6 ms·

It's pretty crazy just how computationally intense biological simulations are. The timesteps are just so tiny and the number of nodes involved is insane...

3np 2y ago
That's life.
FerretFred 2y ago
Yeah, I found this when I was Boinc protein folding for Covid. The work units ran for hours, my towercPC nearly melted, and then I found that this produced mere milliseconds of simulation!
- xattt 2y ago
  > nearly melted I’ll be that guy, but a computationally complex problem won’t push your computer to a temperature beyond design limits.
  - FerretFred 2y ago
    You're welcome to come to my attic office in summer, as they ambient temperature reaches 200F m a sunny day and my obsolete-but-powerful tower PC consumes 350W power and adds to the tropical heat :)
  - LoganDark 2y ago
    > a computationally complex problem won’t push your computer to a temperature beyond design limits. Allegedly. It allegedly won't do that. Practically, running your CPU at design limits for a very long period of time tends to cause the temperature of the rest of the chassis to want to equalize to that temperature, which can be above comfort limits.
  - inglisilihsh 2y ago
    https://en.wikipedia.org/wiki/Hyperbole#:~:text=Hyperbole%20is%20often%20used%20for,or%20ironic%20or%20comic%20effects. https://en.wikipedia.org/wiki/Hyperbole#:~:text=Hyperbole%20...
  - JohnHaugeland 2y ago
    It’s just a turn of the phrase. When someone says that the sky is falling, that also doesn’t need to be explained
  - userbinator 2y ago
    It may if components are marginal... https://news.ycombinator.com/item?id=43008879 https://news.ycombinator.com/item?id=43008879
- jxjnskkzxxhx 2y ago
  Now consider that the Planck time is actually 1e-44 secs.
  - atemerev 2y ago
    Hardly any actual physical simulations simulate every possible moment in time. We just calculate consequences of current events and put them in the queue to happen sometime in the future. And there's evidence that the real world works in a similar way.
  - dekhn 2y ago
    Bonds jiggle and wiggle with characteristic times in femtoseconds, and most people believe it's not necessary to simulate at higher frequencies than that.
    - jxjnskkzxxhx 2y ago
      Yes, I agree with that. My point was just that stimulating a rudimentary model of reality is hard work, but reality is way more complex than even that.
monero-xmr 2y ago
My (religious, spiritual, please-don’t-downvote-because-of-hate) opinion is the aspect of free will which extends even to the lowest creatures means computers will never simulate them accurately
- h0l0cube 2y ago
  No downvotes from me, but where do you think the free will lives? What mechanism produces it? What about that mechanism, if not purely stochastic, is not predictable? If free will can have a real measurable physical effect on this universe, then why can't it be isolated?
  - khimaros 2y ago
    a random number generator in base reality? ;) really i'm just using this as a more accessible (to rational materialists) example of how processes can exist outside of our measurable physical reality.
    - h0l0cube 2y ago
      > processes can exist outside of our measurable physical reality If it can't be measured in reality, how can it be relevant to reality? > a random number generator in base reality? How does randomness pertain to free will?
      - khimaros 2y ago
        they may have measurable effects in our physical reality, eg. the behavior of the individual and even be correlated with other physical processes. however, there may be hidden variables outside of our physical reality that are actually mechanizing the result. some such processes may be non-deterministic, which is why i used randomness as an example. what i'm implying is analogous but opposite to the concept of a philosophical zombie. there may be a ghost in the machine which no measurement can reveal.
        2y ago
jboggan 2y ago
I think it's because we learn biology with a textbook cartoon illustration of the cell with sparse organelles floating in a clear sea of cytoplasm. The packed reality looks more like the halftime bathroom line at the Rose Bowl. Tons of heterogenous proteins and metabolites and mRNA and everything seemingly diffusing to its destination, shouldering past everything else. The sheer combinatorial complexity of the number of neighbors each molecule is imparting force on, combined with the constantly shifting conformational changes of many proteins which change the forces they receive, it's overwhelming. Recall that only recently AlphaFold was able to decently solve protein structures in isolation - add in dozens of shifting neighbors and that's suddenly a dynamic problem. I am not sure where whole-cell simulation is at the moment since I've been away from the field for about 15 years, but I recall a rather difficult multi-month simulation that was trying to model an "empty" volume of cytoplasm away from the organelles, about 1/50th of total cell volume, and with all proteins and metabolites replaced by hard spheres of varying "stickiness". It was considered a huge success to just get a few of the diffusion rates for various compounds in the right order of magnitude. I mean, if you really want to get the fleeting interactions right you need to be modeling individual water molecules. I know there have been large advances in computing in the intervening years, but this was on I think #20 in the TOP500 at the time. Unlike AlphaGo I don't see any immediate avenue for AI to help with this because unlike protein crystal structures there is no wealth of quality training data for cellular dynamics at the molecular level.
- smj-edison 2y ago
  Great point on the chaos! I'm planning to pursue atomic simulation in college, so I've tried to read as much as I can about the field short of the advanced math (hence college, lol). It seems to me that cross cluster synchronization is a massive scaling issue, since you essentially have global state both reduced and broadcasted every couple timesteps. I've been thinking it would be cool to design chips to be realtime safe—that way there's no need to synctronize—and have further away information delayed (just like relativity) to deal with speed of light communication issues. Never heard of fleeting interactions, would you mind to elaborate?
  - jyounker 2y ago
mkoubaa 2y ago
Except this isn't computational biology, it's fluid dynamics with an agent inside running a neural network. You'd think they'd try something like smooth particle hydrodynamics so that it can be done more efficiently on a GPU cluster but I think they are just doing classic CFD

3np 2y ago

That's life.

FerretFred 2y ago

Yeah, I found this when I was Boinc protein folding for Covid. The work units ran for hours, my towercPC nearly melted, and then I found that this produced mere milliseconds of simulation!

xattt 2y ago
> nearly melted I’ll be that guy, but a computationally complex problem won’t push your computer to a temperature beyond design limits.
- FerretFred 2y ago
  You're welcome to come to my attic office in summer, as they ambient temperature reaches 200F m a sunny day and my obsolete-but-powerful tower PC consumes 350W power and adds to the tropical heat :)
- LoganDark 2y ago
  > a computationally complex problem won’t push your computer to a temperature beyond design limits. Allegedly. It allegedly won't do that. Practically, running your CPU at design limits for a very long period of time tends to cause the temperature of the rest of the chassis to want to equalize to that temperature, which can be above comfort limits.
- inglisilihsh 2y ago
  https://en.wikipedia.org/wiki/Hyperbole#:~:text=Hyperbole%20is%20often%20used%20for,or%20ironic%20or%20comic%20effects. https://en.wikipedia.org/wiki/Hyperbole#:~:text=Hyperbole%20...
- JohnHaugeland 2y ago
  It’s just a turn of the phrase. When someone says that the sky is falling, that also doesn’t need to be explained
- userbinator 2y ago
  It may if components are marginal... https://news.ycombinator.com/item?id=43008879 https://news.ycombinator.com/item?id=43008879
jxjnskkzxxhx 2y ago
Now consider that the Planck time is actually 1e-44 secs.
- atemerev 2y ago
  Hardly any actual physical simulations simulate every possible moment in time. We just calculate consequences of current events and put them in the queue to happen sometime in the future. And there's evidence that the real world works in a similar way.
- dekhn 2y ago
  Bonds jiggle and wiggle with characteristic times in femtoseconds, and most people believe it's not necessary to simulate at higher frequencies than that.
  - jxjnskkzxxhx 2y ago
    Yes, I agree with that. My point was just that stimulating a rudimentary model of reality is hard work, but reality is way more complex than even that.

monero-xmr 2y ago

My (religious, spiritual, please-don’t-downvote-because-of-hate) opinion is the aspect of free will which extends even to the lowest creatures means computers will never simulate them accurately

h0l0cube 2y ago
No downvotes from me, but where do you think the free will lives? What mechanism produces it? What about that mechanism, if not purely stochastic, is not predictable? If free will can have a real measurable physical effect on this universe, then why can't it be isolated?
- khimaros 2y ago
  a random number generator in base reality? ;) really i'm just using this as a more accessible (to rational materialists) example of how processes can exist outside of our measurable physical reality.
  - h0l0cube 2y ago
    > processes can exist outside of our measurable physical reality If it can't be measured in reality, how can it be relevant to reality? > a random number generator in base reality? How does randomness pertain to free will?
    - khimaros 2y ago
      they may have measurable effects in our physical reality, eg. the behavior of the individual and even be correlated with other physical processes. however, there may be hidden variables outside of our physical reality that are actually mechanizing the result. some such processes may be non-deterministic, which is why i used randomness as an example. what i'm implying is analogous but opposite to the concept of a philosophical zombie. there may be a ghost in the machine which no measurement can reveal.
      - 2y ago

jboggan 2y ago

I think it's because we learn biology with a textbook cartoon illustration of the cell with sparse organelles floating in a clear sea of cytoplasm. The packed reality looks more like the halftime bathroom line at the Rose Bowl. Tons of heterogenous proteins and metabolites and mRNA and everything seemingly diffusing to its destination, shouldering past everything else. The sheer combinatorial complexity of the number of neighbors each molecule is imparting force on, combined with the constantly shifting conformational changes of many proteins which change the forces they receive, it's overwhelming. Recall that only recently AlphaFold was able to decently solve protein structures in isolation - add in dozens of shifting neighbors and that's suddenly a dynamic problem. I am not sure where whole-cell simulation is at the moment since I've been away from the field for about 15 years, but I recall a rather difficult multi-month simulation that was trying to model an "empty" volume of cytoplasm away from the organelles, about 1/50th of total cell volume, and with all proteins and metabolites replaced by hard spheres of varying "stickiness". It was considered a huge success to just get a few of the diffusion rates for various compounds in the right order of magnitude. I mean, if you really want to get the fleeting interactions right you need to be modeling individual water molecules. I know there have been large advances in computing in the intervening years, but this was on I think #20 in the TOP500 at the time. Unlike AlphaGo I don't see any immediate avenue for AI to help with this because unlike protein crystal structures there is no wealth of quality training data for cellular dynamics at the molecular level.

smj-edison 2y ago
Great point on the chaos! I'm planning to pursue atomic simulation in college, so I've tried to read as much as I can about the field short of the advanced math (hence college, lol). It seems to me that cross cluster synchronization is a massive scaling issue, since you essentially have global state both reduced and broadcasted every couple timesteps. I've been thinking it would be cool to design chips to be realtime safe—that way there's no need to synctronize—and have further away information delayed (just like relativity) to deal with speed of light communication issues. Never heard of fleeting interactions, would you mind to elaborate?
- jyounker 2y ago

mkoubaa 2y ago

Except this isn't computational biology, it's fluid dynamics with an agent inside running a neural network. You'd think they'd try something like smooth particle hydrodynamics so that it can be done more efficiently on a GPU cluster but I think they are just doing classic CFD