We proceed with our series of articles devoted to analyzing performance of modern CPUs in real applications and finding out the effect of processor features on performance. Today we shall redo the second article and analyze the effect of memory subsystem on performance. This time on the example of Intel Core i7.
Testbeds
- CPU: Intel Core i7 950
- Cooler: ASUS Triton 81
- Motherboard: ASUS P6T SE (Intel X58)
- Memory: 3 x 2GB Corsair DDR3-1800 in DDR3-1600 mode
- Graphics card: Palit GeForce GTX 275
- PSU: Cooler Master Real Power M1000
Tests
This time we expanded the memory frequency range to test the most memory clock rates available: from DDR3-800 to DDR3-1600. The former does not exist as such, but we still conducted such a test for the sake of an extra dot on each diagram. We also wanted to find out how the number of channels of the Core i7 memory controller affected performance. Obviously, the fastest mode would be the one with all three channels enabled. However, this requires 3 [identical] memory modules, and not every user can afford that many. Thus, we wanted to see how much you could lose, using just 2 channels.
The problem is that our test method doesn't provide for using 4GB of RAM -- the standard amount is 6GB. Of course, you can install 4GB into one channel and 2GB into another, however, in this case, the 4GB will be accessed in dual-channel mode, while the remaining 2GB will be used in single-channel mode (the Flex Memory technology is still here, even after the memory controller has been moved to the CPU). Thus, performance might decrease not only because of the dual-channel access, but also because of the single-channel access to certain memory areas. The latter assumption is the one we have checked. If we use only 2 x 2GB of RAM, the memory will be accessed in the dual-channel mode for sure. But performance might still decrease because the total amount is reduced by one third.
We can't say our solution was perfect, but it greatly simplified making reasonable assumptions. We decided to test both modes -- 2 channels and 6GB, 2 channels and 4GB -- and compare the results to the reference (3 channels and 6GB) and see if any assumptions could be made. The DDR3-1333 memory we had complied with the "spirit" of the test -- today, it's the most popular, and more or less affordable, DDR3 memory in the market.
After a few attempts to find out the effect of timings on performance, we decided to exclude these tests (known to be the most latency-critical) due to their obvious uselessness. Thus, with regard to this issue, we can say that Core i7 does the same as Phenom II X4.
Traditionally, for our most curious readers, we provide a Microsoft Excel spreadsheet with the detailed test results. There are two additional tabs for your convenience: Compare #1 and Compare #2. Like the tables in this article, these have the four systems compared percentagewise. In Compare #1 percentages are calculated in the same way as in the tables -- relatively to the previous system. In Compare #2 all systems are compared to the reference one (1.86 GHz).
3D visualization
| |
DDR3-800 |
DDR3-1066 |
DDR3-1333 |
DDR3-1600 |
| 3ds max ↑* |
16.2 |
16.96 |
5% |
17.35 |
2% |
18.01 |
4% |
| Lightwave ↓ |
13.05 |
12.93 |
1% |
12.63 |
2% |
12.57 |
0% |
| Maya ↑ |
3.96 |
4.3 |
9% |
4.3 |
0% |
4.58 |
7% |
| SolidWorks ↓ |
52.7 |
51.14 |
3% |
48.95 |
4% |
47.93 |
2% |
| Pro/ENGINEER ↓ |
1048 |
1012 |
4% |
1004 |
1% |
979 |
3% |
| UGS NX ↑ |
2.62 |
2.76 |
5% |
2.89 |
5% |
3.05 |
6% |
| Group Score ↑ |
135 |
141 |
4% |
144 |
2% |
149 |
3% |
| |
DDR3-1333
3-ch. 6GB |
DDR3-1333
2-ch. 4GB |
DDR3-1333
2-ch. 6GB |
| 3ds max ↑* |
17.35 |
14.58 |
-16% |
14.16 |
-18% |
| Lightwave ↓ |
12.63 |
15.33 |
-18% |
15.39 |
-18% |
| Maya ↑ |
4.3 |
3.56 |
-17% |
3.55 |
-17% |
| SolidWorks ↓ |
48.95 |
55.69 |
-12% |
55.88 |
-12% |
| Pro/ENGINEER ↓ |
1004 |
1206 |
-17% |
1076 |
-7% |
| UGS NX ↑ |
2.89 |
2.89 |
0% |
2.89 |
0% |
| Group Score ↑ |
144 |
125 |
-13% |
126 |
-13% |
The up arrow (↑) marks tests, where the highest results are the best, the down arrow (↓) marks tests, where the lowest results are the best.
The first article dedicated to this topic doubted if memory frequency could affect performance considerably. The second article, featuring Intel Core i7, the key rival of AMD Phenom II X4, is not likely to prove the contrary. Yes, performance does grow linearly, but that's far from being adequate to the difference in prices for DDR3-1333 and DDR3-1600.
The dual-channel part is more interesting. According to the second table, there's a serious performance drop which is similar on both dual-channel machines (average an application-wise). This lets us assume that the performance drop is caused by switching to the dual-channel mode, not by reducting the amount of memory. There's only one subtest where the performance drop is different on the two dual-channel systems -- Pro/ENGINEER. Having compared the results, we can assume that 4GB of RAM is not enough for this subtest.
Write a comment below. No registration needed!
