The [Evaluation Flow document](https://docs.google.com/document/d/1xDGFSYxIE0AKsGAI3ccLz1EX3bLHOvDtwl3983G5kYk/edit?usp=sharing) also sheds light on the relative
strength of a "Pure Commercial Flow", as follows. CT uses the placement information generated by physical synthesis (Genus iSpatial). Observe that if we go straight into Evaluation Flow 1 from physical synthesis (without running CT), this will produce a "pure commercial flow" (i.e., CMP) outcome without any use of
Circuit Training. From the data in [Evaluation Flow document](https://docs.google.com/document/d/1xDGFSYxIE0AKsGAI3ccLz1EX3bLHOvDtwl3983G5kYk/edit?usp=sharing),
we see that with the "pure commercial flow", CMP macro placements produce similar timing and power numbers compared to CT macro placements. However, the postRouteOpt wirelength of CT macro placements is at least 18% larger than the postRouteOpt wirelength of CMP macro placements.
Please note that we report this data as part of our study of Circuit Training. It is not intended to "benchmark" any commercial EDA tool, and the data should not be interpreted as providing any sort of "benchmarking" comparison.
**November 27:**
<aid="Question3ext"></a>
We have extended the experiment of [Question 3](#Question3) to assess the difficulty of our testcases. As mentioned [here](#Question3), we take the CT-generated macro placement and then randomly swap the same-size macros. We use the [shuffle_macro.tcl](https://github.com/TILOS-AI-Institute/MacroPlacement/blob/651a36626dd778018c5cf867b419e44f64fb103e/Flows/util/shuffle_macro.tcl#L29) script for this experiment. The following items provide details of the macro shuffling experiments for different testcases.
