Comparison of the performance of Intel processors of different generations

To do this, I decided to test the performance of similar processors of different generations. I decided to compare the models of the mass segment: 8-core processors 2660, 2670, 2640V2, 2650V2, 2630V3 and 2620V4. Testing with such a generational spread is not entirely fair, because between V2 and V3 there is a different chipset, a new generation memory with a higher frequency, and most importantly - there are no direct peers in frequency among models of all 4 generations. But, in any case, this study will help to understand the extent to which the performance of new processors in real applications and synthetic tests has grown.
The selected processor line has many similar parameters : the same number of cores and threads, 20 MB SmartCache, 8 GT / s QPI (except 2640V2) and the number of PCI-E lines is 40.
To assess the appropriateness of testing all processors, I turned to the PassMark test results .
Below is a summary graph of the results:

Since the frequency is significantly different, comparing the results is not entirely correct. But despite this, the conclusions
suggest themselves : 1. 2660 is equivalent in performance to 2620V4
2. 2670 is superior in performance to 2620V4 (obviously, due to the frequency)
3. 2640V2 sags, and 2650V2 hits everyone (also because of the frequency)
I divided the result on the frequency and got a certain performance value at 1 GHz:

Here the results were more interesting and clear:
1. 2660 and 2670 - an unexpected run for me within the framework of one generation, 2670 justifies only that its overall performance is very high
2. 2640V2 and 2650V2 - a very strange low result, which is worse than that of 2660
3. 2630V3 and 2620V4 - the only one logical growth (apparently just due to the new architecture ...)
After analyzing the result, I decided to weed out some uninteresting models that do not have value for further testing:
1. 2640V2 and 2650V2 - the intermediate generation, and not very successful, in my opinion - I remove from candidates
2. 2630V3 - excellent result at, but costs unreasonably more expensive than 2620V4, given similar performance and, moreover, this is already a passing generation of processors
3. 2620V4 - adequate price (comparing with 2630V3), high performance and, most importantly - this is the only model of the latest 8-core processor with Hyper-threading in our list, so we definitely leave
4. 2660 and 2670 for further tests - an excellent result in comparison with 2620V4. In my opinion, it is the comparison of the first and last (at the moment) generation in the Intel Xeon E5 line that is of particular interest. In addition, we have enough stocks of first-generation processors in our warehouse, so this comparison is very important for us.
The cost of servers based on 2660 and 2620V4 processors can differ by up to 2 times not in favor of the latter, therefore, comparing their performance and choosing a server on V1 processors can significantly reduce the budget for buying a new server. But I will talk about this proposal after the test results.
For testing, 3 stands were assembled:
1. 2 x Xeon E5-2660, 8 x 8Gb DDR3 ECC REG 1333, SSD Intel Enterprise 150Gb
2. 2 x Xeon E5-2670, 8 x 8Gb DDR3 ECC REG 1333, SSD Intel Enterprise 150Gb
3 . 2 x Xeon E5-2620V4, 8 x 8Gb DDR4 ECC REG 2133, SSD Intel Enterprise 150Gb
PassMark PerformanceTest 9.0
When selecting processors for tests, I already used the results of synthetic tests, but now it’s interesting to compare these models in more detail. The comparison was made in groups: 1st generation versus 4th.

A more detailed test report allows us to draw some conclusions:
1. Mathematics, incl. and floating point, mainly depends on the frequency. The difference in 100 MHz allowed the 2660 to get ahead of the 2620V4 in computational operations, in encryption and compression (and this despite the significant difference in memory frequency)
2. Physics and calculations using advanced instructions on the new architecture perform better, despite the low frequency
3. Well, of course, the test using memory passed in favor of V4 processors, since in this case different generations of memory competed - DDR4 and DDR3.
It was synthetics. Let's see what specialized benchmarks and real applications will show.
7ZIP archiver

Here the results overlap with the previous test - a direct link to the processor frequency. It doesn’t matter that a slower memory is installed - V1 processors confidently take the lead in frequency.
CINEBENCH R15
CINEBENCH is a benchmark for evaluating the performance of a computer for working with a professional animation program MAXON Cinema 4D.

Xeon E5-2670 pulled in frequency and beat 2620V4. But the E5-2660, which has a not so visible advantage in frequency, lost to the 4th generation processor. Hence the conclusion - this software uses useful additions to the new architecture (although it is possible the whole thing is in the memory ...), but not so much that this is a decisive factor.
3DS MAX + V-Ray
To assess the performance of processors when rendering in a real application, I took a bunch: 3ds Max 2016 + V-ray 3.4 + a real scene with several light sources, mirror and transparent materials, and an environment map.

The results were similar to CINEBENCH: Xeon E5-2670 showed the lowest rendering time, and 2660 could not get around 2620V4.
1C: SQL / File
In conclusion, I enclose the gilev test results for 1C.

When testing a database with file access, the E5-2620V4 processor confidently leads. The table shows the average values of 20 runs of the same test. The difference between the results of each stand in the case of the file base was no more than 2%.
The single-threaded test of the SQL database showed very strange results. The difference turned out to be insignificant, given the different frequencies for 2660 and 2670, and different frequencies for DDR3 and DDR4. There was an attempt to optimize the SQL settings, but the results were worse than they were, so I decided to test all the stands on the basic settings.

The results of the multi-threaded SQL test turned out to be even more strange and contradictory. The maximum speed of 1 stream in MB / s was equivalent to the performance index in the previous single-threaded test.
The next parameter was the maximum speed (of all flows) - the result was almost identical for all stands. Since the results of different runs fluctuated strongly (+ -5%) - sometimes they were at different stands with a significant margin both in one direction and the other. The identical average results of a multithreaded SQL test lead me to 3 thoughts:
1. This situation was caused by an unoptimized configuration of SQL
2. The SSD became a bottleneck in the system and did not allow processors to overclock
3. There are almost no differences between the memory and processor frequencies for these tasks (which is extremely unlikely)
If you have reliable explanations for such results, please share them in the comments.
Also turned out to be inexplicable result on the parameter "Recommended number of users." The average result for 2660 was higher than all - and this is with the low results of all tests.
On this issue I will also be glad to see your comments.
conclusions
The results of several versatile computational tests showed that the processor frequency in most cases turned out to be more important than generation, architecture, and even memory frequency. Of course there is modern software that uses all the improvements of the new architecture. For example, video transcoding is sometimes performed including using instructions AVX2.0, but this is specialized software - and most server applications are still tied to the number and frequency of cores.
Of course, I do not declare that there is absolutely no difference between the processors, I just want to note that for certain applications there is no sense in a “planned” transition to a new generation.
If you do not agree with me or you have suggestions for testing, the stands have not yet been dismantled, and I will be happy to test your tasks.
Economic benefit
As I wrote at the beginning of the article, we offer a line of servers based on the first generation Xeon E5 processors, which are significantly cheaper than the servers on the E5-2620V4 at a cost.
These are the same new servers (not to be confused with used) with a 3-year warranty.
Below I give an approximate calculation:
STSS Flagman RX227.4-008LH server in a configuration of 2 x Intel Xeon E5-2620V4 + 8 x 8Gb DDR4 ECC REG at retail costs today 265,065 rubles.
A similar configuration of STSS Flagman EX227.3-008LH based on 2 x Intel Xeon E5-2660 + 8 x 8Gb DDR3 ECC REG on the stock is available for 175275r.
Habr readers can receive an additional 5% discount when ordering. To do this, you need to select the desired form factor of the case from the list of models on our website . The EX217.3-004LH model is made in a 1U-case, the EX227.3-008LH - 2U, and the EX240.3-008LH is built on the basis of the Tower / 4U case.
In the model’s configurator, you can select the necessary parameters of memory, disk subsystem and additional devices. When sending an application for calculation, you must specify the promotional code HABRAHABR .
Thanks for attention! I will be waiting for your comments and suggestions on the tests.
Article Writing and Testing: Usikoff
Testing 1C: sarge74