Estimation of the number of needed GPUs

W-SLDA Toolkit provides a script that can be used to estimate the number of GPUs that you need to run your code efficiently: tools/td-memory.py. The user must edit the # SETTINGS section and run the code. Example:

# SETTINGS
NX = 128
NY = 128
NZ = 16
codedim=2 # dimensonality of code
nwf=70141 # provide here number if you know it, otherwise the code will use simple estimate
mem_per_gpu = 16.0 # in GB
min_mem_utilization = 2.0 # in GB

Note that the number of wave-functions to be evolved is typically printed by st-wslda code when writing them to files. Optionally you can leave nwf=None, and then the script will use an estimate for this number.
Running the script:

[gabrielw@wutdell tools]$ python td-memory.py 
MINIMAL NUMBER OF GPUs=24

and plot like this will show up:

To obtain a good performance of the code, it is recommended that the memory utilization of each GPU card is about 50% or more of its capacity. In the given example, it is recommended to run the code with the number of GPUs less than 50.