The AMD FX-8150 (Bulldozer), dual AMD Opteron 2384 (Shanghai), Intel Core i5 2500K (Sandy Bridge), Intel Core i7 2630QM (Sandy Bridge), and Intel Core i7 990X (Gulftown) CPUs were used for these multi-core scaling tests. Each processor was tested with 1/2/4/6/8/12 threads enabled, up to the maximum number of logical cores offered by each processor.
All systems were running Ubuntu 11.10 64-bit for this testing with a near-final Linux 3.1 kernel that did contain the IC aliasing patch from AMD that's specific to the Family 15h (Bulldozer) CPUs. The hardware configurations remain pretty much the same as the results from Monday. As always, this testing was all facilitated by the Phoronix Test Suite so it was fully automated and reproducible. You can run a similar test yourself and compare the results directly to the numbers in this article by running phoronix-test-suite benchmark 1110227-AR-AMDSCAL0184. With the numbers on OpenBenchmarking.org, the results from all of these processors were normalized.
With the results normalized by the Phoronix Test Suite / OpenBenchmarking.org, in the graphs about to be displayed, each CPU has a value of 1.0 when only one core is enabled (since all of these tests are very well multi-threaded under Linux, so obviously with just one core it's the slowest) and then with multiple threads/cores, those results are all relative to the single-core result of that individual processor.
In other words, it should be a very fair game in looking at how well each is scaling with multiple threads. For those interested in the raw results, they can be obtained from OpenBenchmarking.org using the 1110227-AR-AMDSCAL0184 ID. Also to note is that when the results are normalized, for tests where normally the results are produced as "less is better" (e.g. time to complete XXX task), those numbers are inverted before being normalized. It is just a matter of checking a box within OpenBenchmarking.org to achieve this, among its many other features it has now and especially more advancements coming soon.