Get best of the week

Everyday life "alchemist"


Comparison of server performance is a complex topic, causing a lot of controversy on how to conduct the comparison “correctly”. In this article, we will talk about the way that processors go before they get into our services.

Modern servers consist of many components, each of which requires its own, unique approach. Server processors are not an exception, which will be discussed later.

New in the hands


Usually, vendors release new processors immediately with rulers, but we clearly know what we want. And so it happened: Intel recently released an update, code-named Cascade Lake Refresh. The lineup contains 14 processors, but we are only interested in three of them:

  • Intel® Xeon® Gold 6240R;
  • Intel® Xeon® Gold 5218R;
  • Intel® Xeon® Silver 4214R.


It is easy to justify the interest in these models: we have successfully used their predecessors:

  • Intel® Xeon® Silver 4114;
  • Intel® Xeon® Silver 4214;
  • Intel® Xeon® Gold 6140;
  • Intel® Xeon® Gold 6240;
  • Intel® Xeon® Gold 5218.

In some cases, vendors release something completely new, without even revealing the name, and send their partners engineering samples of such components marked Confidential.


First of all, processors are tested for compatibility with the available motherboards, since motherboards may not work correctly with processors unknown to them. Fortunately, vendors work closely with motherboard manufacturers, which allows them to release BIOS updates in advance, adding support for microcodes of new processors. Since there is a tendency to increase heat dissipation due to an increase in the processing power of the processor, we carefully check also the thermal regime during testing.

After checking the performance at the physical level, we compile a comparison table based on theoretical data, and then proceed to the performance comparison.
Data on the technical specifications of Intel processors is publicly available and is available on the resource ark.intel.com
CharacteristicXeon® Gold 6140Xeon® Gold 6240Xeon® Gold 6240R
Number of Coreseighteeneighteen24
Number of threads363648
Base frequency, GHz2.302.602.40
Maximum frequency in Turbo Boost mode, GHz3.703.904.00
L3 Cache Size24.7524.7535.75
Number of UPI Lines332
TDP, Watt140150165
Changes to the 6240R compared to its predecessors are obvious: an increased frequency in Turbo Boost mode, more cores, more cache and, as a result, more heat. It is important to note that Cascade Lake Refresh processors, unlike their predecessors, are not supported on four-socket systems.

Now you can begin practical tests of performance.

"Heavy artillery"


Initially, we used the famous SPEC CPU 2017 to demonstrate the difference in performance. SPEC CPU tests are the de facto standard in the field of performance measurement. Most processor manufacturers in their presentations refer to the results obtained using this test suite.

All SPEC CPU tests are compiled from the source code for a particular “test subject,” moreover, the use of optimizing compilers such as Intel C ++ Compiler (icc) and AMD Optimizing C / C ++ Compiler (AOCC) is not forbidden, but rather welcome. This approach allows you to compare processors of different architectures, using all the power of the processor and related tools.

SPEC organization allows you to post test results on a special page. We were interested in two-socket servers with "gold". According to our requests, we found:

  • 38 results for the Xeon® Gold 6240R processor;
  • 55 results for Xeon® Gold 6240;
  • 116 results for Xeon® Gold 6140.

The median values ​​for each test are shown in the table below:
TestXeon® Gold 6140Xeon® Gold 6240Xeon® Gold 6240R
SPECspeed2017_int_base9.0810.310.3
SPECspeed2017_fp_base110.0132.0149.0
SPECrate2017_int_base197.0230.5274.0
SPECrate2017_fp_base186.0209.5241.0
Analyzing the theoretical data on processors, we can safely say that we expect and predict growth of indicators in benchmarks.

In addition to SPEC, we use other, easier to understand benchmarks.

Minimalistic approach


There are no less popular programs for measuring performance, which may not have become the standard, but have won some fame. It's about the PassMark Performance Test and GeekBench .

GeekBench (version 5 is relevant at the time of publication) specializes in measuring processor performance and has an open rating table in which you can see your own results and compare with others. In contrast to the SPEC CPU, GeekBench, in addition to calculating in integers and with a floating point, conducts cryptographic calculations involving special processor instructions .

Three popular families of operating systems based on x86_64 architecture are supported: Windows, Linux and MacOS X. GeekBench also runs on devices with ARM architecture, but only on Android and iOS, which excludes the possibility of running on servers with this architecture.

Test results are provided in points relative to the standard. In GeekBench 5, the Intel Core i3-8100 was chosen as the benchmark, whose result was taken for 1000 points. If the processor scores 2000 points, it means that it is twice as powerful as the standard.

We chose PassMark PerformanceTest as an alternative to GeekBench. Despite the fact that PerformanceTest is adapted only for Windows, it has a wider profile and tests not only the CPU, but also the RAM, disk and graphics accelerator, if any, in the system.

The main difference between GeekBench and PerformanceTest is distribution as executable files. Such tests are similar to closed source software, which can only have general optimizations and not support the new processor capabilities.

Consider the test results using GeekBench 5, which are shown in the table.
IndexXeon® Gold 6240Xeon® Gold 6240R
Single-core score10441089
Single-Core Crypto Score14421456
Single-Core Integer Score9981055
Single-Core Floating Point Score10771100
Multi-core score2620331262
Multi-Core Crypto Score2655426920
Multi-Core Integer Score2621431639
Multi-Core Floating Point Score2612131170
Compared to the SPEC CPU results, GeekBench shows a more noticeable difference in integer calculations. However, the superiority of Cascade Lake Refresh remains undeniable.

"Real" tasks


Some experts are skeptical about benchmarking performance. "Synthetics! Not significant! ” They say, and are right in their own way. Specialists of IT companies consider new processors exclusively in the context of their own systems. At the same time, requirements are made not only for the processing power of the processor, but also for the interaction with RAM and peripherals.

The variety of software products used in the IT field is huge, and each of them has many settings. Preparing a huge number of tests takes an unacceptably long time, and also allows experienced professionals to express their dissatisfaction with the configuration of the test.
New equipment is provided for testing to specialists in the framework of the Selectel LAB project , which allows specialists to conduct their own independent testing on any tasks.
We selected several tasks that allow us to evaluate the performance of both the processor and the whole north. Here is the list:

  • compilation of LLVM project with aggressive optimization;
  • calculation of one million decimal places of π;
  • recoding video with ffmpeg.

The LLVM project consists of many complex programs, so the compilation process of all the project programs perfectly parallel even on powerful servers. In our tests, we use SSDs as storage and 12 DDR4-2666 RAM dies.
TestXeon® Gold 6240Xeon® Gold 6240R
Calculation of the number π3m 20.559s3m 16.246s
Transcoding video with ffmpeg0m 46.646s0m 46.634s
Compiling LLVM project with aggressive optimization15m 35.933s15m 0.370s
Practical tests confirm the results of previously conducted "synthetic" tests. Compilation of LLVM project programs is performed in parallel, so the increase in the number of cores accelerates the process as expected. The calculation of the π number uses only one thread, and in this test you can see a small increase in speed associated with an increase in the maximum processor frequency.

Recoding video is a less trivial task, which is calculated in parallel at the discretion of the ffmpeg application itself. Such a slight difference between the processors is due to the software feature.

Instead of a conclusion


Progress does not stand still and updating the Cascade Lake line was a highly anticipated event. Our testing showed that an increase in the number of cores, an increase in the L3 cache and an increased clock frequency favorably affected the overall processor performance and allowed it to adequately cope with both “synthetics” and tasks close to real conditions.

You can try new processors in your own tasks now - servers are available for order in the control panel .

More articles:


All Articles