I've yet to see a common situation where you can do better than a primary thread and a task pool*. It's extremely hard to get synchronization right across multiple threads especially when they don't have very explicit sync points.

In short, unless you really know you need to mess with fine-grained threading, I'd say do the simple thing and just write a master thread that issues parallelizable work to your task thread pool. It'll be a lot less headache and you can leverage libraries like Intel's TBB et. al. to do the heavy lifting.


*NB for obsessive completeness: there are situations where you can beat out a task pool using fine-grained threading. It's very difficult, though, and not at all common.

There's two reasons to use multiple threads:
A) You've got processor intensive routines that you can run in parallel.
B) You have to call a slow, blocking function, but want it to be asynchronous.

In any case, if you're running more threads than your CPU has 'hardware threads' then it's called over-subscription, and is a waste of performance. The opposite, where you're running less threads than your CPU can handle is called under-subscription, and is a waste of potential. Ideally, your game will make use of the same number of threads that is supported by the CPU (or less, to allow for the OS or other apps).

My idea so far is to have the Game/AI/Physics be one main thread. The graphics be another main thread that updates asynchronously from the game. The sound output [and networking] would be its own thread as well.

So you've got a fixed number of 4 threads (assuming one network connection), which means you're targeting quad-core CPU's only. On dual-core, you'll be over-subscribed, and on hex-core you'll be under-subscribed.

Each of your threads above has a very different workload. The Game/AI/Physics thread is going to have a very large workload, the graphics one a mild workload, and the audio/network threads are going to be almost entirely idle. Ideally, each of your threads would have almost identical workloads.

AI/Physics/Graphics are a case of (A) -- you've got a lot of work to do, which can largely be done in parallel. Ideally, you'd split the workload of each of these amongst each of your threads to balance it over the CPU, and use it's full potential.

Networking is a case of (B) above -- instead of wrapping a blocking API in threads to make it asynchronous, just use an asynchronous networking API to begin with -- so you don't need separate threads dedicated solely to networking.