Introducing the 2025 GPU Hierarchy Testbed

Every benchmark suite begins with a selection of the appropriate testbed. For the GPU hierarchy, we want the fastest possible system to go with our graphics cards, ensuring that nothing else holds the GPU back — as much as we can, at least. The reality is that lower settings and resolutions are less demanding, so 1080p testing in particular simply won't allow the fastest cards to reach their full potential. Faster CPUs routinely come out, but often the gains are only 5–10 percent, so we want to make the most of what's available. Which brings us to our component choices.

Right now, the crown for the fastest CPU for gaming goes to the AMD Ryzen 7 9800X3D. It's "only" an 8-core, 16-thread CPU, but the number of games that truly push more than eight CPU-heavy threads can probably be counted on one hand. More importantly, having a large 64MB L3 cache stacked on top of the existing 32MB L3 cache — that's the "X3D" part of the model name — proves extremely beneficial for a lot of games. Not everything benefits to the same degree, but overall the 9800X3D generally outpaces the more expensive 16-core, 32-thread Ryzen 9 9950X3D. That's because the 9950X3D only has the extra L3 cache on one of the two 8-core chiplets, and the extra traffic between the various chips works against the higher core counts in most games. It certainly helps that the 9800X3D costs $320 less.

The motherboard is an often overlooked component, with some feeling that just about any board will get the job done. On some level that's true, but memory compatibility, system stability, features, and ease of use are all important factors. Our board of choice for this socket AM5 build is the Asus ROG Crosshair X870E Hero, a high-end board with a plethora of options. It's perhaps a bit overkill, but again we're trying to remove any and all bottlenecks.

For the memory, we're using a G.Skill Trident Z5 Neo RGB DDR5-6000 CL28 32GB kit. We could have bumped this up to a 64GB kit, but we felt this was sufficient. Maybe we'll upgrade the RAM when we look at whatever AMD's Zen 6 X3D chips bring to the table probably in early 2027. Note that the higher speed and lower latency of this particular kit ends up costing about twice as much as a basic DDR5-5600 CL38 kit, though at just $75 extra that's not really a huge ask. Still, the X3D chips do make higher memory speeds less of a factor — this kit probably only improves overall performance by 3–5 percent over a budget kit.

For storage, we wanted at least a 4TB SSD, preferrably with PCIe Gen5 support. That's a big ask, frankly, and a Gen4 drive could have shaved the cost down by about $150. Still, we wanted the "best" and that lead us to the Micron T705 4TB. We've had this system for a few months already, so there are drives that may technically be a bit faster and consume a bit less power, but they're also a bit more expensive. Overall, the Phison E26 controller at the heart of the T705 still rates as one of the fastest around.

Keeping the CPU cool isn't nearly as difficult with the 9800X3D as it would be with one of Intel's Core i9-14900K CPUs — which are still generally a bit faster in games than the newer Core Ultra 9 285K. We picked up a Cooler Master ML280 Mirror a while back for only $100, but the price has since doubled (it's now discontinued). Still, it works well, as should most other 280mm or 360mm liquid coolers from reputable brands: Cooler Master, NZXT, Corsair, and some others.

Wrapping up the core components, everything needs power, and with the potential to have GPUs slurping down over 600W all on their own, we want plenty to go around. We picked the Corsair HX1500i, and Platinum rated PSU that adheres to the latest ATX 3.1 and PCIe 5.1 specs. It used to cost closer to $400, but right now it's on sale for just $265, which is an excellent price for a high quality power supply.

We're not using a case, because we have an Nvidia PCAT v2 device that sits between the graphics card and the motherboard to capture power data — PCAT stands for Power Capture Analysis Tool. We've tested other power meters and any differences are negligible, meaning this isn't an Nvidia-biased tool. It allows us to see real power use, while running games. It also integrates with Nvidia's FrameView software, a customized version of Intel's PresentMon, so that we get frametimes as well as GPU clocks, power, and temperatures for every benchmark.

Everything is hooked up to a high-end 4K 240Hz monitor, in this case an MSI MPG 272URX. We mostly like the monitor, though it has a penchant to be extremely oversaturated and blown out on colors with certain GPUs, like the RX 6000-series as well as Nvidia's RTX 50-series when running at non-native resolutions. Weird. Also, the newer MSI MPG 321URX has the same specs and features but a larger 32-inch display, for $200 less, so if you're looking for a display recommendation, that's what we'd recommend today.

That wraps up the hardware side of things. What about the software, meaning the games? We've selected 15 different games, representing a variety of genres and engines. In alphabetical order, they are: Black Myth Wukong, Control Ultimate Edition (DXR), Cyberpunk 2077 (DXR), Dragon Age: The Veilguard, Final Fantasy XVI, Flight Simulator 2024, God of War Ragnarok, Indiana Jones and the Great Circle (DXR), MechWarrior 5 Clans, Monster Hunter Wilds, A Quiet Place: The Road Ahead, Spider-Man 2, Stalker 2, Warhammer 40,000: Space Marine 2, and Witchfire. You'll note the "DXR" on Control, Cyberpunk, and Indiana Jones, indicating we've enabled DirectX Raytracing on those three games — though we didn't use the full RT option in Cyberpunk or Indiana Jones.

Each game gets tested at four different resolution and setting combinations: 1080p medium, 1080p ultra, 1440p ultra, and 4K ultra. Obviously, some of those won't make much sense on certain GPUs, but trying to test every potentially interesting setting at every resolution on each GPU is a fool's errand. Budget GPUs will primarily target 1080p, and extreme GPUs will mostly be of interest at 4K. The 1080p and 1440p native results also serve as a viable proxy for 4K with upscaling — quality mode upscaling at 4K will perform similarly to 1440p native, and performance mode upscaling at 4K will perform similarly to 1080p native.

That's it for the 2025 and 2026 GPU Hierarchy testbed. We plan on keeping it static throughout at least the next year, though drivers will be upgraded as needed and we may change up some of the games over time. If there's anything you'd specifically like to see us test, drop a note in the comments. Thanks for reading, and please share this with your friends.