For this review we are comparing two systems, a Supermicro 2U system based on Stoakley and an Intel system using the older Bensley platform (using a 65nm quad-core Clovertown). In an ideal world, we would be using MPUs with equal frequency to clearly see the improvements in Harpertown and Seaburg, but unfortunately that is not possible. Careful analysis and inference can make up for this frequency difference, but that introduces slightly more uncertainty into our estimates.
The Stoakley platform shipped with 16GB of memory in 8 DIMMs, but we elected to remove half the memory for testing, since it would give the Stoakley system an unfair advantage for many tests. SPECjbb2005 would benefit from the extra capacity and most tests would benefit from the extra bandwidth due to multiple DIMMs per channel. Recall that an FB-DIMM channel can simultaneously read and write to separate DIMMs, so two DIMMs per channel increases the available bandwidth.
Chart 1 – System Configuration
Performance for each benchmark is the average of three runs. For tests susceptible to caching, the initial run was discarded so that there were fewer penalties from compulsory cache misses and smaller run to run variation.
This time around, we have significantly expanded the number and scope of our benchmarks. We’d particularly like to thank Scott Wasson of the Tech Report for sharing MyriMatch and Euler3D. Our full set of benchmarks includes:
- POV-Ray 3.7 beta 22 (64 bit)
- FlamMap (64 bit)
- Valve VRAD
- Sungard (64 bit)
- SPECjbb2005 with BEA JRockit 6 R27.4
- XMLMark 1.1 with BEA JRockit 6 R27.4
These benchmarks span a variety of application areas: 3 that would be considered classical HPC, 2 Java based server benchmarks, and 2 content creation benchmarks.