Not mentioned in the article is the downside of having really wide decoders (and why they're not likely to get much larger) - essentially the big issue in all m

Not mentioned in the article is the downside of having really wide decoders (and why they're not likely to get much larger) - essentially the big issue in all modern CPUs is branch prediction because the cost of misprediction on a big CPU is so high - there's a rule-of-thumb that in real-world instruction streams there's a branch every 5 instructions or so ... that means that if you're decoding 8 instructions each bundle has 1 or 2 branches in it, if any are predicted taken you have to throw away the subsequent instructions - if you're decoding 16 instructions you've got 3 or 4 branches to predict, chances of having to throw something away gets higher as you go .... there's a law of diminishing returns that kicks in, and in fact has probably kicked in at 8