GPU Performance Test 2020/2021: 20 games, 15 GPU's, 4 resolutions - in-depth benchmarking

PC games and hardware are constantly evolving - the same is valid for our benchmarks. With 20 years of expertise, PCGH's aim is to provide you with the best, most accurate benchmarks on the internet. To make the data accessible to a wider audience, we offer an article like this in English for the first time. As we expect the new graphics chips based on Nvidia's Ampere and AMD's RDNA2 soon, we had to balance a lot of thoughts and preparations, which is why we explain the new GPU performance charts 2020/2021 in detail - of course including first results. 3.600+ measurements have been done until today to give you the most comprehensive and robust performance charts. Come and join us on the journey of GPU benchmarks, powered by the well-known "German precision".

Disclaimer on the translation: Since English isn't our native tongue, please look generously over "bumpy" sentences. Most of the text was translated from German to English with the help of www.DeepL.com/Translator. We hope that you still find this article useful and are looking forward to your feedback!

Testing methodology explained

Every 12 months, we're re-imagining the "GPU Leistungsindex" - better known as the GPU performance charts. This is never done recklessly since the new benchmark setup & selection should last a whole year and draw a realistic picture for that time. According to our philosophy, a good benchmark chart should capture every nuance of current PC games. This includes the idea of mapping as many engines as possible and thus as many "attitudes" as possible. Some games run better on a Radeon and others better with a Geforce - that's not a secret! Who would like to have 20 tests where the ranking is always the same? Versatility between the benchmarks is good and exciting.

The PCGH benchmark committee has spent the past months on validation. More than 40 new PC games were on our list. We have constantly monitored their suitability, popularity, performance and maturation through patches over the past months and mixed them with brand new games. The final mix could hardly be more colourful and holds interesting data for every player. Before we head over to the benchmark bars, we'd like to explain in detail what we test, how we test and why we do it exactly this way. Let's start with the complete list of GPU benchmarks 2020-2021:

Note: PCGH emphasizes diversity, that's why every important engine is represented.

Scope & goals

PCGH prefers to test graphics cards with maxed-out details in highly GPU-bound scenarios using tailored benchmark sequences instead of generic in-game fly-by's. This results in a simple but clear rule: If a graphics card reaches fluid frame rates in our demanding tests, users don't have to fear any stuttering in the remaining levels or scenarios. Doing this, we're strictly trying to avoid the CPU limit since that is completely misplaced while doing GPU benchmarks. Although our readers regularly ask us to test the graphics cards also with slower processors for "practical relevance", this is not the subject of a graphics card performance test because it hides the actual performance of the graphics cards.

Let's have a look at this problem with an example. A Titan RTX, Nvidia's flagship since the end of 2018, achieves an average of 60 FPS in a game at Ultra HD resolution. Even 120 FPS are possible using Full HD, so the scaling depends on the resolution and thus GPU performance. These frame rates are made possible by a fast CPU infrastructure which allows the graphics card to unfold. If this graphics card is now installed in an old but widespread quad-core system, the Titan RTX in Ultra HD can still achieve 58 FPS. This looks valid at first glance - but in Full HD, which only displays a quarter of the pixels, only 72 FPS are achieved. In other words: In this test, the speed of the graphics card is no longer tested, but that of the CPU. If you took these values to create a performance chart, enthusiast models couldn't break away much from mid-range cards.

Of course, everyone wants to see exactly his or her system in the benchmarks. But PC hardware can be combined in many different ways and the limits shift accordingly. We therefore do not change our view on testing graphics cards within the GPU limit (if possible). So you can see at a glance which graphics card achieves which performance. With the recently revamped benchmark setup, we are even better able to determine the ideal pairings of GPU and CPU. For this purpose, some games are included in both the GPU and upcoming CPU performance parcours. To get the complement, it is not always useful to use the same test sequence. Many games are both GPU- and CPU-heavy - but not at the same time, the extremes alternate depending on the area. For example, cities with many objects and characters usually generate a high processor load, whereas an overgrown surrounding area usually attracts attention with graphics load. If you pair the measured values, you get the perfect CPU-GPU combinations that can cope with all scenarios.

Hand-crafted sequences

As usual, we use our own savegames with special choreographies to ensure that our goals are achieved. For example, we rarely simply walk straight ahead, even though this is wonderfully reproducible. Usually we implement a bunch of turns to make the streaming systems of the games (and thus the hardware) sweat a little more. This automatically leads to greater variance of the frametimes.

Each of our scenes has been recreated over several weeks, sometimes months, by different PCGH testers with different game versions. Only games with the highest reliability and flawless results qualify for the regular benchmark charts. Titles with weekly updates, server downtimes or erratic results are not suitable as permanent benchmarks. All pure multiplayer titles share this fate. The high chaos factor, dynamic processes, changing weather effects and events as well as irregular server maintenance lead to the fact that we test such games only sporadically and then in a row. The won values are usually outdated after a short time. If such a game were in the performance index, we would not be able to get out of the retests. The alternative, the mixing of old and new values, is out of the question.

PC Games Hardware emphasizes transparency and trust in our benchmarks. To achieve this, we have been providing detailed benchmark instructions for years, so that you can reproduce our tests at home. A crucial piece of the new Performance Index is this article: Benchmark FAQ 2020/2021: Borderlands 3, Death Stranding, Red Dead Redemption 2 and 17 more - Savegames included. Although it is in German, you can find all 20 testing sequences in there, accompanied with videos that show the settings and procedure. We would like to provide you with all savegames, but in times of online constraints and micro-transactions, this is not possible for many games. If you have questions, please don't hesitate to leave a comment.

Real-world frequencies

Without exception, every GPU (and CPU) benchmark done by us is preceded by a warm-up sequence: Each graphics card is put under load for five minutes in the game to be measured before the first benchmark, which results in realistic GPU boost frequencies. This is necessary due to the automatic boost mechanisms found in every modern GPU. Their opportunistic behaviour leads to higher clock speeds in cold state. Benchmarks "straight from the box" without any heating-up phase do not show the real (sustained) performance. The average frequency of each GPU can be found in the following benchmarks under the card's name.

Source: PC Games Hardware In the test we only use reference cards from AMD and Nvidia to draw a baseline. Many years ago we decided to use reference cards for performance charts like this one - and we always end up back there. At first glance, this seems far from practical, as most gamers prefer custom designs (with factory overclocking), but there are good reasons. The first, as banal as it may sound, is that we have to draw a baseline somewhere. Reference cards are perfectly suited for this task, while it is impossible to select a "typical" AIB design that represents a specific graphics card. The choice is contestable in any case. There is another reason for the reference cards or Founders Editions respectively: They are available from the start. What would you say if Nvidia's Ampere and AMD's RDNA2 are released in a few months and we don't test them because there are no AIB models available? Custom designs always need some weeks to be done. This is not a viable option for anyone involved.

Whatever we're benchmarking, each number corresponds to an average value of three, sometimes four runs per resolution. Outliers are not evaluated. That makes at least 12 measurements per game, ergo 240 per graphic card. 15 graphics cards have already completed the new course for this article. Together with the innumerable validation runs before, this ends up in more than 3,600 measurements, each lasting 20 seconds. The index run of a single graphics card with warm-up, clock logs and loading times takes about 4.5 hours - provided you've mastered the tests and nothing goes wrong in between.

Frametimes, Percentiles, Average FPS

We all know and love the Average FPS as performance indicator, but the outliers are also important to determine the perceived performance. Since 2018, PCGH has been using percentiles instead of Minimum FPS - and we still do, but utilizing a better tool. Since this benchmark revision we have been using the powerful measurement and analysis program CapFrameX (CX for short), which was created by hard-working PCGHX members. We don't want to bother you with all the math, so let's just say this much: The percentile given by PCGH is a better alternative to the "old-school" Min Fps values because it is less susceptible to interference - and that's what matters. We will continue to specify both the average frame rate and the "P99", although the latter is now equivalent to the P1 output by CapFrameX. In addition, we have complete frametime logs for each individual measurement, but the amount of this data would fill a whole book and is therefore left out. For now.

Source: PC Games Hardware CapFrameX: Our primary tool. The CX overlay is turned off during GPU testing as well as other overlays - no matter how convenient it is.

2020, year of change

While the methodology follows proven paths, the hardware and software environment has changed dramatically since last time. 15 of 20 benchmark games are new, with release dates from mid-2019 to mid-2020. You remember our goal of always capturing every nuance of the gaming market and the different GPU architectures. Besides brand new titles like Death Stranding and F1 2020, which were released in July, mature games like Borderlands 3, Control, Doom Eternal, Ghost Recon Breakpoint and Red Dead Redemption 2 are mixed in. This "Who's Who" of Triple A games is joined by smaller and indie titles such as the Half-Life remake Black Mesa, the stealth gem Desperados 3 and the RPG Greedfall. The ensemble is completed by five old benchmark friends: The Witcher 3, Metro Exodus, Forza Horizon 4, Battlefield 5 and Anno 1800. On the hardware side, however, nothing is as it once was: For the first time in the history of the PCGH Performance Chart, we use a test system based on AMD hardware.

Future-proof enviroment

A processor limit must be excluded in graphics card tests to get valid results. After long internal discussions, we decided against an Intel Core i9-10900K, despite its undeniably enormous gaming performance. Instead, we opted for a Ryzen 9 3900X. This one underwent a long tuning procedure, in which we not only determined the fastest specimen in the PCGH test lab, but also sharpened the memory to the last subtiming. For best results, we asked Corsair to provide a potent 32 GiB memory kit with Samsung B-Dies, accompanied by a powerful watercooling loop for the CPU. The effort has paid off: Thanks to massive 4.5 GHz on all cores, our "Ryzen 3900XTX" outperforms a stock Ryzen 3900X(T) by about 20 percent. This is mainly achieved by the effective 25 percent increase in memory performance according to the AIDA64 benchmark:

AIDA64 Cache & Memory Benchmark

Ryzen 9 3900XT Ryzen 9 3900"XTX" PCGH Edition

Full-screen comparison

The use of an AMD CPU in the GPU testing system is a novelty. We rely on the Ryzen 3900X not only for the future-proof 12 cores (24 threads), but especially because of its infrastructure: With the X570 I/O hub, we can use and test all upcoming graphics cards with PCI Express 4.0. Since the new system is supposed to last for the next twelve months, the use of an outdated PCI Express 3.0 infrastructure was not an option anymore. To ensure that next-gen games run smoothly, we are also renewing the storage area with a fast PCI Express 4 SSD (NVMe 1.3 ×4). All in all, the new system is optimally equipped for the future. All details about the system can be found in the following table.

The use of Windows 10 x64 v2004 aka May Update is mandatory. Like last year, we let a few weeks and bug reports pass before we updated the benchmark systems. Graphics drivers with WDDM 2.7 support had to wait despite a long test phase, but both AMD and Nvidia have meanwhile released corresponding versions. Special attention should be paid to the "Hardware Scheduling" function, an optional feature that gives the GPU more control over memory management. HAGS, respectively GHS or HWS, so the abbreviations of the function, proved to lift FPS rates a bit on average, so that we leave it turned on with compatible graphic cards. At the moment, only Turing and Pascal based GPU's can take advantage of HAGS.

The lower, the better!

The lower-level API's DirectX 12 and Vulkan are on the rise. This is not only due to the learning processes on the developers' side, improved tools and optimized game engines, but also to Nvidia, because their ray tracing advance uses the DXR ray tracing integrated in the DirectX 12 API. The OpenGL successor Vulkan now also supports ray tracing. With the upcoming consoles and the associated cross-platform games, the lower-level interfaces will certainly continue to gain market share. The recent spread is reflected in our new benchmarks: 14 out of 20 games now offer DirectX 12 or Vulkan support - some of them even require it. In 2019/2020 there were only eight games with DX12 and Vulkan. If there is a choice, we compare the API's before the measurements and then decide which one offers the best performance.

While it was clear in 2019 that no stronger graphics cards would appear in the near future, the situation is completely different in 2020. We had to make sure that our tests included a large buffer for upcoming GPUs. To achieve this, we tested all the games and scenes in 1,280 × 720 with internally reduced render scale to analyze the CPU limit - on both a Geforce RTX 2080 Ti OC and a Radeon RX 5700 XT OC. We also simulated a future graphics card with the help of a heavily overclocked MSI Geforce RTX 2080 Ti Gaming Z Trio. You can find its data as "Geforce RTX 2080 Ultra" in the benchmarks (3DMark Firestrike Extreme result), to get a preview of what's coming in autumn.

Source: PC Games Hardware The PCGH GPU Rig 2020/2021

Share article

GPU Performance Test 2020/2021: 20 games, 15 GPU's, 4 resolutions - in-depth benchmarking In preparation for the upcoming Nvidia Ampere and AMD RDNA 2: PC Games Hardware refreshes the GPU benchmarks to provide you with well-founded information based on current games, API's and engines. After months of evaluation and more than 3.600 measurements, the GPU performance chart 2020/2021 is ready for action. This article explains our testing methodology and gives recent recommendations.

Send via email

1. Page 1 GPU-Benchmarks 2020+: Methodology
2. Page 2 GPU-Benchmarks 2020+: Results & Recap