More benchmark cheating
Apparently, now Apple is guilty of benchmark cheating. Yesterday, I said that the G5s were comperable to the top of the line P4s. This is mostly true, but not to the extent that Apple has claimed. They published well-known benchmarks named Spec, which appeared to show that the G5 was much faster than the P4. However, they used a really-fast-but-dirty single-threaded (ie useless in the real world) malloc implementation for their Spec marks, and disabled SSE2 for the P4 benchmarks, with the result that they were artifically inflated, and the P4 was artifically slowed down. If you compare their (inflated) Spec scores to the ones that Dell published for the same P4 Apple tested, you will see that the G5 loses. Not by too much, but the G5 is not the fastest desktop available, as Apple claims.
Now, this ties in with the Nvidia benchmark cheating earlier. Apparently, we cannot trust any benchmarks. <sigh> How are we supposed to choose hardware if we can’t trust any benchmarks? I guess that we just guess. At least for the Apple ones, we can just discount anything published by Apple themselves, as they didn’t modify the libc to cheat even when others run the benchmarks. Maybe someone will publish some Linux benchmarks on the G5. They should at least be accurate.
Not sure where you saw the info about the malloc implementation, but… It’s not uncommon for compilers to use different allocators for the threaded and non-threaded cases. It’s an obvious optimization: if you’re compiling a non-threaded program, drop the locking overhead.
I haven’t been actually reading all the stories I’ve seen regarding Apple’s cheating on benchmarks, but the ones I’ve seen so far seem to be just bitching; the benchmarks seem valid.