Ask HN: What made VLIW a good fit for DSPs compared to GPUs?

Why didn’t DSPs evolve toward vector accelerators instead of VLIW, despite having highly regular data-parallel workloads