4 ms·

"130,000 processor cores with 1 ns/day" Still a bit to slow... I guess it is progress :)

saalweachter 7y ago
This is the "hard part" with atom-level biological simulations. You need extremely tiny timesteps -- like, sub-picosecond, even -- but many interesting molecular events happen on the timescale of minutes.
krzat 7y ago
Damn I hope they can optimize this away eventually. Game of life can be sped up with hash tables.
- safsafsafa 7y ago
  Game of life doesn't have long range interactions, and doesn't have to deal with electrostatic interactions (which, with parallel B.C., diverge and cannot be calculated in real space).
dekhn 7y ago
when I wrote my thesis in 2001 I got 1ns/week on a much smaller system (1M atoms running on a T3E w/ ~256-512 processors). It's not really clear whether MD simulations like this are true contributions or whether they will ever produce useful results compared to well-parameterized neural nets.
- bocklund 7y ago
  Are NNs producing useful results in this area? I'm on the density functional theory side of things, so much smaller scales than MD. To me, it's not clear whether we are good at quantifying when a neural net is well-parameterized in atomistic simulations. The limitations of MD simulations and all the built in assumptions are well understood. We can reason about where our simulations might fail, even for very large systems. It's always a question about whether the underlying model can capture all the physics that occurs. With a NN or other black box model, there's no way to reason about it except for benchmarking test/validation sets that (hopefully) capture the physics that you care about, but you still cannot really reason about how well the model will extrapolate to multiple interacting physics with different magnitudes at different length scales.
  - dekhn 7y ago
    I don't think anybody has shown that NNs produce better predictions of protein models (yet). The limitations of MD aren't well understood. We don't know the implications of using a polarizable water model other than to say "it should be more accurate". All the statements you make about NNs apply to MDs since MDs are basically feedforward ASTs with the same level of complexity and non-linearity as NNs.

This is the "hard part" with atom-level biological simulations. You need extremely tiny timesteps -- like, sub-picosecond, even -- but many interesting molecular events happen on the timescale of minutes.

krzat 7y ago

Damn I hope they can optimize this away eventually. Game of life can be sped up with hash tables.

safsafsafa 7y ago
Game of life doesn't have long range interactions, and doesn't have to deal with electrostatic interactions (which, with parallel B.C., diverge and cannot be calculated in real space).

dekhn 7y ago

when I wrote my thesis in 2001 I got 1ns/week on a much smaller system (1M atoms running on a T3E w/ ~256-512 processors). It's not really clear whether MD simulations like this are true contributions or whether they will ever produce useful results compared to well-parameterized neural nets.

bocklund 7y ago
Are NNs producing useful results in this area? I'm on the density functional theory side of things, so much smaller scales than MD. To me, it's not clear whether we are good at quantifying when a neural net is well-parameterized in atomistic simulations. The limitations of MD simulations and all the built in assumptions are well understood. We can reason about where our simulations might fail, even for very large systems. It's always a question about whether the underlying model can capture all the physics that occurs. With a NN or other black box model, there's no way to reason about it except for benchmarking test/validation sets that (hopefully) capture the physics that you care about, but you still cannot really reason about how well the model will extrapolate to multiple interacting physics with different magnitudes at different length scales.
- dekhn 7y ago
  I don't think anybody has shown that NNs produce better predictions of protein models (yet). The limitations of MD aren't well understood. We don't know the implications of using a polarizable water model other than to say "it should be more accurate". All the statements you make about NNs apply to MDs since MDs are basically feedforward ASTs with the same level of complexity and non-linearity as NNs.