ScienceMark 2.0 and Pifast
ScienceMark 2.0 leads the way.
It's a trade-off of just how much is available and how much memory bandwidth can be used. Dual-channel Intel chipsets offer up a potential 6.4GB/s at 200MHz. The ABIT IC7-MAX3 Canterwood, heavily tuned, still can only manage to extract around two-thirds of its potential. The Shuttle SB65G2 doesn't appear to feature PAT-like memory optimisations. If it had, it would have matched the Canterwood's performance. Don't be fooled by the Athlon 64-powered Shuttle XPC's low benchmark here.
Memory latency can be considered to be just as important as bandwidth. On-die memory controllers have a nasty habit of showing up the competition. More importantly for our examination is the SB62G2's higher latency when compared with the Canterwood. Higher latency and lower bandwidth, albeit not by much, will, ceteris paribus, translate into slightly lower performance.
That does indeed seem to be the case. Notice how three CPUs on three different chipsets manage to benchmark within 0.5s of each other. Remarkable. The Pifast test involves calculating the constant Pi to 10 million places.
