ComputerBase has updated the test course for graphics cards again. At the beginning, 23 graphics cards from AMD and Nvidia had to prove themselves in the new Ryzen 5000 test system. The results show shifts that also affect Ampere and RDNA 2. It is also clear: 6 GB of storage are at the end.
Even if the critical delivery situation and massively inflated prices for graphics cards remain in their minds as a reminder of the year 2020 for many PC gamers, the graphics card market was also technically and with a view to the competition between AMD and Nvidia more exciting than ever before.
New games, drivers, Ryzen 5000 and Resizable BAR
Nvidia had presented the GeForce RTX 3000 with Ampere and raised the performance bar a good bit higher than its predecessor. And then came AMD’s Radeon RX 6000 with RDNA 2. With a giant leap in terms of performance, efficiency and features, gamers had the choice of a GeForce or a Radeon graphics card for the first time in seven years. The editorial team accompanied the new product launches, which take place almost every week, with a true graphics card test marathon.
All tests were carried out using a test course, which was only intended as a transition course from the start and primarily contained results from high-end graphics cards. A Ryzen 9 3900 XT was used as the CPU. After the first seven tested graphics cards of the new generation, the transition course is already obsolete after only five months. New games, new drivers, new BIOS versions and the launch of Ryzen 5000 as the fastest gaming PC platform also speak in favor of this step.
In this article ComputerBase presents what exactly has been changed and what the results and the differences are.
The total of 17 games tested included 10 new entries, so there are more new entries than long-established titles. And since the games used have a strong influence on the benchmark results, the potential for major changes in the rating is high.
In the past five months there have been numerous new games, especially with the blockbuster products there has been one new release after the other. Assassin’s Creed Valhalla, Call of Duty: Black Ops Cold War, Cyberpunk 2077 and Watch Dogs: Legion have made it into the course. The other new releases are all a size smaller in terms of production costs, but still have absolutely high quality cast with Dirt 5, Hitman 3, Mafia: Definitive Edition, Serious Sam 4 and Star Wars: Squadrons. And with Gears 5, a title is now represented in the course that had actually long since passed. But since the PC version has received the same new graphics optimizations as the version for the Xbox Series X, there is an unexpected comeback.
Ten new games also means that ten other titles had to say goodbye. The long-running Shadow of the Tomb Raider is no longer part of the test course and Modern Warfare has been replaced by the new CoD part Cold War.
New graphics settings and test scenes
The old test course was explicitly designed for high-end graphics cards, while the new one also takes into account slower models. In order not to undermine the fast specimens, the graphics settings in Full HD and WQHD remain at the maximum, but in Ultra HD the level of detail is reduced from time to time so that it does not always have to be the fastest product.
In addition, Cyberpunk 2077 uses a different test sequence for graphics card tests than before. The new savegame is much more demanding, on the publication date of the technology test for Cyberpunk 2077, the editorial staff simply hadn’t gotten to this point.
The changes to the graphics options, the test scene and a few other things mean that the results of the old with those of the new test course can only be compared to a limited extent. If you would like further information on the test procedures, you are welcome to post relevant questions in the forum thread belonging to the article.
16 on 17 games with low-level API
It felt like an eternity for low-level APIs to become established. After it looked like DirectX 12 and Vulkan would fail at the beginning, the modern interfaces for large-scale productions are now clearly in the majority. Of the 17 games, 16 titles support a low-level API, 11 of which do not use DirectX at all 11. DirectX 12 is represented 15 times, Vulkan 3 times.
In all games, the best API for the respective GPU is used. This means that DX11 can compete against DX12 or Vulkan in a game, depending on what works best on which GPU. The re is no agreement in Borderlands 3, Control, Red Dead Redemption 2 and Serious Sam 4, where different APIs are used depending on the hardware.
None of the four RTX games in the course (CoD: Black Ops Cold War, Control, Cyberpunk 2077, Watch Dogs: Legion) are tested in the standard benchmarks with activated ray tracing or DLSS.
Separate benchmarks with ray tracing and DLSS
Nevertheless, ray tracing is included in the new course. Yes, the number of games with ray tracing is still very limited. However, the hardware base that supports ray tracing has grown massively in the past few months. Because not only Nvidia’s Turing understands ray tracing in hardware, the same applies of course to the successor Ampere. And with RDNA 2, AMD can now also accelerate ray tracing in hardware. The same goes for the new PlayStation 5 and Xbox Series X / S consoles. Or to put it another way: every new piece of hardware for “big” games now masters ray tracing.
Ray tracing therefore also plays a role in computer-based graphics card tests. However, a strict distinction is made between classic rasterizer and modern ray tracing benchmarks – the results are not mixed up. This applies to both the individual game benchmarks and the performance ratings.
Hybrid AAA games are in focus
All graphics cards that support ray tracing in hardware are now also tested with ray tracing. The focus is on the new rasterizer-raytracing hybrid titles, while pure pathtracing titles are not used because they are still a rare special case. Call of Duty: Black Ops Cold War, Control and Watch Dogs: Legion serve as ray tracing games. Cyberpunk 2077 is also tested with ray tracing on Nvidia graphics cards, but is not included in the RT rating because ray tracing does not work on AMD GPUs in the game.
DLSS is becoming more and more relevant
In the original version, DLSS was more of a tragedy than a useful feature. But over time and especially with the last major update, it has made a huge leap forward in terms of quality. In the meantime, DLSS 2.0 (test) manages to surprise with a better quality in individual titles than is possible with native resolution, or to achieve an overall equivalent quality.
However, DLSS is a pure Nvidia technology, and AI upsampling does not work on graphics cards other than GeForce RTX. However, as DLSS becomes more relevant, it is no longer ignored in graphics card tests. The feature is only taken into account in tests of a GeForce RTX graphics card; the benchmarks will not be found in Radeon articles. And only compatible graphics cards take part in the corresponding article section – there will be no comparisons of a Radeon without DLSS with a GeForce plus DLSS according to the current status.
CoD: Black Ops Cold War, Control, Cyberpunk 2077 and Watch Dogs: Legion are the games tested with DLSS. Depending on the test model, WQHD or Ultra HD is selected as the target resolution, the DLSS preset is consistently “quality” and therefore the most optically attractive variant. Ray tracing is activated in parallel in the titles.
New test system with AMD Ryzen 9 5950X
In the last GPU test system, the choice of CPU was a compromise. The Ryzen 9 3900XT offered decent gaming performance, but above all a modern platform with PCIe 4.0. Intel’s CPU competition, on the other hand, was able to score with more gaming performance, but the platform with PCIe 3.0 was no longer modern. Accordingly, it was clear from the start that the Zen 2 processor was only a short detour.
AMD’s Ryzen 5000 with Zen 3 now combines the modern platform with the highest gaming performance and accordingly it is currently the best gaming CPU and also the best benchmark platform. The Ryzen 9 5950X and thus AMD’s flagship is used by the editorial team with the factory settings.
The 32 GB DDR4 memory consists of four 8 GB modules that are operated with DDR4-3600 and optimized timing. Windows 10 20H2 and some games are installed on an M.2 NVMe SSD with PCIe 4.0, while others have to make do with an NVMe according to the PCIe 3.0 standard.
From now on with Resizable BAR / SAM
AMD’s Radeon RX 6000 graphics cards already support PCIe Resizable BAR alias AMD Smart Access Memory (test) and Nvidia’s GeForce RTX 3000 offshoots will follow in March. And that’s a good thing, because benchmarks show that the new feature can bring a performance boost, especially in low, but even in high resolutions. This is not the case in every game, and often there is little difference. In some titles on the course, the frame rate even increases by a whopping 20 percent.
For two reasons, the editorial team decided to test all graphics cards – if they are compatible – in the new test course with the Resizable BAR activated. As a result, the Radeon RX 6800, the Radeon RX 6800 XT and the Radeon RX 6900 XT are already running with the new feature, while the GeForce RTX 3000 cards are still running without the necessary BIOS updates and the driver is still missing. To demonstrate what the Resizable BAR is capable of, the Radeon RX 6800 XT was tested with and without a SAM.
If Nvidia provides the missing BIOS versions and the driver, ComputerBase will also test all Ampere models in this mode.
The performance of the games is still measured with the freely available tool CapFrameX, which relies on PresentMon for recording performance. As usual, the most important data set is the known average FPS; the percentile value of the frame rate is used as a further measured value.
The 0.2% percentile is used in most cases, but since some games suffer from irregular and non-reproducible frame output, the 1% percentile is used in these, in which the outliers play a much smaller role. This is the case in CoD: Cold War, Cyberpunk 2077, and F1 2020.
The results of the three games will still be included in the 0.2% percentile rating.
In the new test course, only reference or Founders Edition models are used for the benchmarks that are not overclocked and adhere to AMD’s and Nvidia’s specifications with regard to maximum power consumption. This means that almost all custom designs are at least as fast or a little faster than the variants tested here. But the differences are no longer great.
For a better performance classification, the following table shows the average clock rates achieved in Doom Eternal. The se are measured in 3,840 × 2,160 with maximum graphic details for 30 minutes, but the specified section only shows the last 10 minutes of the measurement series.
The scenario is very demanding for the graphics card, so the clock rates are higher in most of the other test scenes.
On the next page: A comparison of the new and old test course