Oh geeze I went from an UltraSPARC III to a T1 and performance tanked.
Turns out my use case was not large numbers of parallel integer only threads.
Basically no real use-case actually are pure integer like that. Basically it performed badly on pretty much all workloads that anybody cared about.
It was all aimed at data shuffling. Think disk<->network stuff that's all byte and integer based work. Four threads per core to wait on spinning rust that was prevalent in it's day. Certainly terrible for general purpose compute and work loads.