Comment on: Question: Where does winner deduplication live?

It always fires twice. Not a CPU expert, so correct me if I'm wrong: You could shave a vanishingly small gain off a long running computation. I don't think at this granularity you would even be able to have a noticeable effect on a DSP pipeline. Not to mention LLMs, there's tail latency sure, but intuitively its OOMs more than can be accounted for by CAS stalls. Drilling down on the benchmarks will show whether there's anything compounding there soon enough, I guess it's possible. The implications and workarounds are real... this really is a problem that should be solved in hardware level, but I suppose ultimately it's just a cost cutting measure, or profit maximizing, because so much of computing nowadays seem to be about trying to outsmart contention rather than eliminating it. I can only guess why it was introduced. I guess the answer to your question is: either you don't need deduplication because the old data will just look the same as the new data and overwrite itself? ...