Native FP16 is 7% faster.
Taken from Videocardz … When AMD made its FidelityFX Super Resolution technology open source, the GPU manufacturer disclosed that the upscaling technology supports a fallback for older GPUs.
If the GPU does not support half-precision (FP16) compute acceleration, AMD offers a fallback to single-precision (FP32) computations. This can be, for instance, set by manipulating the flags in the AMS FSR samples’ source code.
Older GPUs such as NVIDIA Pascal/Maxwell as well as AMD’s own Polaris series are all using the fallback to FP32 precision. This does not mean that newer GPUs such as Radeon RX 6800XT cannot be tested with the fallback. In fact, this is exactly what CapFrameX has done to see what is the actual advantage of using the native half-precision coding instead of single-precision.
It was concluded that the Navi 21 powered graphics card sees 7% higher performance when using FSR 1.0 in FP16 mode. To force FP32 fallback, CapFrameX edited the source code, which AMD has published on the GPU Open website.
Read the rest at Videocardz