Synthetics: Sandra And Euler3d
SiSoft Sandra 2012
Sandra is a fairly common diagnostic on Tom's Hardware because it's easy to use, gives us reproducible results, and very precisely isolates specific subsystems.
With identical memory modules in both systems, Sandra’s memory test results are fairly predictable. The quad-channel LGA 2011-based system shows 3.82 times the performance of the dual-channel LGA 1155-based system, which is close enough to the theoretical 4x improvement. That's impressive scaling.
The Sandra multimedia benchmark shows a similar delta. The LGA 2011-based P900DX, with four times the theoretical memory bandwidth and number of cores, shows a 3.7x lead over the P500X baseline system. Meanwhile. the arithmetic benchmark shows the P900DX with a 3.6x edge over the more entry-level workstation.
The cryptography results are slightly narrower, with the more expensive system showing a 3.87x gain in AES-256, but only a 2.9x increase in SHA-256. Because both Xeons benefit from AES-NI, they're bound by memory performance, which feeds the processors as fast as they're able to execute instructions. On the other hands, SHA-256 is bound by compute power, since the workload is software-based.
The Sandra results are slightly weighted by the LGA 1155-based machine's Ivy Bridge architecture and higher clock rate. In contrast, the P900DX operates at a lower frequency and centers on Sandy Bridge-E. That's why the higher-end workstation isn't exactly four times faster.
CASE Euler3D
Euler3D is a benchmark developed by the CASE Lab at Oklahoma State University. It simulates a Mach 0.5 airflow over a AGARD 445.6 aeroelastic test wing. This test is 32-bit because the developers feel that 64-bit compilers aren't quite as mature, and prefer to keep results standardized to the 32-bit version. However, the benchmark is fully threaded and multiprocessor-aware, automatically detecting the number of cores and launching the appropriate number of threads.
The P900DX demonstrates a 3.17x increase over the P500X. While this result is still impressive, it illustrates the less-than-4x scaling discussed previously.
