How to leverage gem5's parallel capabilities, especially to fully utilize the high bandwidth of HBMCtrl?

[LujhCoconut]

Based on my understanding and some experimental analysis, whether in SE mode or FS mode, the final memory programming model provides external APIs similar to sendTimingReq. It seems that, regardless of whether it is a single-core or multi-core setup and the type of dataset used, memory accesses involving APIs like sendTimingReq appear to be serial. When I attempted to use HBM2Stack as memory, I expected it to perform well for programs suitable for parallel processing. However, unfortunately, the output I obtained with MemCtrl as a debug flag showed that for each 64B request, HBMCtrl splits it into two 32B requests sent to two pseudo channels. First, the 64B requests are all handled serially; second, the two 32B requests are not submitted simultaneously either. Therefore, my question is: Can gem5 fully leverage the high bandwidth of HBM?