[newgre]

Why did the compiler even chose to fetch DWORDs only in the first place? It's unclear to me why the accumulator (apparently) determines the vectorization width?

[TinkersW]

The accumulator is a vector type, with 64 bit sum you can only fit 4 into a 256 bit register.

After the loop it will do a horizontal add across the vector register to produce the final scalar result.