A Performance-Correctness Explicitly-Decoupled Architecture

  • Michael: Are there any other good examples for EDA? Tunji: Value prediction wouldn't probably work.
  • Eriko: Are there other types of analysis? e.g. for power improvement.
  • Eriko: Does profiling always work? Michael: If control flow changes e.g. depending on input, then yes.
  • Chris F.: Building program skeleton might be even better with JIT.
  • Janini: Wouldn't you rather have two simple cores? Tunji: code might not be parallelizable.
  • Eriko: How about area? Area-wise they don't study if the EDA core is larger than a sigle core.
  • Optimistic core might be turned off if it's not providing useful/enough information or if the buffer is full.
  • Michael: The correct core cannot be simple (e.g. in-order), because some workloads might not benefit from the optimized core and you wouldn't want performance to be terrible for some programs.
  • Eriko: What if OS get's involved? Chris C.: Maybe set thread priorities. Tunji: Issue I/O or acquire lock ahead of time. Siddarth: You can't expose that, because it can't be speculative (you can't be wrong).
  • Eriko: It seems that the EDA approach only works if there is loose coupling.