Just to add that scheduling overhead goes away with SMT (assuming you don’t oversubscribe), but the sharing of caches and branch prediction logic is still an issue as you point out.