Anton 3 is extremely hard-wired to doing electrostatics, whilst this work is much more about subtle quantum interactions – their big trick is to avoid having to do a calculation for every possible subset of four electrons, instead you do it for every subset of three electrons and can store the whole interactions for a thousand electrons in 40GB of GPU memory.

I’ve only ever used quantum chemistry tools as an input into a separate optimiser, modelling the ligand as a quantum object while doing the protein surrounding it with constraints much more like a ball-and-stick model, and where what we were trying to do is fit experimental X-ray diffraction data; I left the company fifteen years ago but it’s still there at the website in my link.

To me, the clincher comes with the results (in bold text) at the top-right of page 10: the computation of one time step takes “1.55 ZettaFLOPs on 9,400 nodes” of Frontier, and that takes 25.6 minutes of wall time (1,540 seconds) as the supermachine executes just North of 1 ExaFLOP per second … that’s what I call putting a supercomputer through its paces (bravo!)! And from this, and the needs of 1-km resolution climate modeling, it is becoming increasingly clear that we need to develop and qualify FP64 ZettaFlopping supermachines quick (let’s not rest on our ExaFlopping laurels!)!

One lingering question (for me, as a non-expert) is whether the specialized architecture of Anton 3 ( https://www.nextplatform.com/2023/12/04/the-bespoke-supercomputing-architecture-that-stood-test-of-time/ ) could help with such biologically-oriented AIMD computation, or if it is aimed at a slightly different conceptualization of the underlying processes, with a differing field of applicability? Either way, this work is very impressive imho, congratulations to the Barca (down under) team for winning this presitigous award, and let’s get ZettaFlopping soon, please!