That's not a real dichotomy. A RISC processor can range from completely lacking vector instructions to including complex matrix multiplications, and a so can a VLIW processor.
As far as why there are architectural differences between DSPs and GPUs, the largest reason is that DSPs are designed for processing one-dimensional data, like sensor readings or audio, where as GPUs are designed to process two-dimensional data, like pictures and video. Adding a dimension greatly increases the complexity of the data-processing algorithms the device will be running.
I read this old paper: https://ieeexplore.ieee.org/document/1176257 that compared VLIWs, superscalars and Vector processors for embedded devices and it turns out Vector processors perform the best (for the given benchmark) with the second-lowest power consumption
That's not a real dichotomy. A RISC processor can range from completely lacking vector instructions to including complex matrix multiplications, and a so can a VLIW processor.
As far as why there are architectural differences between DSPs and GPUs, the largest reason is that DSPs are designed for processing one-dimensional data, like sensor readings or audio, where as GPUs are designed to process two-dimensional data, like pictures and video. Adding a dimension greatly increases the complexity of the data-processing algorithms the device will be running.
I read this old paper: https://ieeexplore.ieee.org/document/1176257 that compared VLIWs, superscalars and Vector processors for embedded devices and it turns out Vector processors perform the best (for the given benchmark) with the second-lowest power consumption