SSE / Non-SSE Considerations and Myths

According to a forwarded internal memo from AMD, substituting the Palomino for a standard Thunderbird in an existing setup will not allow to make optimal use of the SSE potential of the Palomino, meaning that the operating system needs to be freshly installed with the Palomino running as CPU. Likewise, substituting a Thunderbird for the Palomino if the latter was used for the installation of the OS will hurt the performance of the Thunderbird. In one sentence, regardless of which way the OS was installed or whether the CPU was switched in between applications, all benchmark scores came out identical (within the margins of error), in other words, it doesn't matter at all.

We were using the latest beta version (821) of SiSoft Sandra intermittently to the benchmarks for the simple reason of verifying that the CPU speed was, in fact, what the BIOS screen reported. We noticed that, on occasion, the Palomino would reboot at a 10.5 x CPU multiplier which made us suspicious that, in fact we might compare apples with oranges. To avoid any mistakes in clock speed, we ran SiSoft Sandra as internal control to verify the appropriate settings and CPU recognition before running any benchmarks.

The SiSoft Sandra Memory Benchmarks show the Thunderbird (above) taking a lead of about 4.5% over the Palomino (below) .

CPU Arithmetic Benchmark

In the CPU Arithmetic benchmark, our Thunderbird results exactly matched those given as internal reference (Reference CPU 3), the actual results show the Palomino scores. The differences between the Thunderbird and Palomino are negligible which is to be expected in a purely arithmetic environment

CPU MultiMedia Benchmark

In the CPU MMX benchmark, the Palomino actually beats the Thunderbird in Integer performance but shows a slightly lower floating point performance than the Thunderbird (Reference CPU 3)

All in all, SiSoft Sandra shows no differences between the two CPUs where no differences are expected and shows some impact of the SSE instruction set in the multimedia benchmark where we would expect these differences. The lower FPU performance is hard to explain but keep in mind that the SiSoft Sandra (812) is still a beta version. The key point of these benchmarks, however is to show independent verification that the system was running under identical conditions except for the CPU in either case.