This parameter controls how data is divided to stream to/from the GPU in parallel with processing. If you change the default value you should consider the following:
- Low value of ICHUNK:
- (-) Smaller transfer to/from the GPUs,
- (-) Less reuse of data on the GPUs,
- (-) Less work performed with each kernel execution,
- (+) Quicker startup and shutdown.
- High value of ICHUNK:
- (+) Larger transfers to/from the GPUs,
- (+) Increased reuse of data on the GPUs,
- (+) More work performed with each kernel execution,
- (-) Longer startup and shutdown.
Values that are not an integer multiple of CUDA cores may result in decreased performance. The default value of 0 will cause FMS to select the optimum value for the type of data being processed (real or complex) and the properties of the GPUs.