Make an app_config file to run 0.5 CPU and 1 GPU or MW to use 3 CPUs. You have 4 CPUs total so 4 CPU tasks cannot run along side a GPU task that also requires 1 CPU for a total of 5.

nBody tasks are as far as I know multi-threaded, so they will take all of the CPU cores that the hardware has, no matter how many you specify BOINC can use (be it through preferences or app_config.xml file).
Einstein tasks run on the GPU require a full CPU core, because they do part of the calculations on the CPU. Trying to set Einstein to use less CPU through an app_config.xml file will also result in it pausing. Until a full core is freed up again.

If you do not want to run nBody tasks from Milkyway, specify that via their project preferences
(Edit them, uncheck MilkyWay@Home N-Body Simulation, save preferences)

---

Having done a search on this (who does that? d'oh) it appears it is possible to use an app_config.xml file to specify that n-body tasks use less than the max amount of cores.

<app_config>
<app>
<name>milkyway_nbody</name>
<max_concurrent>1</max_concurrent>
</app>
<app_version>
<app_name>milkyway_nbody</app_name>
<plan_class>mt</plan_class>
<avg_ncpus>3</avg_ncpus>
<cmdline>--nthreads 3</cmdline>
</app_version>
</app_config>

Make the above file in Notepad and save as app_config.xml into %Your data directory/projects/milkyway.cs.rpi.edu_milkyway directory.
Save as all files type, in ANSI format. When saved, open BOINC Manager->View->Advanced view->Options->Read config files.

avg_ncpus tells BOINC how many threads to reserve for the job scheduling.
cmdline tells BOINC how many threads the application may use at maximum.

---

Jord is absolutely right, and I've done that with the Milkway Nbody tasks. They can run on any number of threads that you care to set for them, but do keep the numbers you use for avg_ncpus and nthreads equal.

You may find that any existing work you have already downloaded before making the change still runs using the old settings. Have patience - when you have downloaded new work after the change, and starting running it, it should behave as you want.

I think most require an app_config file to limit mt tasks to certain # threads.

LHC and Yafu have some option, not sure if it works.

I don't think these do.
MW
Cosmology
GPUGrid
PrimeGrid - Some apps can be run as mt.
BURP
SourceFinder

Cosmology FAQ has an example app_config file setup. Change the app names and it'll work for other projects. If its unknown what the app name is called then the event log will tell you the exact format.

I typically just copy one from project to project and edit the names as needed.