CPU benchmarks do matter, but it's easy to be misled by them. Every time a new CPU arrives on the market, there's a flood of benchmark charts showing how the latest chips stack up to the old guard. Maybe two chips trade blows across a series of tests, or maybe one comes out as the clear victor. It's important not to get too lost in averages and percentages, though, lest you leave with the wrong impression.

If you aren't careful, the wide swath of CPU benchmarks that are a staple among reviews can sway your buying decision, and not always in the best direction. You shouldn't throw out CPU performance data and select a chip with blind loyalty, no. But you should understand the data you're looking at and how to cut through the noise to get a CPU that's right for you.

Your CPU isn't actually doing the rendering

Cinebench is the benchmark I love to hate

Oh, Cinebench. No CPU review is complete without a run of Cinebench. It's usually the first benchmark you see when looking at a CPU's performance, serving as somewhat of a ground zero for how two CPUs compare to each other. Cinebench is a little deceptive, though. It's wonderfully consistent, and it's a great benchmark to show how two CPUs stack up against each other. It just doesn't mirror anything close to real-world performance when it comes gaming. As Maxon, who makes Cinebench, puts it, Cinebench is a tool "to test if a machine runs stable on a high CPU load, if the cooling solution of a desktop or notebook is sufficient for longer running tasks to deliver the full potential of the CPU and if a machine can handle demanding real-life 3D tasks."

Your CPU isn't doing any rendering in your games, though, and usually not in offline rendering apps, either. That's handled entirely by your graphics card. Most renderers available today are accelerated by a GPU. You'll sometimes see Blender benchmarks pop up in CPU reviews, for instance, despite the fact that even a low-end GPU with CUDA support exponentially outperforms even powerful chips.

These benchmarks serve a good purpose. It's just important to understand what that purpose is. Because CPUs are so inefficient at rendering, this type of workload represents an extreme use case. It serves to answer the question of what performance looks like for a chip when it's pushed to its limits. It's a stress test, not a real-world test. Or, at least, Cinebench isn't a real-world test unless you're using Maxon's RedShift renderer with a CPU, in which case you should probably switch over to GPU rendering.

A broad scope means no specificity

It turns out AES encryption tests aren't that relevant for PC gamers

A step down from stress tests like Cinebench and Blender, you run into general-purpose benchmarks. I'm talking about tools like PCMark and Geekbench, both of which run a multitude of different workloads to get a broad snapshot of how your CPU performs. It's a 10,000-foot view of performance, and it's important to treat it as such. For instance, Geekbench 6 includes everything from text processing and asset compression to CPU-based ray tracing. Those are all things you might do with your CPU, but if you aren't compiling applications with Clang, well, that's not super relevant to the performance you'll get out of a CPU.

It's especially not relevant when talking about games. Tests with a broad scope are great for touching on a little bit of everything; in other words, it shows you how performance scales between two different chips in general. Sometimes that can sway the final result. One of the most notorious examples of this was Geekbench 5, which included a single test for cryptography that accounted for 5% of the final score.

These broad tests aren't bad, but you need to factor in that they're looking at a broad range of tasks, some of which might not be relevant to your final performance. A dash of salt for an easily digestible number is all you need.

Small differences in architecture create big performance differences

These tests matter, but they only tell a small part of the story

With the release of Zen 5 CPUs like the Ryzen 9 9950X3D, AMD included a 512-bit data path for AVX-512 instructions. Support for these instructions has been around since Zen 4, but in an architecture that split the data across two 256-bit paths. Intel CPUs, on the other hand, don't support AVX-512 instructions natively. This is a niche instruction set that shows up in some places, namely AI applications and things like PS3 emulation. But given the sparse support, AVX-512 instructions aren't widely used, and if they are, they're used with an AVX2 fallback that's broadly supported.

There are some tests that use AVX-512 instructions, though. A great example of this is Y-Cruncher, which uses your CPU to calculate Pi. Accelerated by AVX-512 instructions, AMD's latest CPUs are leagues faster than anything Intel and AMD itself have ever put out, including the latest Core Ultra 9 285K. It's somewhat of a hollow victory, though. It looks great on a bar chart, sure, but the vast majority of people aren't using apps with AVX-512 instructions.

When you're running through a CPU review and looking at benchmarks, take the time to read all the commentary around the tests, and if it doesn't clarify what the tests are useful for, it's worth doing a quick Google search yourself. That'll give you a good idea of the tests you should pay attention to and which you can safely discard.

An apples-to-apples comparison requires ignoring real-world use cases

An accurate performance picture is, ironically, not that accurate

Stick with me for a second here. In order to paint the most accurate picture of performance in CPU reviews, the test can't mirror reality. I know that sounds counter-intuitive, but think about it. When you're running CPU benchmarks, it's important to control as many variables as possible. That means making the same BIOS adjustments, ensuring the same programs are running (if any), and marking down every OS version and update to make sure they're consistent across test runs. It also means rerunning tests, discarding outliers that, in real-world use, can still pop up.

BIOS tweaks, different driver and OS versions, and especially background applications all sway performance. If they didn't, it wouldn't be so important to control those variables in CPU reviews. You probably have background apps running on your machine, and that's not a bad thing. You're running Discord while playing games, or you have Splice open when working in a Digital Audio Workstation. Or maybe you're passing clips back and forth between Premiere Pro and After Effects. That's not to mention everything else that could be running at all times, from fan control and RGB software to the Nvidia or AMD apps.

Particularly when you're looking at results with thin margins, it's important to keep these background elements in mind. You won't see two CPUs trading places when there's a large disparity in performance, but a couple of background apps, and Windows or BIOS tweaks, can make two CPUs that offer competitive performance trade places.

All in the right context

It might sound like I'm shooting myself in the foot here as a CPU reviewer that uses most of these tests to review processors. But I hope you can see the point here. No review is individualized to you, and given how broad the tasks that a CPU can handle are, it's important to understand the data you're looking at. That way, you'll be able to laser focus on the performance areas that matter to you and what you're building a PC for.