Linchi Shea, who is a SQL Server MVP who I got to meet at the 2007 PASS Summit has yet another good post about SQL Server 2005 I/O issues. This one measures the effect of read and write cache in disk controllers. It turns out, according to his tests, that read caches in disk controllers don’t help SQL Server very much, while write caches are much more helpful. The nice thing about Linchi is that he does not just accept conventional wisdom at face value, he believes in doing empirical testing and presenting the results.
Joe Chang, who is extremely knowledgeable about SQL Server related hardware, has a very interesting post about storage performance for SQL Server. He also has several other good recent posts about the various TPC benchmarks and SQL Server hardware sizing. I have to agree with him that the current sweet spot for most general purpose SQL Server 2005 installations is a two socket, quad-core Xeon X5460 3.16GHz, 2x6MB L2 cache, 1333MHz FSB system (using 45nm Penryn based technology). This gives you eight cores, with larges shared L2 caches, and a very fast FSB. It also minimizes your SQL license costs (since Microsoft licenses by physical CPU socket).
Something like a Dell PowerEdge 2900 III, can be filled up with 48GB of RAM, some internal drives for your OS and page file, and some HBAs to talk to your SAN, and you will have a very capable eight core machine.