The process of reviewing desktop CPUs at PCMag carries on core traditions that date back to the establishment of PC Labs in 1984: We compare each CPU to others in its category on the basis of price, features, upgrades from previous generations, and in-house performance tests.
To evaluate performance, we use a suite of software benchmark tests and real-world applications and games, carefully chosen to highlight the strengths and weaknesses of a tested CPU to see how it handles different situations. Our desktop CPU testing breaks down into two rough classes: productivity testing, and graphics testing.
In some cases, we make use of standardized tests created by established benchmark developers. We've also created our own tests where needed. We regularly evaluate new benchmark solutions as they hit the market and overhaul our testing procedures, when appropriate, to ensure that we can accurately reflect the effects of the latest technologies.
In November 2021, in advance of the launch of the 12th Generation "Alder Lake" desktop processors from Intel, we rolled out a new suite of benchmarks, replacing those we'd been using for several years prior. This move to a new benchmark regimen came with retests of 15 different desktop CPU processors, spanning the past two years of releases from both Intel and AMD. Our current methodology is outlined below.
Productivity and Content Creation Performance Testing
Our first task is evaluating a CPU's productivity performance in scenarios reflecting everyday use, as well as heavy workloads like image rendering and video editing.
We use a small group of testbed PCs to benchmark CPUs, which are chosen according to the socket type and generation of a CPU. For example, our recent retest covering just the past few years of releases required four different motherboard models from Intel (based on LGA 2066, LGA 1156, LGA 1200, and LGA 1700 sockets), while AMD needed three (two AM4s, and an sTRX4). The one constant component in all our test runs (when integrated graphics testing isn't in play) is the high-end Nvidia GeForce RTX 3080 Ti Founders Edition, which is installed in each testbed we build.
Cinebench R23, POV-Ray 3.7
First up on the list of benchmarks we run is Maxon's CPU-crunching Cinebench R23 test. We run this test at the All Cores setting to make use of all available processor cores and threads, as well as at the Single Core setting. Derived from Maxon's Cinema 4D modeling and rendering software, Cinebench stresses the CPU rather than the GPU to render a complex image.
The result is a proprietary score indicating a PC's suitability for processor-intensive workloads, when used with software that is fully threaded. Think of it as an all-out processor deadlift.
Next up is our POV-Ray 3.7 benchmark, a reliable way to test both single-core and multi-core rendering performance for a single image. POV-Ray renders offscreen and reports back a render time in each case.
Windows 10 does not have Intel's Thread Director enabled, which is why some 12th Gen chips we've tested appear slower in the POV-Ray and Handbrake (below) tests. These issues are not present when the same benchmark is run on a Windows 11 build. We note the operating system differences in each CPU review.
Geekbench 5 Pro, Handbrake 1.4
Primate Labs' Geekbench Pro is another industry-common processor workout. It runs a series of CPU workloads designed to simulate real-world applications ranging from PDF rendering and speech recognition to machine learning. We record Geekbench's Multi-Core Score. (Higher numbers are better.)
Geekbench is especially handy because, unlike other common benchmark tests, it has versions for many operating systems (including Apple's macOS and iOS), which can enable valuable cross-platform comparisons.
Our HandBrake video-editing trial is another tough, threaded workout that's highly CPU-dependent and scales well as you add cores and threads. In this test, we put a stopwatch on test systems as they transcode a standard 12-minute clip of 4K video (the open-source Blender demo short movie Tears of SteelTears of Steel) to a 1080p MP4 file.
We use the HandBrake app's Fast 1080p30 preset for this timed test. Lower results (i.e., faster times) are better.
7-Zip, Blender
We also run the 7-Zip benchmark, using the common compression utility. This test measures how quickly the CPU can compress, and then decompress, a dummy file using the Lempel-Ziv-Markov chain Algorithm, or LZMA. The benchmark result is displayed in MIPS, which is a measure of million instructions per second. The more MIPS, the faster your CPU will be at both creating and extracting compressed files.
Following that, we run Blender, an open-source 3D content creation suite for modeling, animation, simulation, and compositing. We record the time it takes for Blender's built-in Cycles path tracer to render two photorealistic scenes of BMW cars. Lower times are better.
PugetBench for Photoshop and Premiere Pro
Our final productivity tests are Puget Systems' PugetBench for Photoshop and PugetBench for Premiere Pro, which are testing utilities that use Adobe's seminal content creation programs to measure a PC's aptitude for multimedia editing.
PugetBench executes a broad range of general and GPU-accelerated Photoshop and Premiere Pro tasks, including opening, resizing, rotating, and saving an image; editing video; adding filters that include Lens Correction, Smart Sharpen, and Field Blur; and more.
The PugetBench Overall Score is a numeric value based on a 50/50 split between the general tasks. (Higher numbers are better.) This benchmark exercises both the system's CPU and its graphics chip or card, as well as its memory and storage subsystems. As with POV-Ray and Handbrake, we expect higher scores for PugetBench tests on 12th Generation Intel chips with testbeds running Windows 11 than those running Windows 10, due to the assist of Thread Director.
Graphics Performance Testing
Judging graphics performance requires using tests that are challenging to every CPU, yet yield meaningful comparisons across the field. We use some benchmarks that report proprietary scores and others that measure frames per second (fps), the frequency at which the graphics hardware renders frames in a sequence, which translates to how smooth the scene looks in motion. We test performance both with a discrete graphics card (our GeForce RTX 3080 Ti staple card mentioned earlier) and with the CPU's integrated graphics, in the latter case if it is so equipped.
Discrete Graphics Testing (With GeForce Card)
The first graphics test we employ is UL's 3DMark. The 3DMark suite comprises a host of different subtests that measure relative graphics muscle by rendering sequences of highly detailed, gaming-style 3D graphics. Many of these tests emphasize particles and lighting.
We run two of 3DMark's DirectX 12 (DX12) tests on all CPUs: Night Raid, and Time Spy. The former is the more modest of the two workloads, suitable for PCs with integrated graphics. It's meant for lower-power, mainstream systems and renders at a simulated resolution of 1,920 by 1,080 pixels. Time Spy is much more demanding, suitable for high-end PCs with the latest graphics cards. It renders at a simulated resolution of 2,560 by 1,440 pixels. The test leverages features of the DX12 API, including asynchronous compute, explicit multi-adapter, and multi-threading.
Each test yields an Overall Score, which is what we report. (We don't break out the graphics and CPU scores.) Higher numbers are better.
Then we break out the games. At the end of the day it's hard to beat full retail video games for judging gaming performance. Assassin's Creed: Valhalla, F1 2021, and Tom Clancy's Rainbow Six: Siege are modern, high-fidelity and high-performance titles with built-in benchmarks that illustrate how a system handles real-world video games at various settings.
Rainbow Six: Siege and F1 2021 in particular are sensitive to both the RAM speeds and maximum single-core boost clock when run in 1080p.
We run all three games at their maximum graphics-quality presets, at both 1080p resolution and 3,840 by 2,160 pixels (4K). Our performance charts in reviews detail the exact settings used.
Integrated Graphics Testing
If a CPU we're testing comes with an integrated graphics processor (IGP) on board, we'll also run the chip through a series of less-demanding synthetic and gaming benchmarks to see how it handles without the assistance of a dedicated GPU. This matters for shoppers who will not install a dedicated graphics card in their systems, but rely on the IGP.
These IGP tests include all three of the games mentioned above (Assassin's Creed, F1, and Rainbow Six) with the quality levels shifted down to "Low," run at 1080p and 720p resolutions. We also run one more pass of 3DMark, this time selecting only the Night Raid run, which is more suited for low-end systems than Time Spy.
Overclocking and Cooling Testing
Finally, after all the benchmarks have been completed with stock profiles applied (aside from XMP II on the memory), we then run our CPUs through a 10-minute stress test in Cinebench R23 to see what their maximum thermal output is during sustained load. We record these results using the component monitoring software CoreTemp.
Once this run is complete and the result recorded, we then move on to overclocking the CPU. We will use a variety of techniques to achieve this depending on the CPU: First, all CPUs are run once on the "Automatic Overclocking" profiles available in Intel's Extreme Tuning Utility (XTU) and AMD's Ryzen Master Utility, depending on the chip maker in question. These tools evaluate everything from the model to the silicon quality before auto-applying an overclock that sits within the recommended guidance of Intel or AMD.
After this, a more aggressive set of overclocks will be manually applied at the BIOS level to see where the chip tops out before crashing during either games or productivity tasks. We achieve the highest stable overclock possible through this process, and then report the benchmark results of the games and applications that were run with the higher clock speed applied to record our final results.
All of our AMD and Intel testbeds contain closed-loop, all-in-one liquid coolers that are 240mm or larger (360mm in the case of the latest Intel 12th Generation testbed, and with AMD's Ryzen Threadripper HEDT platform). This means our results may not be representative of possible overclock profiles or temperature levels for your desktop, as variables in radiator size, fan count, case design, and silicon quality can affect the end result.
Authoritative Guides to Buying the Best CPUs
All this testing is a key part of how PC Labs informs the content and choices within our roundups and reviews on PCMag.com. Performance is just one factor in our overall picture of a CPU's worth, which also includes assessments of its feature set, value for the money, and availability in the marketplace.
Our main best CPUs guide is the place to start for a wide-ranging overview of the top choices across categories (as well as a deep explainer on all things related to CPU buying). We also have a gaming CPUs roundup for those of you who have that specific use case in mind, while those hunting for a good deal should read our guide to the best budget CPUs that you can buy at the moment. We update these guides in sync with major releases to the market, and when we bestow an Editors' Choice award to a given CPU.