I chose several benchmarks to gather, solely on the basis of what seemed a good idea at the time. The goal was to get a picture of how well the newer processors do on the old tasks, not just on optimized armlets but also when running unoptimized code through the PACE emulator. This seems relevant because the vast majority of programs will fall into the latter category for the foreseeable future; on average, if it works, that's good enough for the typical development team given tight schedules.

Some of these tests were run by other people in other contexts. When that's the case, the score is a hyperlink to the web page I found it on. If there's no hyperlink, I probably ran the test myself.

Systems that could run both armlet and unoptimized tests have both reported, the unoptimized reported as a PACE score, in green italics.

Notes

• QuickBits on the Palm IIIc makes a remarkable difference. This is a tiny utility that I thought was only speeding up my screen. I only tried turning it back off after noticing my benchmarks didn't match up to the reported ones.
• PACE means "non-ARM-let" code, or non-optimized. Most applications will fall into this category.
• I was very suprised at the differences in scores between my measurements with the applications and those reported by others on the web. In some cases, the scores were essentially bang-on (in which case I didn't bother reporting them again), but in other cases, such as the Tungten T scores, the differences were significant.
• The Palm IIIc is the only low-rez device in this review; the others have 320x320x 16bit or greater resolution while the Palm IIIc has 160x160x 8bit, which means it has one-eighth the screen memory to move around.
• The Treo 90 has a 12-bit low-rez screen (with poor color too). Consequently the screen test should be slower than on the Palm IIIc, due to 50% more data to move, but it's far slower than that would suggest. I'm not sure if this is due to having to split bytes, but the results are internally-consistent; the kB/sec of screen and memory writes are about equal, as are the per-loop times.  It's just an amazingly poor video performance per-loop.
  
• Sony's faster processors do indeed seem to be faster than the Tungsten by a proportional amount, but the real shocker is the SJ33 speeds on video and memory, where it not only easily out-performed the Tungsten and the XScale Clies in emulation (PACE) mode, but nearly equalled them in their native (armlet) modes.