Agree that OoO is the big cost. But we can also mitigate that without VLIW: SIMD/vector reduces the instruction count by ~5x, and energy by a similar factor. A

Agree that OoO is the big cost. But we can also mitigate that without VLIW: SIMD/vector reduces the instruction count by ~5x, and energy by a similar factor. And a portable API such as Highway also helps us move the same code from x86 to Arm or RISC-V with just a recompile :D