I have two Rosetta Beta 6.05 tasks on my computer now. The <workunit> entries in client_state.xml show <rsc_memory_bound> = 3,500,000,000 = 3.26 GiB. If the admins configured this correctly, it should reflect the
possible (not necessarily the actual) peak memory usage.
On large hosts with many CPUs and sufficient RAM, boinc's scheduling of tasks based on this advisory peak memory usage can indeed lead to lower than necessary utilization. Scheduling should rather take a value somewhere between peak and average RAM usage into account on such larger hosts. In order to do so, RAM can be overcommitted somewhat, especially if there is swap space as reserve. Overcommitment is achieved by hand-editing global_prefs_override.xml to have <ram_max_used_busy_pct> and <ram_max_used_idle_pct> higher than 100. (I learned this recipe from
@Endgame124's Raspberry Pi thread.) Of course, overcommitted RAM is risky and needs to be done only if the real RAM utilization has been (and is) monitored by the user very well.