A GPU is a graphic processor specifically designed for graphics. It is different than the CPU which is general purpose. A GPU is more efficient for graphic processing tasks. 3D rendering programs use the GPU as hardware acceleration, which is noticeable faster than doing rendering on the CPU. An ASIC is an application specific chip, for DL and ML, it means efficient and very specific implementations of the algorithms to get the fastest and/or lowest power processing. An FPGA is reconfigurablebut in essence an ASIC but with worse speed and power performance than an ASIC in the same tech (for the same design). Both ASIC and FPGA are for hardware acceleration, speeding up ML and DL. But an ASIC comes with a huge engineering cost. An FPGA is an off-the-shelf product. Low volume products, development and prototyping are uses for an FPGA. An ASIC would beonly (financially) interesting for high volume products.
What will be the relative importance of ASIC and FPGA for training machine learning / deep learning models in the near future compared to GPU?