GPU computing speed is faster. Conversion from float16 to float32 before training, it's depends on several factors in case of speed such as where the data loading before conversions in CPU or GPU. If you perform a conversion task on the CPU then the GPU will be reserved for further tasks. But it would be better directly on the GPU.

Other things also noticeable for parallel tasks are the GPU is faster such as deep learning model training, graphics rendering, and matrix multiplications. But for Sequential tasks such as complex logic, OS operations, and web server running of course CPU is typically faster.

GPU is better than CPU, you should prefer to use GPU especially for tasks which requires more processing time to avoid the load on CPU, GPU will be able to process your task quickly and also your CPU performance won't effect

Converting data types on the CPU can be efficient if your data is initially stored in CPU memory. CPUs are optimized for a wide range of general-purpose tasks, including data type conversion. If the data is not yet loaded onto the GPU, it makes sense to perform the conversion on the CPU to avoid unnecessary data transfer between the CPU and GPU