Speedup comparison between precompiled and uncompiled UMATs

Hello, I've been trying to calculate the speedup between UMATs that have been precompiled using the 'abaqus make' command and those that are compiled during analysis. However, I can't find any documentation or technical notes that explicitly compare the two methods, and my own work is inconclusive (e.g., some analyses are faster when precompiled and some are slower). My intuition is that precompiling the UMATs should be on average faster, but I'm not sure if this is correct. Does anyone have additional data/guidance on this topic?