Data Prefetch Logic – What is it Worth?

Pages: 1 2 3 4 5 6 7 8 9

Application Benchmarks

In the end the only thing that really matters to most people is what this feature does for ‘real world’ performance. Though many publications seem to be moving away from the ‘industry standard’ benchmarks in favor of individual application benchmarks, I think SYSmark and Winstone still have a place, even with the complaints leveled against them. In this situation I am not comparing two different architectures, so there can be no questions of bias or favoring of one architectural feature over another. For that reason, I believe these are perfectly valid benchmarks that can show a more ‘real world’ performance impact of the Data Prefetch logic than the synthetic benchmarks provided earlier.

With that said, here are the results from the most current of these benchmark suites. Note that the ‘Average’ shown for the Winstone tests is the result of throwing out the highest and lowest scores, then averaging the middle three. This is my preferred method of averaging the Winstone scores because I believe it provides a better picture of what the typical performance increase will be, rather than the default reporting method of listing the highest score achieved. SYSmark does its own averaging of three runs:

The results here are a little mixed, which seems to illustrate some of the issues I have raised about these benchmarks in my reviews (see below). Both test suites in SYSmark 2002 and the Business Winstone 2001 results indicate that the Data Prefetch logic provides a greater benefit than increasing the FSB by 33%. This is pretty impressive, however Content Creation Winstone 2002 shows the opposite. Furthermore, SYSmark 2002 shows that Data Prefetch improves scores by about 10% overall, while both Winstone tests indicate it is around 5%. Remember, however, that the Tualatin part seems to have a slightly worse latency than the Coppermine part, so this could have some limiting effect on the overall system performance.

In my reviews of both of these benchmarks, I indicated that due to the methodology used the SYSmark results tend to show a greater improvement for clock speed related differences than the Winstone results do, while minimizing the effects of memory size. I might have used more memory (i.e., 512MB), but my tests have shown that above 256MB there isn’t much improvement in any of the test results for these benchmarks, so the relative scores would remain very similar.

Pages: « Prev   1 2 3 4 5 6 7 8 9   Next »