Memory performance
This is our first use of Sandra 2002's revised memory bandwidth measurement in a head-to-head processor test. Through buffering, SIMD extensions, and general magic, Sandra 2002 coaxes quite a bit more memory bandwidth out of both types of processor, which isn't entirely realistic, but then these are synthetic tests. What the new Sandra does do is show us the true limitations of the processor, platform, and memory. Have a look:

All of the DDR266 systems are deadlocked at 2000MB/s, or right at the peak of PC2100 memory's available real-world bandwidth. Only the RDRAM system is faster, but characteristically for RDRAM, it delivers much less peak performance than its theoretical 3.2GB/s spec seems to promise.

We'll use Linpack to look at the same question a little differently. Behold the trippy graph:

As always, Linpack demonstrates visibly the cache sizes of the various processors. The newer Pentium 4 chips, with their 512K L2 caches, perform extremely well, peaking out at nearly 900 MFLOPS at larger matrix sizes than the Athlon XP. At the far right of the graph, we're dealing entirely with main memory access, and in that department, every flavor of Pentium 4 delivers a bit more bandwidth than any Athlon XP. And again, the RDRAM setup offers more bandwidth than DDR266.

But then again, these tests are specific, synthetic benchmarks that don't always translate into real-world performance.