In our Vega technology overview we talked a bit about AMD having the ability (with Vega) to have several render targets in a scene to render at lower precision (fp16 versus fp32). AMD calls this feature fp16 Rapid Packed Math (floating point 16). According to AMD this will increase performance.
Rapid Packed Math basically halves the floating point calculations for data request, resulting in a faster turnaround time of that request/data, however with less precision and thus quality in some form. Considering lowering image quality in specific segments will pretty much always create more FPS, we'll just have to wait and see how the feature will evolve. Basically half-precision would be applied in segments where it really isn't needed. AMD has not revealed anything specific as to what and where excatly the feature can be used.
Two developpers have shared that their games will be supporting Rapid Packed Math fp16, Publisher Bethesda and Ubisoft. These initial titles are Wolfenstein 2 The New Colossus and Far Cry 5. Futuremark also is working on a new scene called "Serra" that will Support it.
NVIDIA has not announced plans of supporting the new option, it is a hardware feature, meaning that owners of current NVIDIA GPUs and current AMD GPUs won’t be able to benefit from it.