Hi Lili,

Jose is correct, TeraChem is likely the best bet if you want GPU acceleration, however speed alone shouldn't dictate what program you use. One major concern is cost: TeraChem, Gaussian, Vasp and QChem are not free and your institution likely has a subscription already to one of these. Another thing you should be concerned about is if your chosen program has the functionals implemented that you would like to study, if you're just doing B3LYP or something of that nature you won't have to do much research but if you're looking for a more exotic functional you should make sure it's there first before purchasing.

Additionally, there are other ways to speed up a DFT calculation than GPU acceleration. I don't know much about TeraChem, but I do know Orca is a free computational chemistry program with methods implemented to speed up DFT and post-HF calculations. The resolution of identity approximation for both Coulomb and exchange integrals in ORCA significantly speeds up DFT calculations with little to no effect on the accuracy, even though it's not implemented on GPUs. Additionally, in terms of post-HF methods DLPNO, F12, and RI methods are available for MP2, CCSD(T), etc that work to speed up and increase the accuracy of results with smaller basis sets (in the case of F12). These methods may be on TeraChem, I don't know but if they are they would be great with GPUs, but I just wanted to make you aware of other considerations that could help with your decision.

TeraChem run totally in GPUs, some of the others programs runs partially in GPUs, so TeraChem is faster ...

Hi Lili,

Jose is correct, TeraChem is likely the best bet if you want GPU acceleration, however speed alone shouldn't dictate what program you use. One major concern is cost: TeraChem, Gaussian, Vasp and QChem are not free and your institution likely has a subscription already to one of these. Another thing you should be concerned about is if your chosen program has the functionals implemented that you would like to study, if you're just doing B3LYP or something of that nature you won't have to do much research but if you're looking for a more exotic functional you should make sure it's there first before purchasing.

Additionally, there are other ways to speed up a DFT calculation than GPU acceleration. I don't know much about TeraChem, but I do know Orca is a free computational chemistry program with methods implemented to speed up DFT and post-HF calculations. The resolution of identity approximation for both Coulomb and exchange integrals in ORCA significantly speeds up DFT calculations with little to no effect on the accuracy, even though it's not implemented on GPUs. Additionally, in terms of post-HF methods DLPNO, F12, and RI methods are available for MP2, CCSD(T), etc that work to speed up and increase the accuracy of results with smaller basis sets (in the case of F12). These methods may be on TeraChem, I don't know but if they are they would be great with GPUs, but I just wanted to make you aware of other considerations that could help with your decision.

I think GPU

I think that if you give more information about the system that you are going to study (not just reactions) we can help you with better suggestions. I agree with Daniel Beckett.

Note that even if a code can use GPUs, it might not be considerably faster. Many codes advertise their GPU support with speedups that are barely over 1, still less than 2. This makes little sense, since you will have to shell out 1000€ - 2000€ for every single GPU, and for that price you can just buy more nodes. Proper GPU codes get orders of magnitude speedups.

Like Daniel said above, there are plenty of codes that can do what you ask for very quickly. There are also several that you can get for free, like Orca (free for academic use), and PySCF and Psi4 that are free software. You can also find codes for periodic calculations.

I like DFTB , ADF and quantum expresso.both DFTB and quantum expresso are faster and give similar to the ADF.

I agree with Daniel Beckett and Susi Lehtola answers. More information about the system will also be helpful, as some programs will not work for some applications (periodic boundary conditions, unusual functional).

Also good idea is to first estimate things with lower level method, for this Tight Binding is perfect. The DFTB+ program is excellent, specially for periodic systems. Grimmes XTB has very nice conformational searching capability and supports all atoms.

If your system is periodic things change: Some programs have superficial support for periodic systems or none at all (ORCA), in others you control almost everything (FHI-AIMS, Crystal17, or DFTB+ for Tight Binding approximation).

I did some test with TeraChem and the results depend on the system, sometime ORCA calculations could be faster, depending on the system and basis chosen (ORCA prefers Ahlrichs basis sets and could struggle with Pople ones). In general ORCAs big advantage is also really good algorithms for SCF itself - it could half the number of iterations needed for convergence. Sometimes TeraChem was faster.

If you need to use GPU VASP and TeraChem seem as a good options. Just bear in mind that it will be more expensive (program cost + decent GPU) and may be problematic (e.g. compilation with GPU support, not enough GPU memory).

I prefer ADF-DFTB code for larger molecules using DFT. It can calculate hundreds of atom within minutes.

which car is faster? It depends on how much do you have in your pocket, for what you want to use it, and how large your garage is.

Thank you very much!

Lili

In my experience there are not so much difference in the speed of DFT calculations with CP2K, Quantum Espresso (QE) and VASP (it mostly depends on the basis set and pseudo potentials you choose) but if I want to choose between all quantum chemistry calculation packages I would prefer CP2K for DFT calculations. It uses a special package for sparse matrix multiplication (DBCSR) which efficiently performs multiplications, is mpi and openmpi parallel and also supports CUDA (but does not increase the speed so much). It is a little bit faster than QE in doing AIMD (using auxiliary density matrix method) and geometry optimization. However, it has less features compared to QE but interfaces with many other packages which can perform other types of calculations. Its installation is a bit hard compared to other types of packages.

Nevertheless, you can increase the speed of all these packages by compiling them with ELPA, LIBINT, LIBXC and other types of compilers such as intel parallel studio.

DFTB+ version 19.1 (released on 01/07/2019) has incorporated GPU acceleration using MAGMA library and is very fast.