i7 2600K with DDR3 1600 MHz

Table of Contents Introduction Runtimes 1 core 2 cores 4 cores 8 cores Summary Inferences

Introduction

This page registers the performance achieved with the Intel i7-2600K (stock clocking), sporting 4 DDR3 modules of 4 GB each at 1600 MHz. The machine used is part of blueCAPE's IT pool.

Notes

HyperThreading was turned on in the BIOS settings.

The use of mpirun is merely as a helper application. The avxtest* binaries are not running cooperatively.

Keep in mind that these results are not statistically balanced, since they are the result after a single run.

Runtimes

These were executed on CentOS 6.5 x86_64, using a custom build of GCC 4.8.4. Built with the native options:

    g++ -O3 -march=native avxtest.cpp -o avxtest
    g++ -O3 -march=native avxtest64.cpp -o avxtest64

1 core

32-bit:

   ./avxtest.sh

• x86:
  • Time taken (ms): 26080.0
• AVX:
  • Time taken (ms): 3280.0

64-bit:

   ./avxtest64.sh

• x86_64:
  • Time taken (ms): 25750.0
• AVX:
  • Time taken (ms): 7230.0

2 cores

32-bit:

   mpirun -n 2 ./avxtest.sh

• x86:
  • Time taken (ms): 26710.0
  • Time taken (ms): 26780.0
• AVX:
  • Time taken (ms): 3390.0
  • Time taken (ms): 3400.0

64-bit:

   mpirun -n 2 ./avxtest64.sh

• x86_64:
  • Time taken (ms): 26450.0
  • Time taken (ms): 26520.0
• AVX:
  • Time taken (ms): 7350.0
  • Time taken (ms): 7400.0

4 cores

32-bit:

   mpirun -n 4 ./avxtest.sh

• x86:
  • Time taken (ms): 27850.0
  • Time taken (ms): 27900.0
  • Time taken (ms): 28000.0
  • Time taken (ms): 28060.0
• AVX:
  • Time taken (ms): 3540.0
  • Time taken (ms): 3550.0
  • Time taken (ms): 3520.0
  • Time taken (ms): 3550.0

64-bit:

   mpirun -n 4 ./avxtest64.sh

• x86_64:
  • Time taken (ms): 27420.0
  • Time taken (ms): 27500.0
  • Time taken (ms): 27530.0
  • Time taken (ms): 27660.0
• AVX:
  • Time taken (ms): 7760.0
  • Time taken (ms): 7690.0
  • Time taken (ms): 7820.0
  • Time taken (ms): 7780.0

8 cores

32-bit:

   mpirun -n 8 ./avxtest.sh

• x86:
  • Time taken (ms): 40810.0
  • Time taken (ms): 40790.0
  • Time taken (ms): 40720.0
  • Time taken (ms): 40870.0
  • Time taken (ms): 40830.0
  • Time taken (ms): 40690.0
  • Time taken (ms): 40740.0
  • Time taken (ms): 40670.0
• AVX:
  • Time taken (ms): 7210.0
  • Time taken (ms): 7250.0
  • Time taken (ms): 7220.0
  • Time taken (ms): 7200.0
  • Time taken (ms): 7010.0
  • Time taken (ms): 7260.0
  • Time taken (ms): 7110.0
  • Time taken (ms): 7150.0

64-bit:

   mpirun -n 8 ./avxtest64.sh

• x86_64:
  • Time taken (ms): 40050.0
  • Time taken (ms): 39910.0
  • Time taken (ms): 39970.0
  • Time taken (ms): 39920.0
  • Time taken (ms): 39900.0
  • Time taken (ms): 39940.0
  • Time taken (ms): 39940.0
  • Time taken (ms): 39920.0
• AVX:
  • Time taken (ms): 18580.0
  • Time taken (ms): 18440.0
  • Time taken (ms): 18360.0
  • Time taken (ms): 18660.0
  • Time taken (ms): 18150.0
  • Time taken (ms): 18650.0
  • Time taken (ms): 18700.0
  • Time taken (ms): 18550.0

Summary

Architecture/Mode	1 core	2 cores (std-dev)	4 cores (std-dev)	8 cores (std-dev)
x86 (ms)	26080.0	26745.0 (35.0)	27952.5 (82.2724133595)	40765.0 (66.3324958071)
x86_64 (ms)	25750.0	26485.0 (35.0)	27527.5 (86.4219300872)	39943.75 (44.9826355386)
AVX float (ms)	3280.0	3395.0 (5.0)	3540.0 (12.2474487139)	7176.25 (78.0924932372)
AVX double (ms)	7230.0	7375.0 (25.0)	7762.5 (47.1036092035)	18511.25 (173.812938241)
-	-	-	-	-
Core frequency (MHz) (cpufreq-aperf)	3775	3673	3469	3469
downscale ratio (c1/cx)	1	1.0277702151	1.0882098587	1.0882098587
x86	1	1.02549846626	1.07179831288	1.56307515337
x86_64	1	1.02854368932	1.06902912621	1.55121359223
AVX float	1	1.03506097561	1.07926829268	2.18788109756
AVX double	1	1.02005532503	1.07365145228	2.56033886584

Inferences

1. The downscale ratio on the x86/x86_64 calculations are all almost exactly on the mark for the downscale range, with the exception when relying on the HyperThreading feature.
   • Note: the frequencies were revised with cpufreq-aperf, to compare with the ones at cpu-world.com.
2. On this machine that has HyperThreading turned on, it was possible to assess that... this feature has a considerable good impact on improved performance, but only when simply using x86 FPU; while when using AVX, it lost a lot of performance, which implies that some tuning on the compiler (or CPU?) side is needed for better scheduling.
   • On 32-bit x86 FPU, it meant roughly 37% more processing power: 2*27952.5/40765 - 1.0 = 0.3714
   • On 64-bit x86 FPU, it meant almost 38% more processing power: 2*27527.5/39943.75 - 1.0 = 0.3783
3. For an additional reference, cpubenchmark.net gives an index of 8525 to this CPU, but keep in mind that this index accounts for HyperThreading. The estimated index without HyperThreading is at a 1.3 ratio, namely 6558.
4. An interesting result is that the 64-bit timings are only slightly faster than the 32-bit ones.