With any system you will want to see a combination of synthetic testing and real-world. Synthetics give you a static, easily repeatable testing method that can be compared across multiple platforms.
For our synthetic tests we use Everest Ultimate, SiSoft Sandra, FutureMark's 3D Mark Vantage and PCMark Vantage, Cinebench as well as HyperPi. Each of these covers a different aspect of performance or a different angle of a certain type of performance.
CPU Raw Performance
For CPU Raw Performance we want to look at the theoretical performance numbers. This means how many GigaFlops you can get. We also test for memory bandwidth. As memory controllers are moved onto the CPU and away from the Northbridge, we see memory performance increasing but also becoming much more CPU dependent than motherboard dependent.
To test memory and raw CPU performance, we use a combination of SiSoft Sandra and HyperPi 0.99.
Version and / or Patch Used: 2009 SP3c
Developer Homepage: http://www.sisoftware.net
Buy It Here
You can see that in terms of sheer CPU power Intel has the upper hand here even comparing dollar-for-dollar performance.
Version and / or Patch Used: 0.99
Developer Homepage: http://www.virgilioborges.com.br
Product Homepage: http://www.virgilioborges.com.br
Download It Here
HyperPi is a front end for SuperPi that allows for multiple concurrent instances of SuperPi to be run on each core recognized by the system. It is very dependent on CPU to memory to HDD speed. The faster these components the faster it is able to figure out the number Pi to the selected length. For our testing we use the 32M run. This means that each of the four physical and four logical cores for the i7 (four total on the PII x4 955 and Core i5) is trying to calculate the number Pi out to 32 million decimal places. Each "run" is a comparative to ensure accuracy, and any stability or performance issues in the loop mentioned above will cause errors in calculation.
Okay, now this sort of surprised me here. In fact I re-ran the test four times to make sure I was getting the right results. The Core i5 stomps on everyone in this test. My only thought here is that the overhead of running two counts of HyperPi on each core is simply too much. So when you cut it down to one each, then we see the real power behind the Nehalem architecture.