The folks up north at Primate Labs have some benchmarks of a PowerMac G5 and a Core 2 Duo iMac running under Tiger and Leopard. They used their cross-platform benchmark Geekbench for the tests.
The iMac did pretty well. Under Tiger, it got an overall score of 2699, with 1000 on the scale representing a 1.6GHz G5. Under Leopard, the iMac's score went down a bit to 2618, but in 64-bit mode, it was 2878. Although the results varied from one test to another, the trend was pretty clear: the iMac performed the tests included in Geekbench a bit slower under Leopard 32-bit, but faster under Leopard in 64-bit. The only exception was memory performance, where Leopard was 5 percent slower, regardless of the number of bits.
When it was the G5's turn, things didn't look quite as rosy. This machine had a score of 1013 under Tiger, but performed a good 10 percent slower under Leopard in 32-bit mode (898) and nearly 15 percent slower in 64-bit mode (853). Ouch! The blame for the poor 64-bit performance lies with the integer test: this one completes a whopping 30 percent slower than under Tiger.
This warrants two conclusions:
- There are certain performance regressions in LeopardOn average, 64-bit computing is a win on Intel, but it mostly hurts the PowerPC
There is a very important caveat, however: in the absence of knowledge to the contrary, I'm assuming that Geekbench doesn't do anything that is 64-bit in nature. In other words, it does calculations using (at most) 32 bits at a time, and doesn't need to address more than 4GB of memory. Applications that do need to run 64-bit calculations and/or address more than 4GB of memory could still be faster in 64-bit mode than in 32-bit mode on the PowerPC G5. Also, overall performance depends on more than just the CPU: offloading more work to the video card can make all the difference.