CPU shader with AVX-512: happy jumping
HTML-код
- Опубликовано: 22 авг 2024
- This test shows the performance of a CPU with AVX-512 instructions when you let it do embarrassingly parallel GPU tasks, such as running a shader.
The c++ code can be downloaded from github.com/jan...
The corresponding glsl shader on an NVIDIA GeForce GTX 1060 GPU gives 200 FPS, i.e. about 20x faster.
The shader is based on Inigo Quilez' happy jumping, see
• LIVE Coding "Happy Jum...
How much better do you wager an i9-10980XE with 18 cores, 36 threads, with mesh bus cache frequency at 3.2 GHz, and RAM clocked at 4000 MHz in quad-channel with 1 T command rate and 15-15-15-32 primary memory timings, clocked at 4.0 GHz in AVX-512 would perform compered to the chip in your setup?
P.S. Very impressed with your results BTW!
Hi, thanks for the nice response. As for your question: I don't really know. You should try it to get the most accurate answer. The cpu shaders scale pretty well with the amount of threads, so there I would expect your setup to be around 2.5 times faster, and per thread your setup will also perform faster but I don't know enough about hardware to give a reasonable guess here.
How does this compare to without AVX?
Hi, I didn't implement this shader system without AVX so I can't tell for sure, but in the past I did another experiment where I made some similar shader-like environment on the CPU in c++. There I got as result that SSE2 was 3.8 times faster, and AVX2 was 6.1 times faster than running the shader without vector instructions. I don't have timings for AVX-512.
@@janmaes2712 you need to do AVX2 version at least.