I guess you're talking about scatter / gather? All vector memory instructions in the GCN ISA are scatter / gather-type instructions, i.e. you provide one pointe

I guess you're talking about scatter / gather? All vector memory instructions in the GCN ISA are scatter / gather-type instructions, i.e. you provide one pointer (or buffer index, or texture coordinates) per SIMD lane. Docs say that the performance depends on the actual distribution of pointers, i.e. for best performance you should ensure that the pointers are consecutive (use SoA instead of AoS layouts, etc.), but I imagine that the drop in performance is somewhat gradual and graceful as your pointers become more scattered. I don't think the details are really documented though, and I wouldn't be surprised if those details changed between GCN generations.