FluidX3D running AMD + Nvidia + Intel GPUs in "SLI" to pool together 132GB VRAM скачать в хорошем качестве

FluidX3D running AMD + Nvidia + Intel GPUs in "SLI" to pool together 132GB VRAM

I made this #FluidX3D #CFD simulation run on a frankenstein zoo of 🟥AMD + 🟩Nvidia + 🟦Intel #GPU​s! 🖖🤪 This RGB SLI abomination setup consists of 8 GPUs from 3 vendors in one server: 1x Nvidia A100 40GB (2 domains) 1x Nvidia Tesla P100 16GB (1 domain) 2x Nvidia A2 15GB (1 domain each) 3x AMD Instinct MI50 (1 domain each) 1x Intel Arc A770 16GB (1 domain) I split the simulation box with 2322×1857×581 = 2.5 Billion grid cells (132GB VRAM requirement) up into 9 equal domains of ~15GB each, which run on 8 GPUs. The A100 is fast enough to take 2 domains while the other GPUs each get 1 domain. This is 5 completely different GPU microarchitectures seamlessly communicating over PCIe 4.0 x128. Under #OpenCL they are all created equal and don't care what vendor the GPU is which computes the neighbor domain. This demonstrates that heterogenious #GPGPU compute is actually very practical. FluidX3D users can run the hardware they already have, and freely expand with any other hardware that is best value at the time, rather than being vendor-locked and having to buy more expensive GPUs that bring less value. The demo setup itself is the Cessna-172 in flight fir 1 second real time, at 226 km/h airspeed. 159022 time steps, 11h27min runtime consisting of 9h16min (compute) + 2h11min (rendering). Setup: https://github.com/ProjectPhysX/Fluid... Cessna-172 3D model: https://www.thingiverse.com/thing:814... I created the FluidX3D CFD software from scratch and put the entire source code on #GitHub, for anyone to use for free. Have fun! https://github.com/ProjectPhysX/FluidX3D Huge thanks to Tobias Ribizel from TUM Campus Heilbronn for providing the hardware for this test!