LHC@home

Would be a great RAM saver if they are.

yes, but efficiency/performance is bad with high core WUs.
here is a rather old post, but probably still valid:
https://lhcathome.cern.ch/ATLAS/forum_thread.php?id=568&postid=5208#5208

so probably there will be no 16, 24, 32 core ATLAS available in near future.

There was actually a performance gain at 11 cores comparable to the best gain at 4 cores in her chart.
Wonder why they then set the new limit to 8.
Maybe that data set is obsolete and the new WU version shows a steady decline in efficiency past 8 cores with no gains at 11.

There still is a RAM gain potential by common operating system and OS caching for ATLAS 16, 24 or 32 core if that means it's bundling 2x, 3x or 4x individual multicore WU's as the multi-core Theory application does.
8 core Theory just has 8 slots running 8 disentangled jobs.

So 16, 24 or 32 core ATLAS would have 2, 3 or 4 slots for 8 multi-core jobs.

Where is the 8core limit?

Maybe it was just pragmatic as most people don't have more than 8 cores?

I have 2 cores on the Linux boxen, an AMD Opteron 1210 and an AMD E-450. I have a Windows PC with an AMD A10-6700 which is sold as 4 cores, but the Windows 10 Task Manager says it has two cores and 4 logical processors. It uses only 2 cores, on any BOINC project.
Tullio

Where is the 8core limit?

Maybe it was just pragmatic as most people don't have more than 8 cores?

Could be pragmatism although there is an Intel Phi machine (user pomegranate) attempting to run LHC@home now so maybe they'll consider increasing the parallelism.

The 8 core limit is in David Cameron's post on current ATLAS WU information.
Cameron (Message 29560):

Each task can use from 1 to 8 cores according to what is set for "Max # CPUs" in the LHC@Home project preferences.

He says it was set because of their efficiency testing and posted the graph of results.
Cameron (Message 32802):

Above 8 it gets even worse so this is why we set 8 to be the maximum cores.

Implying that some algorithmic change will be needed to adapt to the Intel Phi if it looks theoretically possible to make gains with more parallelism.

Tried 3x 24's and 3x 16's to test if there were undocumented core counts available and all but one of them failed to start (white screens in VBox manage). The one that started was a 16 which had a string of errors and finally listed it's Virtual Cores as 1..16 with 10, 11 and 12 missing. It made it to the Cern login screen and only used 1% CPU for 2 hours.

I'll try an 11 and 12 core job and see if they are still supported but his remarks about efficiency with the current application design means these would probably not be useful.