This small article touches upon a single problem: how memory performance affects the speed of a computer based on a quad-core AMD Phenom X4 9850 processor. However, it would be reasonable to assume that tendencies discovered in this article will be true for at least all Phenom X4 products.
We shall not describe our testbed configuration, software, and provide other reference materials here, because our tests are quite short and we compare only one system - it's easier to publish links. We follow the standard test method 2008 for all processors. The testbed configuration is the same as used in our Phenom X4 9850 review. We just forced memory to work in DDR2-533 mode (266 MHz real frequency) via BIOS. In the standard mode, Phenom X4 9850 works with DDR2-1066 memory. So the memory frequency was halved. Quite a change, isn't it? Enough to make test results interesting. Let's have a look.
3D Modeling and Rendering
| |
DDR2-533 |
DDR2-1066 |
Gain |
| 3ds max |
| CPU Render |
6.83 |
6.88 |
1% |
| Graphics |
3.34 |
3.36 |
1% |
| Hardware Shaders |
8.33 |
8.50 |
2% |
| Maya |
| GFX |
2.37 |
2.51 |
6% |
| CPU |
4.90 |
5.15 |
5% |
| Render |
0:05:45 |
0:05:45 |
0% |
| Lightwave |
0:09:20 |
0:09:20 |
0% |
| Group Score |
91 |
93 |
2% |
Maya is the most sensitive program, not only in CPU operations, but also in graphics (in case of Maya it's OpenGL). All three programs used in these tests almost unanimously confirm that rendering speed does not depend on memory performance (1% in 3ds max falls within a measurement error).
CAD/CAM
| |
DDR2-533 |
DDR2-1066 |
Gain |
| UGS NX |
| Total CPU |
3.28 |
3.37 |
3% |
| Total Graphics |
1.58 |
1.65 |
5% |
| Pro/ENGINEER |
| CPU-related tasks |
769 |
760 |
1% |
| Graphics-related tasks |
976 |
937 |
4% |
| SolidWorks |
| Graphics |
78.56 |
75.16 |
5% |
| CPU |
55.01 |
54 |
2% |
| Group Score |
91 |
94 |
3% |
It's another example of rare unanimity: all three programs need fast memory to accelerate graphics. By the way, as in case of Maya, we deal with OpenGL here.
Compiling
| |
DDR2-533 |
DDR2-1066 |
Gain |
| Visual Studio |
0:33:30 |
0:32:31 |
3% |
To tell the truth, this result is far from being impressive. Even though we know that this compiler loves cache.
Professional photo processing
| |
DDR2-533 |
DDR2-1066 |
Gain |
| Photoshop |
| Blur |
0:06:30 |
0:06:12 |
5% |
| Color |
0:01:24 |
0:01:24 |
0% |
| Filters |
0:06:08 |
0:06:06 |
1% |
| Light |
0:02:16 |
0:02:14 |
1% |
| Rotate |
0:02:58 |
0:02:48 |
6% |
| Sharp |
0:02:56 |
0:02:46 |
6% |
| Size |
0:01:02 |
0:01:00 |
3% |
| Transform |
0:02:30 |
0:02:22 |
6% |
| Group Score |
92 |
95 |
3% |
It's a mixed situation: Blur, Rotate, Sharp (Unsharp Mask), and Transform operations appreciate fast memory, but complex artistic filters, color space conversions, lighting effects, and resize operations are indifferent to it. Funny: theoretically, Resize should have responded to faster memory. On the whole, we get only 3% of performance gain. Very little.
Scientific applications
| |
DDR2-533 |
DDR2-1066 |
Gain |
| Maple |
0.0256 |
0.0258 |
1% |
| Mathematica |
| Internal |
2.9140 |
3.0160 |
4% |
| MMA |
1.1177 |
1.1359 |
2% |
| MATLAB |
| LU |
0.0879 |
0.0849 |
4% |
| FFT |
0.1705 |
0.1602 |
6% |
| ODE |
0.2743 |
0.2683 |
2% |
| Sparse |
1.8310 |
1.7797 |
3% |
| 2D |
0.3407 |
0.3327 |
2% |
| 3D |
0.5715 |
0.5690 |
0% |
| Group Score |
64 |
65 |
3% |
The benchmark built into Mathematica is more sensitive to memory system than MMA. If you remember, the built-in test uses parallel computations, and MMA practically does not. A coincidence? The average result is again 3%.
