The big win is memory. Half the memory/cache usage and half the bandwidth. For SIMD you'll want 32 bit floats. But in many cases instead of reaching for SIMD yo

The big win is memory. Half the memory/cache usage and half the bandwidth. For SIMD you'll want 32 bit floats. But in many cases instead of reaching for SIMD you should go to compute shaders, which are definitely going to be 32 bit. And of course any graphics shaders will be using 32 bit floats.