Tests of FP16 ops fail for the PlaidML plugin in bizarre ways. An example error message is

Relative comparison of values expected: -4713.58544921875 and actual: -54740496384 at index 0 with threshold 0.0099999997764825821 failed

This is especially notable as the most negative FP16 value is -65504, so it would not be possible to achieve this value in FP16.

However, CPU tests for these same ops with fp16 precision do pass.