LuxMark 3.1
OpenCL is a platform for harnessing GPU power for activities other than real-time 3D rendering to screen, also known as GPGPU. Unlike NVIDIA's CUDA platform, OpenCL is open source and can be ported to anything with processing power. So drivers are available for CPUs as well, both from Intel and AMD.
A popular tool for testing OpenCL performance is LuxMark. We haven't run this on many workstations before, so we only have one comparison amongst our past reviews. We ran the Sala scene on CPU only, GPU only, and then both.
Unfortunately, Intel's Core i9 does not currently appear to be supported by its OpenCL drivers, so we were not able to run the LuxMark 3.1 render on the Intel system's CPU, or on the CPU and GPU combined, using OpenCL. However, we did run the barebones C++ version of the LuxMark benchmark on both systems.
You get 82 per cent more performance form the Ryzen Threaderipper's 1950X's 16 cores compared to a Ryzen 7 1800X's 8 cores. In C++ mode, the Intel Core i9 7980XE is receiving a 16 per cent benefit from its two extra cores. The GPU scores are very interesting. The dual-GPU AMD Radeon Pro Duo in the V25R gives it a 12 per cent OpenCL advantage over the AMD Radeon Vega Frontier Edition, although the NVIDIA GeForce GTX Titan Xp is 14 per cent faster still.
Interestingly, it's swings and roundabouts where overall OpenCL performance is concerned between the V25R and S16T systems. What the latter gains in CPU performance it loses in GPU, giving an almost identical overall score. But the Radeon Vega is likely to be a better choice than the Radeon Pro Duo for a general workstation, because its modelling performance is significantly better.
LuxMark is a synthetic benchmark – it doesn't correspond directly to an application actually used in a production environment. It also uses OpenCL, the GPGPU API that is openly available to all hardware with drivers to run it.
But NVIDIA has its own proprietary GPGPU API called CUDA, which stands for Compute Unified Device Architecture. This name makes it sound as generally available as OpenCL, but in fact only NVIDIA graphics cards support it. There are a number of CUDA-enhanced 3D renderers out there, such as Octane Render, Redshift, and V-Ray.
The full list can be found on NVIDIA's website. But support for both CUDA and OpenCL is available in the latest version of the Open Source Blender, so let's turn to that application next.
Blender 2.79: Gooseberry Production Benchmark
Blender is a free and open source 3D creation suite. It supports the entirety of the 3D pipeline—modeling, rigging, animation, simulation, rendering, compositing and motion tracking, even video editing and game creation. The latest version at the time of writing, 2.79, supports rendering on the GPU as well as the CPU. In GPU mode, it will render using OpenCL with AMD graphics cards, and CUDA with NVIDIA graphics cards. For this test, we used the Gooseberry Production Benchmark. Project Gooseberry is the code name for the Blender Institute's 6th open movie, Cosmos Laundromat — a 10-minute short, the pilot for the planned first-ever free/open source animated feature film. The benchmark renders a single frame from this film in intermediate quality.
The render results are a little surprising. The Intel CPU is 62 per cent faster than the AMD one, but the latter is 11 per cent faster than when running the render via OpenCL on the Radeon Vega Frontier Edition. The surprising bit came when we tried to run the render on the Titan Xp. This was using CUDA rather than OpenCL, and is clearly not very well optimised at the default settings (32 x 32-pixel tiles) because it took over ten times longer than the CPU.
However, a bit more research showed us that GPU rendering on Blender can be very sensitive to tile size, so we tried a much larger 256 x 256-pixel tile size instead. This improved the AMD Radeon Vega Frontier Edition's performance by 27 per cent, so that now it was quicker than the Threadripper CPU, but allowed the NVIDIA GeForce GTX Titan Xp to be nearly ten times faster. Although both GPUs are sensitive to tile size, the Titan Xp's performance is much more significantly improved by tweaking this setting, showing just how important optimising your settings can be. As before, the Intel-NVIDIA combination wins on raw performance, but you need to configure things correctly.
Corona 1.3 Benchmark
Corona Renderer is a new high-performance (un)biased photorealistic renderer, available for Autodesk 3ds Max and as a standalone CLI application, and in development for Maxon Cinema 4D.
The development of Corona Renderer started back in 2009 as a solo student project of Ondřej Karlík at Czech Technical University in Prague. Corona has since evolved to a full-time commercial project, after Ondřej established a company together with the former CG artist Adam Hotový, and Jaroslav Křivánek, associate professor and researcher at Charles University in Prague.
Despite its young age, Corona Renderer has become a production-ready renderer capable of creating high-quality results. The Corona Benchmark outputs a single ray-traced frame from a sample production.
As with most of our other CPU tests, the Intel Core i9 is ahead of AMD's Ryzen Threadripper, in this case by 32 per cent.
Overall, it's no surprise that Intel's 18 cores beat AMD's 16 cores in most rendering tasks, despite the slightly lower clock speed of each one. However, the margin varies from 16 per cent in the LuxMark 3.1 C++ test to a whopping 62 per cent in Gooseberry, so there's no clear picture of how dominant the Core i9 is – it really depends on what you are running.
The picture is the same with the AMD graphics compared to NVIDIA's, and in fact even more pronounced. The GeForce GTX Titan Xp clearly has quite a bit more grunt than the Radeon Vega Frontier Edition, but if the GPGPU renderer isn't well optimised, this doesn't mean very much. AMD has put a lot of work into its OpenCL support, and its ProRender plugin for many professional applications really has a lot of potential to save massive amounts of production time.
Be sure to check out our sponsors store EKWB here