Nehalem Performance Preview

April 7, 2009 by

Pages: 1 2 3 4 5 6 7 8 9 10 11 12

SPECfp_rate2006

While SPECint includes a wide range of workloads, SPECfp focuses on floating point code, which tends to come from the scientific world. SPECfp includes code drawn from applications such as weather modeling, fluid dynamics, quantum chemistry and speech recognition. SPECfp consists of 17 benchmarks written in C, C++, Fortran and a mixture of C and Fortran.Our testing methods are similar to those identified previously for SPECint. We used binaries supplied by Intel, with close to optimal compiler settings. All performance results presented are base, rather than peak. The SPECfp tests were run only once on both Harpertown and Nehalem, although the scores were checked for reasonableness against existing valid submissions.
Figure 24 – SPECfp_rate2006 Performance
povray above scored an incredible 921.233 on Nehalem, although the graph does not extend out that far.Comparing the Nehalem and Harpertown results in SPECfp really highlights the extent to which the platform hamstrung Intel’s prior efforts in HPC. The benchmarks in SPECfp are notoriously dependent on the memory subsystem and Nehalem delivers in spades. Nehalem’s performance is a staggering 2.5X higher than Harpertown (2.2X ignoring povray).
Figure 25 – Nehalem SPECfp_rate2006 Power Consumption I
Unfortunately, illustrating the power consumption for all the SPECfp benchmarks requires three charts instead of two.
Figure 26 – Nehalem SPECfp_rate2006 Power Consumption II
Only one more chart left to go.
Figure 27 – Nehalem SPECfp_rate2006 Power Consumption III
The power consumption in the floating point benchmarks is even more varied than their integer counter parts, with quite a few interesting patterns appearing. Calculix shows three peculiar periods where the power oscillates wildly between 340-360W, while dealii and milc have recurring patterns of power rising and falling that suggests an outer loop. What would be a fascinating experiment would be to correlate this power consumption data with a control flow graph of execution, because it’s very likely that there is a close correspondence between the two – it would be nice to be able to actually pin-point which lines of source code are responsible for increasing or decreasing the power consumption.

Pages: « Prev   1 2 3 4 5 6 7 8 9 10 11 12   Next »

Discuss (52 comments)