It's wonderful that the entire paper is available freely from the journal. Most journals these days are pay-for-access (and none of the payment makes it to the authors).
https://github.com/Luthey-Schulten-Lab/Minimal_Cell
https://github.com/Luthey-Schulten-Lab/Lattice_Microbes
This is an impressive result. They've found a nice level at which to simulate. Down at the atom level would take too long. They got this to work at the chemical reaction in 3D space level, low enough that the model maps directly to the biochemistry.
They don't seem to say how many GPUs they used, but the article seems to indicate small numbers, not entire data centers. Current speed is about 1/30 of real time for one minimal cell, which is pretty good.
> The GPUs used for spatial simulations included NVIDIA Titan V and NVIDIA Tesla Volta V100 GPUs, which took 10 h and 8 h to simulate 20 min of cell time, respectively.
So apparently they tried 3 different GPU configurations:
1. A single Nvidia Titan V, for a non-mixed model, which took 10-hours to simulate 20-minutes.
2. A single Nvidia Tesla Volta V100, for a non-mixed model, which took 8-hours to simulate 20-minutes.
3. No GPU's, for a well-mixed model.
Regarding mixing: A well-mixed model is one where all chemicals are assumed to have spatially-equal concentrations throughout a phase. By contrast, a non-mixed ("spatially resolved", in the paper) model models chemicals as having concentrations that can vary from location to location, making them more realistic but also far more expensive.
For anyone surprised that they only needed up to 1 GPU per simulation: their models appeared to be largely based in rate-kinetics; it's not like they were simulating atoms or really even molecules, but rather more like they were simulating concentrations of molecules.
To note it, they probably couldn't get a linear-speedup by using more GPU's. The thing's that they needed the GPU's to solve interactions over a space; if they had multiple GPU's, then they'd have needed to actively connect them on each iteration, slowing down the simulations. For a small number of GPU's (like maybe 2 to 4), they might try for a near-linear speedup if they fudge the boundaries a bit (allowing computational-artifacts at the edges between them).
This paper's got a ton of background research that's gathered up a lot of relevant data and models. The computational-results are nice, too, in providing a perspective on simulation-costs.
Once there're viable cell-simulations, we can do stuff like have computers predict, say, medicines by just simulating random molecules and optimizing them for desired-effect. A really powerful tool to look forward to!
Ah so it's sort of like CFD but with chemical interactions thrown into the mix?
Without specifically checking, I'd guess that they probably ignored some of the stuff involved in normal CFD, e.g. pressure-driven flows. In fact, their well-mixed model (their simpler model that didn't need a GPU) basically ignores fluid-dynamics entirely, since it ignores spatial-variation and so there're no fluid-dynamics to model. Their non-mixed model (their more complex model that did use a GPU) probably relied on diffusive-transport, without regard for stuff like pressure-driven flows.
So, yeah, CFD -- lighter on the mechanical-dynamics (like spatial-flows) and heavier on the chemical-dynamics (like chemical-reactions).
https://blogs.nvidia.com/blog/2022/01/20/living-cell-simulat...
But that was merely a note for "How this could have gotten on Nvidia's site" that isn't "Nvidia paid for it all."