Hi.
That sounds very interesting. Let me try to answer your questions:
1) ATK uses Intel MKL for matrix operations. Provided that the CRAY XC series offer MKL, the ATK software should work well on such a system.
2) The ATK computation time vs. number of CPUs depends very much on the type of system you are considering. For bulk calculations with DFT we parallelize efficiently over k-points and more, but at some point the Hartree potential will start hurting you. For device calculations, we take advantage of sparse methods and k-point parallelization to drive down the time per SCF step. Obviously, device calculations with semi-empirical methods will scale differently than with DFT.
3) The number of atoms than can be computed with 8 GB available RAM will depend a lot on a) the physical system (which elements? how big a supercell? etc.), and b) the DFT method (how many basis functions per atom? should the calculation be spin-polarized, noncolinear, or include spin-orbit interactions? etc.). The scaling is not gonna be linear. It will most likely be quadratic wrt. total number of basis functions in the calculations. The curse of matrix algebra
What sort of systems are you thinking of running? It would be easier to estimate the scaling you will get if you could provide an example of a calculation (in private if support@quantumwise.com is used).
Best,
Jess
