Turn off Intel Speedstep. Also try disabling any "turbo" performance features in BIOS.

Its fairly common for one (or more) thread to get loaded more than the others. Multiprocessor load balancing is not perfect since plugin loads vary. Also if you have 10 plugins on one track/bus they are all processed on a single thread. If your project had all other tracks with a single plugin for example, the thread with 10 plugins would consume more.
The same would apply depending on how expensive the plugins are.
In short its unrealistic to expect fully balanced cores in a real world situation. I hope to get some time to revisit this area at some point since there are some ideas we have to improve this...

If one or two cores are getting massively loaded and not much is happening on the others try deleting or renaming Sonar's aud.ini file and let it create a new one.
 
I found Splat when first installed (X2 and X3 were already installed) was thrashing two cores to the point of dropouts while completely ignoring the others. Letting Sonar create a new aud.ini from scratch solved the problem.

Noel Borthwick [Cakewalk]
I Also if you have 10 plugins on one track/bus they are all processed on a single thread.

So, no pipeline parallel processing in Sonar. Thanks for info

Well a chain of plugins is by definition serially processed since each plugin needs its predecessor's output.
What would you process in parallel? :)

Noel Borthwick [Cakewalk]
Well a chain of plugins is by definition serially processed since each plugin needs its predecessor's output.
What would you process in parallel? :)

The next chuck of audio data. Usual pipeline parallelization. I do not claim it make sense for most Sonar use cases, but during offline rendering that can consume all available cores and so speedup the process. So I was unsure either Sonar sound engine has something like that. And it is hard to observe on my 2 core computers ;)

I'm running an i7 4770k and regularly load 12 to 16 instances EW Play with Hollywood orchestra along with a couple of Kontakt multiple output instances.

My CPU usage is usually pretty low but distributed more or less evenly. Processor 1 always has a bit more of a load on it. Play works out of the box but Kontakt needs to be set to support multiple threads . Perhaps the vstis are part of the issue.

Surely that would only work if you knew there was no relationship between the two blocks of audio (which you couldn't necessarily - think reverbs, delays or look-ahead plug-ins)?

Noel Borthwick [Cakewalk]
Its fairly common for one (or more) thread to get loaded more than the others. Multiprocessor load balancing is not perfect since plugin loads vary. Also if you have 10 plugins on one track/bus they are all processed on a single thread. If your project had all other tracks with a single plugin for example, the thread with 10 plugins would consume more.
The same would apply depending on how expensive the plugins are.
In short its unrealistic to expect fully balanced cores in a real world situation. I hope to get some time to revisit this area at some point since there are some ideas we have to improve this...

That was very informative thanks. Plugins like Maschine and Kontakt apparently have multicore support? Is it best to set them to use a single core at present?

You can try it but it really depends on the project. If you have a lot of other stuff going on, multithreading at the plugin level can potentially conflict with the hosts multi-threading and give you worse performance.