If by "sleeping" you're referring to an ACPI sleep state, you're probably getting killed by the amount of physical memory you have installed.
That's probably true. It's fully populated, so keeping that alive is likely consuming a good bit of the sleep power. The fans don't power down either, though. That's fine, I just won't run the super-long-time-between-checkpoint units on it very much (cpdn has some of those for Linux hosts - I like them, but also have a core or two of my 24/7 home server throwing cycles after them - which may change when I get home solar online, depending on energy produced vs used balanace). Those WUs tend to be cache-brutal as well, so only having one or two is fine. They're a workload where you can watch the total machine throughput (in instructions retired per second across all cores) drop as you add more of them, because of cache pressure.
72 GB is overkill for Rosetta@Home. If you pulled one stick per channel, you'd reduce your power usage, and would still have 48 GB available. Your memory frequency may also increase with fewer DIMMs per channels. Dropping down to 24 GB (1 DIMM per channel) would reduce power usage further, although you'd probably start swapping with some of the heavier WU's.
Going to 48 is worth considering, going to 24, no. There are plenty of >1GB WUs I get, and I've been seeing some 2GB units. I'd rather be able to fit them, though the question of total system throughput under different workunits is an interesting question. I've seen this system "flatline" around 18 running threads of Rosetta and such, but I'm not seeing performance reversions on this set of WUs. Also, being a server board, I'm not sure what the memory frequency would do if I pulled sticks out. I'm not sure I particularly care, having 72GB of RAM is just sort of fun. Most I've ever had in a personal system by a good margin - I upgraded a box to 48GB recently before donating it for an indefinite period for video crunching work, previously 32GB was my high water mark.
Also on the topic of power efficiency, if your goal is performance/watt rather than raw performance, I'd suggest restricting your maximum clock speed to below your base clock. This will effectively disable Turbo Boost, but should increase your performance/watt. I'm guessing setting the max frequency 1-2 ticks below the max base clock will produce the best efficiency, without dropping throughput too much.
Power efficiency is gained by turning it off during the hours the sun isn't up, and when it's cloudy.
On a sunny day, I can't use all the power my panels generate as-is, and I'm working on adding 50% more to give me better winter running capability (I'd like to get off my generator for the most part in the winter). At some point, I might start charging the EV out there, but I'll see. Power efficiency just isn't a concern of mine though. This is an old, power hungry box that's retired to a place that is power consumption limited to do some good before getting fully retired. My panels spend most of the day in use-it-or-lose-it mode, where I just can't consume everything they could potentially generate. This is pretty standard for off-grid systems, because if you're paneled for winter, with short, cloudy days, you have more power than you know what to do with during the long, sunny summer days. This is my way of attempting to do something useful with it - it's either that or cryptocurrency mining, and this seemed an awful lot more "globally useful."
I can also radically improve power efficiency by moving the computer outside, and not having to pump hot air out of my office. I just haven't gotten around to building a sheltered plywood enclosure for servers and then figuring out how to mount it somewhere.
EDIT: You may also want to consider joining
Ibercivis, which is currently running a project that
tests whether some existing drugs are effective at treating COVID-19. In addition to spreading your compute power around more COVID-related projects, Ibercivis WU's are memory-light (all of my WU's have peaked under 100 MB), which makes it easier to run with only 24 GB of RAM.
Sure, I'll throw a few cycles at them. The Rosetta workunits have been running tighter-than-I-can-deal-with deadlines a lot of the time, though they seem to terminate earlier than expected a lot of the time. WCG is soaking up the spares, but I'll toss ibercivis in and let it fight for some cycles.
