Skip to main content

Part 2: How Many CPU Cores Do You Need?

Time To Follow-Up

A few months ago, we looked into the effectiveness of using different numbers of CPU cores with various types of software. We received a lot of good feedback from that article, and there were some interesting suggestions from the community that we've taken to heart in this follow-up.

Primarily, there was a concern that part one might have been flawed technically, as the Core 2 Quad Q6600 we used in our testing does not share all 8 MB of its L2 cache between its four CPU cores. Intel's Q6600 instead has two separate 4 MB cache repositories, each shared between one pair of CPU cores. This means the quad- and triple-core results would have demonstrated the CPUs utilizing 8 MB of total cache, while the dual- and single-core results show that they were likely benefiting from 4 MB. Indeed, the benchmarks may have been reflecting the difference in L2 cache availability more than performance attributable to enabled processing cores.

To remedy this, we are using a different CPU this time around: AMD's Phenom II X4 955 BE. There are a number of reasons why the Phenom II is ideal for these tests. First of all, its 6 MB of L3 cache is shared between all four CPU cores, so the cache's impact on results will be kept to a minimum. Secondly, since there are now X2, X3, and X4 versions of the Phenom II CPU based on the same die, we will have the opportunity to test the validity of the method we use to simulate fewer CPU cores. By comparing simulated results to an actual retail CPU with fewer CPU cores, we will know more definitively whether disabling CPU cores in the operating system is a truly legitimate test.

At the end of these tests, we will be able to compare the Phenom II X4 results with the ones achieved by Intel's Core 2 Quad Q6600 to see if the impact of shared CPU cache is dramatic or minimal.

A few readers were also interested in simulating a scenario where multiple applications are running at the same time, in order to gauge the benefit of additional CPU cores while multitasking. We therefore ran a new test to analyze this type of scenario, too.