Processing 32 inputs at 1-byte granularity means that you'll need to expand each of those 32 inputs across 5 registers, and means you lose the ability to trivia

Processing 32 inputs at 1-byte granularity means that you'll need to expand each of those 32 inputs across 5 registers, and means you lose the ability to trivially shift them along (would need 15 blends to align the 5 input bytes to 3 output bytes for the 32×i8 approach, vs a single vpsllvd or some shuffle for 8×i32), or do any arithmetic wider than 8-bit (i.e. no 64-bit multiply at the core of the method in the article). More lanes just for the sake of it makes no sense (not counting unrolling as you can unroll anything to any amount (give or take register pressure)).