I am seeing potentially conflicting information.

In the programming guide, Table 33 doesn't indicate that FP64 is supported as an input data type.

In section 5.4.11.3, I encounter the statement "Tensor Cores support double-precision floating point operations on devices with compute capability 8.0 and higher...".

Since 8.6>8.0, we have a possible contradiction.

What am I missing?