[gmx-users] Computer processor and graphics card choice

[Guillem Prats Ejarque]

> Hi Guillem,

> I can answer your last question - the 10xx series does indeed work well
> with GROMACS (currently running GROMACS 2016 with a GTX1070), although I
> would be inclined to get the GTX1070 if your budget can stretch to it.

> Andrew

---

>> 80k atom MEM Benchmark System - 2fs time step - gmx2016.B2
>> GTX1070 - i7 6700k 4.00 GHz - 4 core - 42 ns/day - net cost per node (pc
>> chassis) ~ 950 EUR
>> GTX1070 - Xeon E5 2630v4 2.20 - 10 core - 58 ns/day - net cost per node (19"
>> rack) ~ 1600 EUR

>> GTX1080 will give you about the same performance but it is much more expensive, GTX1060 (6GB) is about 10% slower

> Note that this does not apply in general; here it is (most likely) the
> case because this specific simulation (with the GROMACS version used)
> is CPU-bound already with the 1070. If the GPU is "fast enough" for
> the workload, it does not help usign a faster GPU. Take another system
> (e.g. a CHARMM setup) and you might see significant speedup with the
> 1080.

> The 1060 being only marginally slower is also specific to the load
> balance in the MEM benchmark! With some other workload that's more
> GPU-bound on the 1060 + whichever CPU you were referring to, the 1070
> will give much larger simulation speedup.

> Certainly, if ns/day/buck is a major consideration (rather than high
> perf per node), the 1060 or 1070 are better value for the money!

> Cheers,

> Szilard


---


> Hi,

> CPU-wise, I think the 6700k, 6800k or 6850K would be better options.
> If you plan to have a second GPU, the 6700k and 6800k will not be
> ideal (too few PCI-E lanes to run both GPUs at full speed unless you
> get a special mobo). While the 2620v4 has 8 cores, it also has a low
> TDP, so it's unlikely those will run anything more than 2.5 GHz which
> won't be faster than six i7 cores at 3.3-3.5 GHz.

> If you don't need to buy right away, you might also consider the
> upcoming AMD CPUs which will be announced soon. Performance could end
> up being pretty good.

> The GTX 10xx series GPUs indeed work very well. Given the high clock
> the 1060 will certainly be faster than the 970. Andrew is right, the
> 1070 will give a further, significant bump in speed (>1.4x).

> (Note that these performance comparisons refer to the
> GPU-only/nonbonded kernel performance!)

> Cheers,

> Szilard


---

> Hi Guille,

> I'm not an expert on this, but it'll also depend on what kind of MD
> simulations you want to run. Coarse-grained forcefields, for example,
> don't benefit much from a GPU.

> So what kind of simulations do you have in mind? Atomistic? QM/MM? How
> many particles? ...?

> Peter


Thank you all for your valuable comments. Looking at that, the Xeon one really does not seem to be worth it. Regarding the other processors that you suggest, i7-6800K seems interesting to me. I do not plan to use more than one GPU.  About the graphic card, probably I will choose the GTX1060(6GB), because GTX1070 is getting out of my budget.


Mostly, what I want to simulate is a protein-ligand complex on an AMBER forcefield, and also some QM/MM simulations. My system is small, between 30 to 50 thousand atoms. I don't know if it is more CPU or GPU-bound.


Thanks again,


Guillem