It only takes 32 AMD GPUs and 33 hours to run a 40 billion cell simulation of the Concorde supersonic plane
Note: This article originally appeared on pcgamer.com, click here to read the original piece.
That AMD-powered GigaIO SuperNODE server really slaps.
Slapping one of the most intense Computational Fluid Dynamics (CFD) simulations ever on the world’s largest single-node GPU server sounds like a fun weekend activity, honestly. That’s exactly how Dr. Moritz Lehmann spent their time recently with a Concorde focused FluidX3D simulation, and the results are intensely beautiful.
There are a few configurations of the GigaIO SuperNODE server. One has 24 Nvidia’s A100 graphics cards stuffed in it, whereas the particular design ProjectPhysX got to use contains 32 AMD Instinct MI210 64GB GPUs. And damn do they slap.
You can check out the benchmarks over on Dr. Lehmann’s ProjectPhysX Github page, but the general gist is that it took just 33 hours to run the simulation—including “29 hours for 67k timesteps at 2976×8936×1489 (12.4mm)³ cells, plus 4h for rendering 5×600 4K frames.” If that all sounds a little complicated to you, all you need to know is that this was a damn complex simulation containing 40 billion cells overall, and Flight Sim has nothing on it.
The very fact it was able to run the simulation in a single weekend is a marvel unto itself, as the original poster notes “Commercial CFD would need years for this”.
For those who overlooked the intensely pixelated scene caused by Reddit’s compression algorithm, and headed for a look at the 4K version on YouTube it’s clear something went very right.
Thanks to OpenCL (Open Computing Language), ProjectPhysX says “FluidX3D works out-of-the-box with 32-GPU scaling”, which made the whole process a lot easier with no code changes required on their part.
The purpose of the test was to get an idea of how well the “newly implemented free-slip boundaries” are working in the simulation. These are meant to make the model more accurate than previous implementations of so-called “no-slip boundaries”.
In the comments, ProjectPhysX gives a little explanation of the difference between the two: “No-slip enforces that the fluid velocity at the wall is the same as the wall velocity. Walls ‘drag along’ the fluid, causing more friction. Free-slip means the fluid velocity perpendicular to the wall is not restricted, the fluid can freely glide along the wall without friction.”
All in all, it’s a very exciting use of the GigaIO SuperNODE, and one that I’m sure has a number of exciting implications for the future of aviation. The better we can simulate these things, the less Icarus-style experiments we have to do in the real world.
View source version on pcgamer.com: https://www.pcgamer.com/it-only-takes-32-amd-gpus-and-33-hours-to-run-a-40-billion-cell-simulation-of-the-concorde-supersonic-plane/