10. Half-precision floating point and stochastic rounding
The IPU supports IEEE half-precision floating-point numbers, and supports stochastic rounding in hardware. The IPU extensions to TensorFlow expose this floating point functionality through the functions described below. See the TensorFlow Python API for more details.
10.1. Controlling the half-precision floating-point unit
You can configure the floating-point behaviour of the hardware using the
floating_point_behaviour category of an
esr bit enables the stochastic rounding unit. Three of the remaining
options control the generation of hardware exceptions on various conditions.
nanoo bit selects between clipping or generating a NaN
when a half-precision number overflows.
10.2. Resetting the global random number seed
The stochastic rounding unit and the TensorFlow stateful random number generators both use a common global random number seed to initialise the random number generator hardware. Each IPU device has its own seed.
By default this seed is set randomly, but it can be reset by using the function
Due to the hardware threading in the device, if the seed reset function is used
target.deterministicWorkers Poplar Engine option will need to be set
This can be done using the
compilation_poplar_options option of an