Intel Core i9-11900K vs. AMD Ryzen 9 5900X. Who’s the winner?

Methodology: performance tests

Have you already created a perfect picture of the current generation of desktop processors or are you still willing to reconsider some things? For the start of HWC processor tests, we’ve picked the hottest hardware from Intel’s mainstream offer – the Core i9-11900K, which is going to compete against the AMD Ryzen 9 5900X in 40 chapters full of information. The older Core i9-10900K has also successfully been re-tested by the new methodology.

Gaming tests

We test performance in games in four resolutions with different graphics settings. To warm up, there is more or less a theoretical resolution of 1280 × 720 px. We had been tweaking graphics settings for this resolution for a long time. We finally decided to go for the lowest possible (Low, Lowest, Ultra Low, …) settings that a game allows.

One could argue that a processor does not calculate how many objects are drawn in such settings (so-called draw calls). However, with high detail at this very low resolution, there was not much difference in performance compared to FHD (which we also test). On the contrary, the GPU load was clearly higher, and this impractical setting should demonstrate the performance of a processor with the lowest possible participation of a graphics card.

At higher resolutions, high settings (for FHD and QHD) and highest (for UHD) are used. In Full HD it’s usually with Anti-Aliasing turned off, but overall, these are relatively practical settings that are commonly used.

The selection of games was made considering the diversity of genres, player popularity and processor performance requirements. For a complete list, see Chapters 7–16. A built-in benchmark is used when a game has one, otherwise we have created our own scenes, which we always repeat with each processor in the same way. We use OCAT to record fps, or the times of individual frames, from which fps are then calculated, and FLAT to analyze CSV. Both were developed by the author of articles (and videos) from GPUreport.cz. For the highest possible accuracy, all runs are repeated three times and the average values of average and minimum fps are drawn in the graphs. These multiple repetitions also apply to non-gaming tests.

Computing tests

Let’s start lightly with PCMark 10, which tests more than sixty sub-tasks in various applications as part of a complete set of “benchmarks for a modern office”. It then sorts them into fewer thematic categories and for the best possible overview we include the gained points from them in the graphs. We then have the total score for single and multithreaded performance from Geekbench 5. Lighter test tasks are also represented by tests in a web browser – Speedometer and Octane. Other tests usually represent higher load or are aimed at advanced users.

We test the 3D rendering performance in Cinebench. In R20, where the results are more widespread, but mainly in R23. Rendering in this version takes longer with each processor, cycles of at least ten minutes. We also test 3D rendering in Blender, with the Cycles render in the BMW and Classroom projects. You can also compare the latter with the test results of graphics cards (contains the same number of tiles).

We test how processors perform in video editing in Adobe Premiere Pro and DaVinci Resolve Studio 17. We use a PugetBench plugin, which deals with all the tasks you may encounter when editing videos. We also use PugetBench services in Adobe After Effects, where the performance of creating graphic effects is tested. Some subtasks use GPU acceleration, but we never turn it off, as no one will do it in practice. Some things don’t even work without GPU acceleration, but on the contrary, it’s interesting to see that the performance in the tasks accelerated by the graphics card also varies as some operations are still serviced by the CPU.

We test video encoding in HandBrake and benchmarks (x264 HD and HWBot x265). x264 HD benchmark works in 32-bit mode (we did not manage to run 64-bit consistently on W10 and in general on newer OS’s it may be unstable and show errors in video). In HandBrake we use the x264 processor encoder for AVC and x265 for HEVC. Detailed settings of individual profiles can be found in the corresponding chapter 25. In addition to video, we also encode audio, where all the details are also stated in the chapter of these tests. Gamers who record their gameplay on video can also have to do with the performance of processor encoders. Therefore, we also test the performance of “processor broadcasting” in two popular applications OBS Studio and Xsplit.

We also have two chapters dedicated to photo editing performance. Adobe has a separate one, where we test Photoshop via PugetBench. However, we do not use PugetBench in Lightroom, because it requires various OS modifications for stable operation, and overall we rather avoided it (due to the higher risk of complications) and create our own test scenes. Both are CPU intensive, whether it’s exporting RAW files to 16-bit TIFF with ProPhotoRGB color space or generating 1:1 thumbnails of 42 lossless CR2 photos.

However, we also have several alternative photo editing applications in which we test CPU performance. These include Affinity Photo, in which we use a built-in benchmark, or XnViewMP for batch photo editing or ZPS X. Of the truly modern ones, there are three Topaz Labz applications that use AI algorithms. DeNoise AI, Gigapixel AI and Sharpen AI. Topaz Labs often and happily compares its results with Adobe applications (Photoshop and Lightroom) and boasts of better results. So we’ll see, maybe we’ll get into it from the image point of view sometime. In processor tests, however, we are primarily focused on performance.

We test compression and decompression performance in WinRAR, 7-Zip and Aida64 (Zlib) benchmarks, decryption in TrueCrypt and Aida64, where in addition to AES there are also SHA3 tests. In Aida64, we also test FPU in the chapter of mathematical calculations. From this category you may also be interested in the results of Stockfish 13 and the number of chess combinations achieved per unit time. We perform many tests that can be included in the category of mathematics in SPECworkstation 3.1. It is a set of professional applications extending to various simulations, such as LAMMPS or NAMD, which are molecular simulators. A detailed description of the tests from SPECworkstation 3.1 can be found at spec.org. We do not test 7-zip, Blender and HandBrake from the list for redundancy, because we test performance in them separately in applications. A detailed listing of SPECWS results usually represents times or fps, but we graph “SPEC ratio”, which represents gained points—higher means better.

Processor settings…

We test processors in the default settings, without active PBO2 (AMD) or ABT (Intel) technologies, but naturally with active XMP 2.0.

… and app updates

The tests should also take into account that, over time, individual updates may affect performance comparisons. Some applications are used in portable versions, which are not updated or can be kept on a stable version, but this is not the case for some others. Typically, games update over time. On the other hand, even intentional obsolescence (and testing something out of date that already behaves differently) would not be entirely the way to go.

In short, just take into account that the accuracy of the results you are comparing decreases a bit over time. To make this analysis easier for you, we indicate when each processor was tested. You can find this in the dialog box, where there is information about the test date of each processor. This dialog box appears in interactive graphs, just hover the mouse cursor over any bar.


  •  
  •  
  •  
Flattr this!

Leave a Reply

Your email address will not be published. Required fields are marked *