PCMark2002 – A First Look

Pages: 1 2 3 4 5 6

CPU Test Results (Cont.)

After seeing the results on the previous page, I got curious about the working set size of these tests. Well, one way to find out is to run two or three identical processors with a different amount of cache, and it turns out that the PII 300, Celeron 300A and Celeron 300 fit the bill perfectly. Both the Celeron 300 and PII 300 are based upon the Deschutes core, while the 300A uses the Mendocino core. The Deschutes core did not include L2 cache, so the PII 300 has 512K of ½ speed external cache, while the Celeron 300A has 128KB of integrated, full speed L2 cache. The Celeron 300 has no L2 cache, of course. All three have 32KB of full speed, split (16K data/16K instruction) L1 cache. So, how does this look to PCMark2002?

PII 300

Cel. 300A

Cel 300





Zlib Compression




Zlib Decompression




Text Search




Audio Conversion




3D Vectorize




CPU Overall




It would appear that the working set size for all tests fit well within 128KB cache, and some fit into 32KB of cache! Only the Zlib compression/decompression and Audio Conversion tests show any degradation on the Celeron 300, and even then the Audio Conversion different is not that significant. In the overall score, we see only a bit more than a 11% difference between Celeron 300 and Celeron 300A. From this, I have come to two conclusions.

  1. The CPU tests in PCMark2002 are fairly small in size, and if they fit into the L1 cache the performance measured will essentially be that of the processor core with little dependence on memory speed. However, it also means that a processor with a small L1 cache (such as the P4) may be at a disadvantage.
  2. One of the biggest complaints about SPEC CPU2000 is that it is ‘bandwidth dependent’, meaning none of the tests fit within the L1, L2 or L3 cache of any current processor, so all memory references must go to system RAM. As can be seen here, unless one can guarantee all of the tests will fit within even the smallest L1 cache, the only alternative is to force the working set size to be larger than any cache, or you will get skewed results on the relative strength of the processor core. In this case, we can see that two identical cores (PII and Celeron 300) get different CPU scores because of the cache, and that the full speed L2 cache of the Celeron 300A provides a small boost over the PII 300, when they should all be essentially the same (if we are measuring only the CPU performance).

It will be interesting to run these tests on other processors from Intel and AMD to see how they fare. As noted earlier, these results can be very useful, but it is necessary to really understand what is going on underneath the covers of both the processors being tested and the benchmark itself to really come to any useful conclusions.

Pages: « Prev   1 2 3 4 5 6   Next »

Be the first to discuss this article!