I’m using an old computer with Blender and I plan to buy a new computer.
To render (simple) scene I’m currently using CUDA with GPU (gtx770) and CPU (i7-4770) together
My first thought was about a computer with an RTX 3080 and an i9-10900K
But the RTX 3080 is very hard to buy (due to shortage) and both RTX 3080 and i9-10900K are expensive.
So I checked AMD solutions and I found that the configuration Radeon rx 6900 xt with and AMD ryzen 3950x is not too expensive and not too bad in term of computation power.
But I wonder if it’s a good idea ? Does Blender work also well with AMD ? With both CPU and GPU together ? Do you know any limitations ?
I worked with both, AMD and Nvidia. The most obvious thing I took away from that: Dont go AMD for GPU rendering.
For gaming, AMD GPUs are alright, but for rendering with OpenCLs current state (in blender) is simply abysmal compared to Nvidias Optix.
Check https://opendata.blender.org/
RTX 3080 is a smart pick, go for that if you can! 3060TI / 3070 works too price/perf wise. But has less Vram ofc.
As you already said its an availability problem more than anything else.
For CPU, I’d prolly go AMD. 3950x or 5950x. Very similar price/performance wise. (For my regional price tags)
Big no for GPU.
Not advisable with a high-end GPU anyway. If you have multiple frames to render just go CPU rendering in one blender instance and GPU rendering in another. Its a bit faster even with the new tile stealing by Lukas.
For my regional prices: 1500€ for that kind of performance with 16gb (6900xt) VS a way better perf for just 350€ more and 24gb (RTX 3090) is a definitive “abysmal” for me.
(The 3080 is simply unavailable for me here, or almost as expensive as the 3090)
OpenCL render kernel compilation takes longer and is more often than Optix shader compilation, in my experience from a year ago. (owned a Vega 64)
So depending on your scene size and significance of your scene changes (changing Shaders, i.e.) you need to compile the render kernels again. This adds up quickly in an viewport-render workflow.
So what you see on Open Data and guru3D is also only the final Benchmark renders, not necessarily your true render times when working.
With AMDs current API implementation in blender (OpenCL) I wouldnt even buy one at 70% of its current price.
Nvidias Optix is just that good in comparisen.
The main difference is the Ray Tracing for Nvidia is supported (Optix) and for AMD it isn’t (yet but will - based on some comments made by AMD devs sitting in these forums)
So if you want to use Cycles, it works, but not at 100% of its current capability, If you want to try AMD’s proRender with Blender, then you’ll get a “near” 3090 perforance.
The BMW isn’t best scene, (BMW not great due to its limited complexity, and AMD has one… issue … the kernel recompile that eats a bit of the initial render)
