Talk:Anton (computer)

Comparison to Folding@home

This material was placed on main page, but doesn't really fit there as written (it is argumentative in tone, and a digression from this topic). But there may be useful material to use on some "comparison of MD systems" article and/or in the Folding@home article:

However, the Folding@home distributed-computing project is substantially more powerful than Anton. Currently computing at about 6.5 petaFLOPS,

Folding@home performs protein folding molecular dynamics for disease research.

According to Vijay Pande, founder of the Folding@home project; "In many ways, I think they are still several years behind us. If you take a look at the proteins they simulated [in October 2010], it's the systems we first folded several years ago (eg villin). The systems we're publishing now fold 1000x slower than villin and they would not be able to study the folding of those systems with their current approach. With that said, it's really neat to see what dedicated hardware can do and they've done very nice work."

While a NatureNews article on Anton dated October 14, 2010 states that "Anton has simulated changes in a protein's three-dimensional structure over a period of a millisecond — a time-scale more than a hundred-fold greater than the previous record.",

Folding@home in January 2010 completed simulations of proteins in the 1.5-millisecond range.

Note: It appears that Folding@home achieves millisecond timescales by performing large ensembles of of relatively short trajectories as a means of sampling the space (as opposed to Anton's single long trajectories). Moreover, Folding@home uses an implicit solvent model which is computationally simpler than an explicit solvent. Both projects certainly have scientific value, but the timescales mentioned are not directly comparable from either a computational or chemical perspective. LotLE×talk 02:03, 21 September 2011 (UTC)

Actually, Folding@home uses a combination of both implicit and explicit molecular dynamics simulations. Folding@home runs on CPUs, GPUs, and PS3s. Implicit models are often performed by GPUs and also to some degree by the PS3s, while the CPUs do explicit. See Folding@home cores which is an article about the software that actually does the scientific simulations for Folding@home. F@h has achieved 1.5 millisecond timescales. Tiny simulations are necessary because F@h does distributed computing and Internet connections are usually slow. To take a well-cited quote from the F@h article: "Often how a protein misfolds can't be determined in a single simulation. Folding@home thus often runs thousands of slightly different simulations of the same protein, so as to capture the diversity and complexity.[20][38] Although some of these simulations can either work themselves into impossible atomic configurations or end up in the correctly folded shape, others will illustrate how that protein misfolds.[38]" Both are comparable; both Anton and F@h do molecular dynamics on proteins. The approaches are different, but they're really doing the same thing in the end. Jessemv (talk) 00:51, 23 September 2011 (UTC)

Thanks for the clarification, Jeessemv. FWIW, here is the specific long-trajectory performed on Anton:

David E. Shaw, Paul Maragakis, Kresten Lindorff-Larsen, Stefano Piana, Ron O. Dror, Michael P. Eastwood, Joseph A. Bank, John M. Jumper, John K. Salmon, Yibing Shan, and Willy Wriggers, "Atomic-Level Characterization of the Structural Dynamics of Proteins," Science, vol. 330, no. 6002, 2010, pp. 341–346.: A 1.013 millisecond simulation of the native-state dynamics of bovine pancreatic trypsin inhibitor (BPTI) (Figure 4A).

I happen to know that this trajectory was continued past the length in that publication, and now somewhat exceeeds 1.5ms. The scientific value of related ensembles vs. long trajectories is complex, and there is no single answer to which is "better." They each are able to reveal certain relevant features of the dynamics of complex macromolecules. 72.13.242.56 (talk) 19:54, 23 September 2011 (UTC)

Also of possible note: NTL9 is 2.3 kbp, while BPTI is 4.9 kbp. So the molecule simulated at similar scale on Anton is somewhat more than twice as many atoms as that done by Folding@home. LotLE×talk 19:59, 23 September 2011 (UTC)

Ah. Is there a paper supporting the 1.5-millisecond simulation time? Here is F@h's paper, and in there it states about the 1.5-ms timescale. Also, I'm no expert in protein molecular dynamics, but from my readings it seems that F@h performed many different simulations to each understand the different variables involved, where Anton performed just one. I get this from this quote in the F@h paper: "Trajectories were simulated via the Folding@Home distributed computing platform at 300, 330, 370, and 450K from native, extended, and random-coil configurations using an accelerated version of GROMACS written for GPU processors, for an aggregate time of 1.52 ms" whereas I see this in the Anton paper: "We used a 1-ms MD simulation at a temperature of 300 K to reproduce and interpret the kinetics of folded BPTI." Perhaps I missed something in all the technicalities discussed in the paper, but it seems to me that F@h had multiple 1.5-ms simulations completed at the same time, whereas Anton had one. F@h's was a 39-residue protein, whereas Anton's was a 58-residue one. I'm confident though both systems are making some serious progress in molecular dynamics. — Preceding unsigned comment added by Jessemv (talk • contribs) 01:46, 24 September 2011 (UTC)

Hi Jessemv: I am in a slightly sensitive position since I have worked with DESRES, and have some knowledge of Anton that is confidential. In the comments above I have been careful only to mention things available in publication. But without disclosing anything proprietary I can add a couple clarifying points: (1) The 1.5ms simulation by F@h indeed used an ensemble of states of its protein at different temperatures. The principle there is that at different temperature that protein may be able to reach different conformational states that a different energy level would preclude (or make less likely). The 1.5ms, nonetheless, is an aggregate length of all temperatures simulated, not 4 separate 1.5ms simulations at different temperatures--i.e. each temperature state might have been simulated for around 0.4ms (I don't know the distribution of temperatures used though). (2) In long-trajectory simulations such as those Anton performs, techniques are available to similarly explore the conformational space of a molecule. One such technique is "free-energy perturbation" and another is "simulated tempering" (also "replica exchange")--both of these are "alchemical" (more-or-less) in the sense that they alter the energy state of a system during a trajectory to try to get it to reach states it might not otherwise reach (or that might take excessively long timescales to reach). As a physical analogy, this might be as if you heated and cooled the test tube every microsecond to "jigger" the folding and unfolding of your protein (which may or may not be possible by any known physical techniques, but in concept anyway). Not every simulation is best served by such "jiggering" though, it really depends on the molecule involved and the goal of the particular research. LotLE×talk 19:38, 24 September 2011 (UTC)

Ah. I did not realize that there were confidential aspects to Anton. I do thank you for clarifying simulated tempering. It appears that both F@h and Anton do this then. In Folding@home, these kinds of calculations are usually done on PlayStation 3s, but other types of calculations are also done to explore the protein's conformational space. The Pande Group (who run Folding@home) construct Markov State Models which I believe help with this as well. See here, here, and this project which I was recently working on on my graphics card. My position is one of just a F@h contributor. I donate CPU and GPU power to the project, but I know nothing beyond what is publicly available, which is just about everything about F@h save only the source code to their middleware clients, since they don't want people trying to cheat or run it in very unsupported ways. However, the software used by Folding@home for their calculations is publicly available, as is all results results from them. I find all these approaches to disease research very fascinating! Jessemv (talk) 23:36, 25 September 2011 (UTC)

Claiming that Folding@Home is equivalent in performance to Anton because they both have run the same aggregate chemical time is like saying that the Ford Taurus is equivalent to the Space Shuttle because they've both traveled the same number of aggregate miles. Anton can run a single trajectory very fast, which is quite useful for finding pathways between states. Highly parallel statistical things like Folding@Home can do some high throughput exploration of the energy landscape around some known state--it is of limited utility in identifying pathways. They are both useful tools. Also, the comparison of aggregate chemical time is flawed because I believe you are comparing all of Folding@Home's chemical time against one hero run on one Anton machine taking on the order of a couple months. There are *many* Anton machines which have been running for at least a couple years now.

Regardless, why is there all this discussion of Folding@Home on the Anton page? Should I head over to the Folding@Home page and litter it with comparisons to Anton? — Preceding unsigned comment added by 74.73.228.91 (talk) 04:46, 20 October 2011 (UTC)

Hi, I don't believe I've said that Anton and F@h are "equivalent". They are not, and they have very different approaches. Like you said, Anton does one thing very fast, while F@h uses other approaches that are very parallelizable, but they both try to get to the understanding of how proteins fold. The content that is currently in F@h is backed up by valid sources, in this case Dr. Vijay Pande. If you read the source I believe it indicates that pathways are explored pretty well, and further information this can be found here although there are many other places as well. I'm not a biochem major; I don't understand some of the technical details of molecular dynamics so please forgive my ignorance. Above I made a comparison statement and was corrected, so I'm aware that it was indeed flawed. However, I very much trust Dr. Pande's comparison, as with all his credentials I'm pretty sure he knows what he's talking about, and like I said his post backs up the comparison to F@h in the article. I'm sorry that you don't think that the content should be there. It is Wikipedia; you're welcome to change it or something. However, if you look at Rosetta@home's page I see that they have comparisons to other projects, something that is obviously important since the article ought to discuss the aspects of its subject. Perhaps it has the wrong tone or something, your welcome to fix that if you want. But everything it says is backed up, so its not Original Research. Perhaps we should add more Anton content so that the comparison takes up a smaller fraction of the content. Also, I have already including comparisons to Anton on the Folding@home page here which also has some summaries as to how effective F@h's statistical approaches are. I have no doubt that Anton will make some really great accomplishments and Dr. Pande has also praised Anton's hardware and work. Sincerely, Jessemv (talk) 05:20, 20 October 2011 (UTC)

I just finished some attempts at better neutrality in that section. A lot of information from the source was removed, but perhaps it is better. I'm actually looking forward to reading more about Anton in the article. Jessemv (talk) 05:38, 20 October 2011 (UTC)

If you read this fascinating scientific paper, published in Current Opinion in Structural Biology in 2010 by members of F@h's Pande Group, below Figure 1 it discusses the difference between single long timescales and the kind of statistical work F@h is doing. I found the paper really interesting, and surprisingly much of it is pretty easy to understand. I'm sure that Anton's researchers are aware of the details of protein folding, but this publication has some insights and I thought that I should mention it. Jessemv (talk) 03:42, 21 October 2011 (UTC)

I would like to issue a general apology to all parties that I may have offended with the non-neutral content. It was wrong of me to place this content in the way that it was worded on the Anton page. This has been pointed out to me, both by Wikipedians and anonymous IPs. I should not have made a comparison between Anton and F@h without looking at both sides of the issue. On the Folding@home page, the neutrality of the comparison to Anton was recently called into question which I have addressed. My research for this fix has involved reading of both F@h and Anton publications, from which I have since learned how powerful and useful Anton really is. In light of this discovery, I realize now that it was definitely not appropriate for that biased comparison to be in the page. I thus detract it. Anton and F@h should be compared in a neutral way, with facts backed up by scientific publications instead of blogs and forum posts. Such comparisons should not "cast a shadow" over Anton, for it is a very useful tool which like F@h will continue to provide unique insights into the complexities of protein folding. Humbled, Jessemv (talk) 02:24, 9 November 2011 (UTC)

Comparison to Other Molecular Dynamics Systems

While I don't disagree that is is interesting to see how F@h and Anton differ in their approaches to computation and aggregate ensembles, this addition basically turns this article on Anton into one about Folding@home, which is really WP:UNDUE, and generally off-topic. I put the material here on the talk page, but it really needs a home somewhere different to be appropriate. LotLE×talk 00:37, 28 October 2011 (UTC)

Currently Anton and the distributed computing project Folding@home are the two most powerful molecular dynamics systems. Folding@home currently runs at 6.4 petaFLOPS from large host of CPUs, GPUs, and PS3s, and has achieved aggregate ensemble simulation timescales similar to those Anton has, specifically achieving the 1.5 millisecond range in January 2010.

Anton and Folding@home simulated and study protein folding very differently. Anton simulates a few long trajectories, while Folding@home takes a more paralleled approach by statistically combining many shorter simulations. Folding@home's approach can accurately simulate timescales equal to or greater than Anton's, and can reproduce Anton's long simulations very well. However, in 2011 it was found that F@h's Markov State Models can find important features missing in Anton's traditional analysis. Dr. Vijay Pande, director of the Folding@home project, looks forward to see how Anton and F@h can be used together, noting that a long-term simulation on Anton followed by more thorough sampling in Folding@home would be very beneficial.

Jessmv: I think what would be more useful that losing focus within this article would be to have a general article called Comparison of molecular dynamics systems that could encompass Folding@Home, Anton, and other software (or hardware) systems as well. This could allow for comparisons of a variety of dimensions, and I think the existing articles and categories provide some guidance. For example:

• Long trajectories vs. aggregate ensemble techniques
• Explicit vs. implicit solvent
• Types of hardware supported and required
• Force fields implemented within a system
• Longest trajectories attained by particular systems
• Notable simulations with published results from various systems

I'm sure there are other dimensions too, but those are a few that occur to me. It might make sense to present much of this comparison in one or more columns, although citations are obviously important, and incorporating more descriptive text within the envisioned article would help also. LotLE×talk 01:29, 28 October 2011 (UTC)

Hmm. I'll think about that. Once I bring Folding@home up to Good Article status I might make that page. We'll see. That's an interesting idea actually. Thanks. Jessemv (talk) 01:37, 28 October 2011 (UTC)

No, at this point I've changed my mind. I don't think I'll do that. For one, I'm not a biochem major, so I can get lost in some of the technical molecular detail, and an article like that would intrinsically involve a lot of that. Two, I've already fought the "this page has been flagged for deletion" battle, and though I won I'm not enthusiastic about doing it again. Three, with my work on the Folding@home article, and my participation and interest in the project, I may not be able to stay neutral. The information for such an article can likely be found in the Wikipedia articles of the MD systems involved, so I'll let someone else pull it all together. It would be a fascinating article, and one that would be worthy of being in an encyclopedia, but I just don't think I'm qualified to create it, that's all. Jessemv (talk) 02:12, 9 November 2011 (UTC)

References

Cost

So I've read through most of the references and some of the paywalled ones, and nowhere did I see an answer to these questions: how much does an Anton cost to build, and then how much does it cost to run? I can make guesses at both based on ASIC costs and power consumption, but I'd rather not. Does anyone know? --Gwern (contribs) 01:35 20 April 2012 (GMT)

I've sent an email to one of the addresses on their site asking for build & run costs. --Gwern (contribs) 16:51 18 May 2012 (GMT)

I'd also be interested in seeing those numbers. Hopefully you can also find a reliable source somewhere that you can use as a citation to whatever figures they give you by email. Best, Jesse V. (talk) 17:11, 18 May 2012 (UTC)

Useful citation material

Anyone who is interested in expanding this article might be interested in the paper "Anton, a special-purpose machine for molecular dynamics simulation", written in 2007. this might be a publicly-working link to a PDF. • Jesse V.^(talk) 14:36, 15 September 2012 (UTC)

Help

how to download trajectories computed by anton? I saw many articles said they download the trajectories, but i can't find the link to the trajectories on the publication page. — Preceding unsigned comment added by 108.236.198.181 (talk) 03:19, 21 May 2013 (UTC)

There is an arithmetic error

It says right there that "Folding@home has achieved aggregate ensemble simulation timescales similar to those Anton has, specifically achieving the 1.5 millisecond range in January 2010". This is incorrect because 1,5 millisecond is not similar to 17.000 nanoseconds = 17 microseconds. 1,5 milliseconds is 3 orders of magnitude ( 1000 times ) bigger than 17 microseconds ( 17.000 nanoseconds ).

cleanup/merge

I'd like to pull Unforgettableid in to this discussion in particular, since he's raised concerns about notability and autobiography. The involvement of Anxdo is a fair point, but it's also unclear how to address the concerns raised by your flag. The content itself is largely neutral and focused on factual details, and my understanding is that's ultimately what matters. Are there specific passages you're concerned about?

As far as notability, the point about sources is fair (although I've seen worse and these are reputable journals), but it strikes me as more of a "needs citations for verification" situation than a general notability one.

On a related note, perhaps Anton (computer) should be merged into D. E. Shaw Research, as the notability of the two are tied to each other. That could also make NPOV sourcing easier to address. Desmond (software) should probably stay separate as it's more widely used. Proxyma (talk) 04:37, 22 November 2013 (UTC)

Disagree about merging Anton (computer) and DEShawRes. There is one Anton with nominally public access so there is some validity is seperating them (as in parent/child) cowburn —Preceding undated comment added 21:55, 2 November 2015 (UTC)

Re: Anxdo, this user hasn't been active since 2011. Looks like a grad student that was not directly involved in the company. That is not a current concern and the content of the article looks fine, imo. A quick search shows recognition by a reputable scientific journal at Supercomputer sets protein-folding record supporting notability. I would support replacing the {{coi}} with something that more specifically addresses the issues perceived to be lacking here. I don't see any compelling reason to merge. --mikeu ^talk 06:56, 28 December 2015 (UTC)

Link mismatch in cited material

The Link to the pdf given here is incorrect. It should refer to the Anton 2 paper refered to by the DOI 10.1109/SC.2014.9 but the pdf seems to be describing Anton 1 and has different authors. I am removing the direct link to the pdf. — Preceding unsigned comment added by Lpd-Lbr (talk • contribs) 10:53, 9 May 2019 (UTC)

References

Resolution

This topic now seems sufficiently neutral and non-partisan. It contains interesting information about the device. I will remove the warning in a few days if there are no objections. — Preceding unsigned comment added by Cowburn (talk • contribs) 14:10, 13 August 2020 (UTC)

Anton 3 ASIC to be described at Hot Chips 2021

DE Shaw has a talk in the upcoming Hot Chips conference, concerning the 3rd version of their ASIC. The talk is listed in the advance program summary on the conference website, no login required. For now there is only a title with authors.

Hot Chips 2021 - August 22-24, 2021

On-Line - A symposium on High-Performance Chips

(Conference day 2, 11:30)

The Anton 3 ASIC: a Fire-Breathing Monster for Molecular Dynamics Simulations, J. Adam Butts and David E. Shaw, D.E. Shaw Research

Robert Munafo (talk) 14:55, 14 May 2021 (UTC)

References

Anton 3 Described and In-Service

Anton 3 has been placed in service since this article last received content updates.[1] The newest version apparently shows significant capacity upgrades and the article is likely due for an update. 2603:6011:5B02:3CE6:A52B:D954:D18D:E3B1 (talk) 01:29, 2 August 2023 (UTC)