Molcas Forum

samarendra

Member

Registered: 2016-04-13

Posts: 1

Has anyone used MOLCAS on GPUs?

Hi,

We are trying to setup MOLCAS on our GPU-equipped machines and just had a test run. But it seems like the molcas binary was not using the GPUs at all!

So I guess we have missed some steps. I tried searching in the manual and on the Internet, but couldn't find any such information.

If you have setup and run MOLCAS to use GPUs (NVIDIA's specifically), please help us with any steps that differ for a GPU-based installation from a CPU-based one.

Thank you.

PSWe have used version 8.0 and compiled from the source using ./configure -setup

Last edited by samarendra (2016-04-13 10:38:57)

valera

Administrator

Registered: 2015-11-03

Posts: 127

Re: Has anyone used MOLCAS on GPUs?

Hi,
Molcas can use BLAS libraries from NVIDIA: NVBLAS (recommended!), and CuBLAS. Some basic information is here
http://www.molcas.org/documentation/man … 0000000000

The main problem with running Molcas on GPU is: the size of the matrices should be very large to see any advantage of GPU usage
http://www.molcas.org/VV/poster_Turku2.pdf
It is the case only for very large active spaces. But in this case the memory of GPU is another bottleneck...

BerndS

Member

Registered: 2015-11-06

Posts: 2

Re: Has anyone used MOLCAS on GPUs?

Hi,

just two remarks concerning CUDA:

1. There is nowadays no lib64 directory anymore
2. The wrapper in MOLCAS uses INTEGER*4 pointers. CUDA has changed to INTEGER*8 pointers for 64-bit architecture, so I had to change the code a bit.

Cheers,

    Bernd

Last edited by BerndS (2016-08-15 16:20:07)