Posted on by

GPU Test System Update for 2025

Introduction

TechPowerUp is ready for the next generation of graphics cards that are about the hit the shelves in 2025. The new year promises to be action-packed for the PC hardware industry, particularly PC graphics. NVIDIA is expected to launch its next-generation GeForce RTX 50-series Blackwell. AMD is planning its Radeon RX 9000 series RDNA 4. Intel already made a stab with its Arc B580 Battlemage, and is likely to launch more GPUs in this series. Testing this avalanche of graphics cards isn’t easy, it entails well over 100 tests per graphics card, and each new GPU launch has half a dozen custom designs from board partners. The most crucial part of the testing process is the Test System, and making sure it’s of the right configuration.

For our graphics card reviews, the Test System is a gaming PC set up to provide a consistent platform for all graphics hardware being tested. Choosing a good processor for gaming is key, as is picking the right amount of system memory, the right motherboard, and the right storage to ensure all the tests load in a flash, without hiccups, consistently. After taking a close look at what both Intel and AMD have to offer with their Core Ultra 200-series Arrow Lake and Ryzen 9000 Zen 5, we have decided to make the AMD Ryzen 7 9800X3D the heart of our Test System, but we’re also making other hardware changes.

Our game tests include Alan Wake 2, Assassin’s Creed Mirage, Baldur’s Gate 3, Black Myth Wukong, Counter-Strike 2, Cyberpunk 2077, DOOM Eternal, Dragon Age Veilguard, Elden Ring, F1 24, Ghost of Tsushima, God of War: Ragnarök, Hogwarts Legacy, Horizon: Forbidden West, Like a Dragon 8, No Rest for the Wicked, Ratchet & Clank: Rift Apart, Resident Evil 4, Silent Hill 2, Warhammer 40,000: Space Marine 2, STALKER 2: Heart of Chornobyl, Starfield, Star Wars Outlaws, and The Last of Us Part One. We will add more game tests over the course of 2025, and replace older ones as we see fit.

To prepare for the arrival of next-gen graphics cards, starting in January 2025, we retested all existing graphics cards on the new machine, and so we decided to present the performance numbers of these cards as an appetizer.

Hardware

Below are the hardware specifications of the new 2025 H1 GPU test system.

As mentioned before, with the 2025 GPU Test System we’re making the change to AMD Ryzen processors after many, many years of Intel. With the disappointing launch of Arrow Lake Core Ultra 200 Series, and the various platform difficulties, switching to Ryzen was easy. Shortly after Intel’s launch, AMD came out with the Ryzen 7 9800X3D, which 1-up’s gaming performance across the board, and it’s running on the battle-tested AM5 platform. So, as soon as the CPUs went on sale we ordered one, and failed—stock evaporated within minutes. Thanks to help from AMD and Mindfactory Germany we were able to still secure one of these highly sought-after CPUs and were able to start setting up the rig.

Earlier in 2024, we switched to MSI motherboards for our GPU Test System, and we’ve been having an excellent experience with them. That’s why it was an easy choice to go with the MSI X870E Carbon. This board gives us PCI-Express 5.0 support for the GPU x16 slot, and we can still run Gen 5 on the SSD without compromising on GPU bandwidth—let’s hope that we’re actually seeing Gen 5 capable graphics cards next year—we’re ready. I also like the POST code indicator on the board, so that it’s easy to see boot progress and diagnose any issues swiftly. The slot is metal-reinforced and looks like it can take quite some abuse over time. Even though I admittedly swap cards hundreds of times each year, probably even 1000+ times, it has never been any issue—insertion force just gets a bit softer, which I actually find nice. Last but not least, the cover on the board ensures that all those graphics card changes can’t scratch the board near the left side (has happened to us in the past).

We’ve been using Thermaltake TOUGHRAM for several years now, and it has served us very well. This time we’re using a 2x 16 GB kit running at 6200 CL28, with manual primary timings and slightly increased voltage. In our previous Raptor Lake system, we used DDR5-7200 memory. For the Ryzen system, we’re intentionally reducing the memory speed to DDR5-6200. However, we’re making sure that the memory controller runs at a 1:1 ratio. On AMD, this setup improves memory performance compared to a slower ratio, which might allow for higher memory speeds but would also increase latency within the memory controller.

Our cooling setup remains unchanged: the Arctic Liquid Freezer III paired with Arctic MX-6 thermal paste is a tried and tested combo that works very well. What definitely helps is that the Liquid Freezer uses an offset for mounting, to ensure the cold plate is better aligned with the actual heat sources on Ryzen processors.

No change in case either, the darkFlash DRX70 Mesh offers plenty of length for even the biggest graphics card, which could be relevant when we start testing RTX 5090 custom designs. I also like that the case has a vertical mount option for the GPU, so we can get you clear shots of the RGB options these cards offer.

Seasonic has been our go-to choice for the GPU Test System for like a decade. While our previous Seasonic Vertex GX 850 W has been perfectly sufficient for everything we ever threw at it, including overclocking RTX 4090, we wanted to be ready in case GPU manufacturers go crazy. That’s why I upgraded to a Seasonic Focus GX 1000 W, which on top of the added power delivery capability features a semi-passive mode, which helps keep noise levels down in the lab during those long benching sessions.

Software

  • Windows 11 was updated to 24H2
  • The AMD graphics driver used for all testing used is 24.10.1 WHQL. While that might seem old, the problem here is that the next release, 24.12.1 wasn’t until Dec 5th, at which point we were half-finished with testing Radeon cards.
  • All NVIDIA cards use 566.14 WHQL. Here, too, near the end of all testing, NVIDIA released 566.36, but that version is the first version that introduced the NVIDIA App, which turns out is slightly problematic and costs a bit of performance in some setups.
  • Intel GPUs were tested with 101.6314, thanks to a leading release cadence from Team Blue. Additionally, the Arc B580, which came out just a few days ago, uses the 101.6252 press driver.

Games

This time I asked the community if they had feedback/ideas for the game test selection. While no huge revelations came out of it, it does give me a sense of what’s important to you guys.

The following titles were removed:

  • A Plague Tale Requiem: getting old, not very popular, even though it brings a unique engine to the test mix
  • Avatar: Replaced by Star Wars Outlaws, which uses the same engine
  • Cities Skylines II: Bad game/engine resulting in very low performance, despite promises no major performance improvements, not very popular
  • Dead Space: Replaced by Dragon Age: Veilguard, which uses the same engine
  • F1 23: Replaced by F1 24
  • Lord of the Fallen: Replaced by other UE5 games
  • Remnant II: Replaced by other, newer, UE5 games
  • Spider-Man Remastered: Getting old, making space for other titles, Insomniac engine still represented by Ratchet and Clank, which is newer

The following titles were added:

  • Black Myth Wukong (UE5)
  • Dragon Age: Veilguard (Frostbite)
  • F1 24 (EGO 4.0)
  • Ghost of Tsushima (in-house engine, Nixxes port)
  • God of War Ragnarök (in-house Jetpack engine)
  • Silent Hill 2 (UE5)
  • Warhammer: Space Marine 2 (Swarm engine)
  • Stalker 2 (UE5)
  • Star Wars Outlaws (Snowdrop)
  • All other games have been updated to their latest available version

I considered Final Fantasy 16, but decided against it because it’s not a big commercial hit, and the engine is quite rare. We have three UE5 games in the mix now, which gives it quite some weight for the averages, but that’s reasonable I think, considering that so many games are using it already and more will come out in the near future.

No plans to add Call of Duty, because of the always-online nature, which enforces game patches that mess with performance—at any time. Also, the way Activision Blizzard is distributing it as a single title with DLCs for every game and multiple restarts just to launch the game is terrible.

Microsoft Flight Simulator 2025 launch was a disaster, they also rely too much on online, so nothing that I can test reliably.

Indiana Jones is interesting, ID engine, Vulkan, requires RT hardware. It came out too late though, but I’ll include it in next retest, in the RT titles list, replacing DOOM Eternal.

Just like in previous years, The Witcher 3 runs the DX11 version, as permanent inclusion, to represent performance in DirectX 11.

I considered dropping Hogwarts Legacy, which is getting kinda old, but decided against it, so that it can represent Unreal Engine 4, which is still very popular. I picked Hogwarts, because it is the most successful (offline single-player) Unreal Engine 4 game.

The full list of games now consists of: Alan Wake 2, Assassin’s Creed Mirage, Baldur’s Gate 3, Black Myth Wukong, Counter-Strike 2, Cyberpunk 2077, DOOM Eternal, Dragon Age Veilguard, Elden Ring, F1 24, Ghost of Tsushima, God of War Ragnarök, Hogwarts Legacy, Horizon Forbidden West, Like a Dragon 8: Infinite Wealth, No Rest for the Wicked, Ratchet and Clank, Resident Evil 4, Silent Hill 2, Warhammer 40,000: Space Marine 2, STALKER 2: Heart of Chornobyl, Starfield, Star Wars Outlaws, The Last of Us Part 1, The Witcher 3: Wild Hunt.

Ray tracing

Only relatively small changes here. Of course every game was updated to the newest version, and new titles are F1 24 (replacing F1 23) and Silent Hill 2.

GPU Compute

While not included in the B580 reviews, I have plans to expand GPU Compute testing with a test that runs GPT2 or a similar large-language model. I will keep Blender, Topaz Video AI and Stable Diffusion, just update them to newer versions.

Gaming performance results are on the next page. If you have more questions, please do let us know in the comments section of this article.