Task #2675: bonded CUDA offload task
Update mdrun-performance.rst to clearly express the nature of task
We now have aspects of both PP and PME tasks that are independently targetable to the GPU for that task, or not (respectively bonded and fft calcs).
It is worth formalizing definition of task vs whatever we call "part of a task", and clarifying how to use things in
Our current docs are probably both repetitive and unclear, as we have both the docs on command-line options to mdrun (
Running :ref:`mdrun <gmx mdrun>` within a single node) including discussion of
-gputasks, and more things below (
Types of GPU tasks).
#6 Updated by Joe Jordan 7 months ago
Sorry I will try to get to this soon. Helpful on this front was a conversation Christian and I had with the Scientific Computing dept of a major Pharma company, who expressed doubt that this guide is useful since it does not seem to change much between versions. I will try to find time to do some benchmarking and more clearly express what (modern) best practices are.