how many atoms atk can deal with

[njuxyh]

Hi all:

i face a big device, the number of atoms in the two electrode cell and the central region are about  2000, the electrode is cu.

and i want to optimize it first and then calculate the transmission spectrum,
So I am wondering whether atk could deal with so big system?

if could, try the DFT or semi-classic method in atk???   

anyone could help me ??

thanks  very much!!

[Anders Blom]

2000 atoms is a lot, esp. for a d-metal. The question is, however, do you really need to optimize the whole structure? If it's pure Cu you could optimize very quickly using a classical potential like the EMT available on ATK.

Then to run it, you may try DFTB rather than DFT. In general, ATK 2014 will include several improvements in terms of memory usage and speed that make this size of systems more realistic to handle. Still, you will need some serious computer resources.

[njuxyh]

Hi  Prof. Anders Blom:
i just want to relax the central region of such a device, and the central region consists of GNRs.

Another question is if i do not relax, only calculate the transmission spectra in atk, atk can deal with?

[Luis M. Villamagua C.]

Hi njuxyh,

Density-function-theory (DFT) based models are computationally  prohibitive for systems having more than about 200 atoms [1].

Depending on the selected method, calculations can be performed on structures with several hundred (DFT/DFTB and Hückel) to tens of thousands (tight-binding) of atoms [2].

[1] H. Raza e E. C. Kan, «An extended Hückel theory based atomistic model for graphene nanoelectronics», J Comput Electron, vol. 7, n. 3, pagg. 372–375, set. 2008.
[2] «Graphene studies with ATK», QuantumWise. [online]. Available at: http://quantumwise.com/applications/graphene/16-graphene-studies-with-atk.

Hope it helps
Luis

[Anders Blom]

Good references, Luis. But the story is not always trivial - actually the number of atoms is not the main problem, it's the number of electrons, and not even that: what matters is the number of orbitals. So 100 d-metal atoms is a very different story than 100 Na atoms. Also, ATK is generally capable of handling more atoms in a device configuration than for a bulk system. This is because in a device, there is no diagonalization (except for the equivalent bulk part, but that can be turned off). So it's certainly possible to do a "long device" with >500 carbon atoms, even with DFT (at least with smaller basis set).

Besides, today it's not uncommon to have 32-64 GB of RAM per node in a cluster, and that's a very different thing than trying to run on a laptop with 2-4 GB. So, in the end, all depends on a combination of many different factors. Thus one can not give any number to say how many atoms you can run. I have for instance done a selfconsistent bulk calculation of 1000 atoms on my laptop, using a small basis set in DFT. And we recently tried 10,000 atoms in tight-binding, it takes 1.5 minutes (non-selfconsistent) in ATK 2014!

[njuxyh]

yes,  it should do an estimate for the memory will be enough or not.

is there a formula to estimate how the memory will be use?

ps: the ATK 2014 is now available version 13.8.1 or a newer version??

[Anders Blom]

No, but you can use the tool described here:
http://quantumwise.com/documents/tutorials/latest/MemoryUsage

ATK 2014 is almost ready for a first public preview. I hope during next week - but that's a goal, not a promise.

[njuxyh]

thanks very much,

i have learned about the methods to reduce the memory usage.

and i am looking forward to the newer 2014 version.

[njuxyh]

No, but you can use the tool described here:
http://quantumwise.com/documents/tutorials/latest/MemoryUsage

ATK 2014 is almost ready for a first public preview. I hope during next week - but that's a goal, not a promise.

Hi Prof. Anders Blom:
when i want to relax the device, the memory usage is taken into account with this method??

thanks very much!!

[Anders Blom]

The relaxation in itself does not increase the memory requirement, at least not by any significant amount. So whatever memory you need for a "usual" calculation for a particular calculation, you be need a similar amount for the relaxation.

[njuxyh]

OK, i got it ,

thanks a a lot !

[njuxyh]

No, but you can use the tool described here:
http://quantumwise.com/documents/tutorials/latest/MemoryUsage

ATK 2014 is almost ready for a first public preview. I hope during next week - but that's a goal, not a promise.

Hi Prof. Anders Blom:

following the previous questions about  the  memory usage estimation. the method  in the linking manual is only for a serial calculation(one cpu),
if i calculate with  multi-cpu, then the total estimated memory should multiply*cpu number?

for example, if the estimated memory usage is 11G with the method in the linking manual. but in a practical calculation, i use 16 cpus in a nodes for a parallel calculation, then i should make sure the node has 11G*16 memory available?
is that so ?

thank you very much!

 

[Anders Blom]

Yes, if you have all 16 MPI processes on the same node, this is the case. So for a big calculation, this limits the amount of MPI you can put on a node. However, you can perhaps reduce the memory requirement by e.g. not saving the self-energies? And then try the sparse recursion methods in ATK 2014, although the "Memory estimate" will still show 11 GB since it can't predict the sparsity pattern.

[njuxyh]

thanks very much!

 i have learn about this issues a lot.