This depends upon your metric.

For raw performance:

---

You could use the High Performance Linpack across all cores to get

a measure an upper bound speed in FLOPS (Floating Point Ops per sec.)

that you could expect in a real scientific application. Usually you

will get nowhere near this though in a real application.

There are also the Spec benchmarks both for integer and floating point.

If you want polygons per second for graphics, you'll need a different

benchmark.

I really like to use the stream benchmark using either MPI or OpenMP

with large vectors to measure the max memory bandwidth. This is usually

the bottleneck for applications, getting data in and out of the processor.

You might want to measure the power usage while running a benchmark

to get a flops/watt , since energy seems to have become a limiting factor these days.

Modern processors are manufactured not just to speed up the execution of tasks however, reducing energy consumption is a critical issue in the design of these systems to prolong battery life. Some of these processors are working on real time systems which is critical to meet the deadline of jobs. So many factors should be addressed to evaluate the performance of such systems.

If an accurate result on the threads is required, consider to use a real-time operating system.