve.
The point is to show scalability, and the three-GPU version runs much faster than the one-GPU version. I haven’t benchmarked this particular setup, but we’ve done extensive benchmarking on a 16-node cluster.
I also increased the sampling to 10x the normal amount to bring down rendering performance to be able to record it properly – the tool I use simply can’t record 60Hz.
The irony is, the better multi GPUs scalability the less impressive result for GPU based VR since single GPU setup goes down very “effectively” as well. Actually, “sampling to 10x” is close to the minimum sampling density for high quality VR and it is a routine interactive sampling density I’m getting from CPU based VR engine. Anyway, it is a freeware and definitely many people will appreciate your efforts.
Regards,
Stefan
Stefan,
I don’t quite get your point. Equalizer is a framework to write parallel OpenGL application, not a volume rendering framework.
It might be that the engine you are referring to outperforms the very simple approach chosen by our example application. I somehow doubt that a CPU can be faster than a GPU, but I am not a volume rendering specialist.
The 2D loadbalancing and task decomposition is applicable to any OpenGL-based renderer which runs too slow on a single GPU. SLI and Crossfire do something similar on a much lower level with more limited results.
Cheers,
Stefan.
>I somehow doubt that a CPU can be faster than a GPU,
It depends on algorithms. Any inherently divergent algorithms do not map effectively to very wide SIMD architecture of today’s GPU. For example CT reconstruction maps very well and 9800GTX outperforms dual E5440 by factor [3x…10x]. Rasterization graphics is another example of GPU total domination; one of VR technique such as VR texture mapping fits well to GPU. However ray-tracing or/and volumetric ray-casting/tracing do not map well due to their inherently divergent nature (code path of each ray has tendency to be unique even for neighboring rays). Thus, the phrase “faster than a GPU” makes a little sense without reference to algorithm. I just run into volume rendering demo as an example of Equalizer capabilities and was a little surprised that such modest VR performance is considered as impressive.
Regards,
Stefan