Investigations into Socket 939 Athlon 64 Overclocking

Application Performance

We'll start with some general application performance, courtesy of Winstones 2004. Winstones runs a bunch of scripts in a variety of real-world applications. The problem is that many of the scripts simulate user input and operate at speeds that no human can approach. Rendering an image, encoding audio or video, etc. can take time. Word, Excel, and Outlook, on the other hand, are almost entirely user-limited. While the fastest systems do perform higher, in every day use the typical office applications are going to run so fast that differentiating between the various overclocked settings is difficult, if not impossible.

We get a decent performance increase from overclocking, but nowhere near the theoretical maximum. Going from 1.8 GHz to 2.8 GHz represents a 64% CPU performance increase, although other factors would almost never allow us to realize that gain in benchmarks. In the Business Winstones test, we see a range from 21.9 to 27.6, a 26% increase. The Content Creation test gives a slightly larger increase, ranging from 28.3 to 39.7 - 40% more performance. If you like to think about it this way, the lack of performance scaling in the Business test can also "simulate" the user-limited aspect of office applications.

Similar in some ways to Winstones performance, PCMark attempts to gauge system performance. The results are a little more theoretical, as PCMark takes 5 to 10 minutes to run compared to 20 to 30 minutes for the Winstones tests. PCMark also includes some 2D and 3D graphics tests, which make the GPU somewhat important to the overall score. With Windows Vista moving to more hardware acceleration for windowing tasks, though, that's not necessarily a bad thing.

The difference between the slowest and fastest scores for our configuration is about the same as Winstones. PCMark04 goes from 3851 to 5567, a 45% increase. PCMark05 shows less of a difference, ranging from 3259 to 4146 (27%). PCMark05 is also the sole benchmark that we couldn't run to completion on the 2.8 GHz overclock. A couple of the tests failed every time. Both of the PCMark tests serve as great stress-tests of CPU overclocks, which is one of the reasons why we included the results. The failure to run complete PCMark05 at 2.80 GHz means that we definitely won't run this particular system at that speed long-term.

In case the graphs don't convey this fact well enough, our standard application scores benefited very little from the use of higher quality RAM. While the 2T command rate on the 9x300 value configuration did worse than the 9x289 value configuration, nearly all of the other tests show increasing performance, even with slightly lower memory speeds and latencies. The biggest gap between the value and performance RAM was in Business Winstones at 2.4 GHz, and even then, it was only a 5% margin of victory.