On 11/6/18 1:11 AM, Ben Melosh wrote:
Thanks for the help Robert,
I will look into these different options. I am wondering what specifically is problematic about using very different or very large magnitudes of items in different blocks? Is it a problem with processing power? I am hoping for specifics so I can better understand what is going on in the model and improve it.
It is a problem of machine precision - when using the 64 bit float values, 1 + 1e-17 is still just 1, so having a block with values of the order of magnitude 1e20 and another or the order 1e3, the second block is essentially zeros from the global matrix point of view.