The cards don’t go on sale for another week, but samples of the much-anticipated Nvidia GeForce RTX 5090 are in the wild and in the hands of reviewers around the world. (See our initial unboxing of the card from a few days ago.) And we’re all banging away furiously at these cards with our special benchmarking hammers: new and familiar synthetic test programs, GPU-accelerated productivity applications, and a mix of AAA games to push the limits of Nvidia’s new silicon. Testing the new flagship RTX 5090 is a mix of old challenges (how well does it do on classic rasterization?) as well as new ones: How well do the AI features assist in accelerating new games? How about for the hands-on AI tasks that people might do today?

Here at PCMag, we’re in the midst of recalibrating and rebuilding our database of graphics card tests. We’re doing that in light of the new GeForce RTX 50-series generation of Nvidia cards, with AMD Radeon RX 9000-series RDNA 4 cards soon in the offing and Intel making some real progress at the lower end of the market with its Arc line. That means lots of retesting of old cards with new tests, and experimenting with new tools and processes.

(Credit: John Burek)

Take, for one thing, DLSS. Nvidia’s upscaling and frame-rate-boosting tech has taken different forms over its relatively short life. It started as an upscaler (rendering frames at a lower resolution, then enhancing them to appear higher-res), moved into frame generation (the short explanation: inserting AI-generated frames between classically rasterized ones), and is now poised to be a somewhat different animal still in some PC games, inserting multiple AI-generated frames between “real” ones to rocket-boost frame rates. (Our senior analyst will be taking a quick first look at the new DLSS 4 in a companion piece here shortly to follow.)

AI assistance and other side tweaks are complexifying the GPU market, to be sure. Meanwhile, though, there’s a more familiar question: How well does the RTX 5090 flagship card of the moment do for the basic stuff? With our new test suite, we pitted the GeForce RTX 5090 Founders Edition–Nvidia’s own version of its flagship GPU–against its former flagship card, the GeForce RTX 4090. We haven’t gathered enough data yet for a full, editor-scored review of the RTX 5090–we need more time with the card, as well as to test a critical mass of older cards in our new fashion. (We received ours less than 72 hours ago.) Nvidia hosted an Editor’s Day earlier this month at CES 2025 that unveiled some of the technologies behind the RTX 50 series, and that briefing made clear the direction that future testing would need to take. So it’s off the benchmarking races.

Below is a rundown of some of the numbers we churned so far. (Our new testbed PC comprises a top-of-the-line 16-core AMD Ryzen 9 9950X CPU cooled by a 360mm Cooler Master cooler, a Gigabyte X870E Aorus Master motherboard, 32GB of Crucial DDR5 clocked at the AMD EXPO 6000 setting, two PCI Express 4.0 SSDs from Crucial, and a 1,500-watt Corsair power supply.) Lots of raw charts follow. The TLDR, though? This card is looking like a beast by any measure. It even cows the power-monger RTX 4090, which we’ll compare it to throughout.

Synthetic Graphics Tests: A Festival of 3DMark

First, let’s take a look at some industry-standard graphics benches we ran. While UL’s various 3DMark subtests are mostly only useful in relation to other 3DMark scores, they’re also a good “reality check” on the hardware at hand, and a key grounding point for subsequent testing.

For those unfamiliar with the suite, here’s a rundown what each test specializes in…

• Time Spy Extreme is a legacy test that pushes gaming PCs to the max. It’s been used for years as a quick 4K-gaming-graphics stress measure.

• Steel Nomad is the newer, spiritual successor to Time Spy, also running at 4K. According to UL, it simulates “the latest heavy PC games.”

• Speed Way is a newer 3DMark test meant for high-end gaming PCs on Windows 10 and 11. It exposes DirectX 12 Ultimate’s ray-tracing features to measure a PC’s performance specifically with ray-traced workloads.

• Port Royal is a lighter-weight ray-tracing benchmark that employs Vulkan and enables cross-comparability between platforms.

• Solar Bay is an even lighter-weight ray-tracing benchmark that also uses Vulkan. It’s here, but it’s less relevant than Speed Way or Port Royal, looked at in a vacuum.

• The UL 3DMark Nvidia DLSS Feature Test gives a rough before-and-after result with a DLSS-capable graphics workload. It’s useful, as you’d expect, only with Nvidia RTX-class cards. Our test numbers here compare DLSS 3 loads versus no DLSS 3; note that the DLSS 3 “on” results use frame generation (that is, sticking an AI-generated frame between each typically rendered frame).

Here’s a peek at the RTX 4090 versus the 5090 on the 3DMark suite…

We won’t calculate every percentage, but it’s clear that at least on these synthetic tests, the RTX 5090 looks to be quite a step above the RTX 4090, especially in the classic 4K Steel Nomad and Time Spy Extreme trials. The DLSS Feature test, set for DLSS 3, also shows a good-size leap.

A Few GPU-Accelerated Faves: Adobe’s Creative Suite, Blender, and More

Next up: Some key apps that benefit (to a greater or lesser degree) from GPU acceleration. Adobe Photoshop and Premiere Pro are tested using a utility from workstation maker Puget Systems, dubbed PugetBench for Creators. PugetBench runs each program through a litany of typical creative tasks: applying image filters, picture manipulation, performing file renders, and the like.

The modeling program Blender, meanwhile, now has a nifty benchmark utility that lets you run three 3D-model renders (dubbed Monster, Junk Shop, and Classroom) on the CPU or GPU. (Here, we chose the GPU option, of course.) And V-Ray 6 lets you run renders in Chaos Group’s engine used by many 3D modeling programs. It runs both RTX and CUDA versions of the test in turn.

Photoshop didn’t see a boost from the RTX 5090, at least in this workload. As we see more cards tested and the variances, we may well drop this benchmark from our suite. Premiere Pro, however, saw a healthy boost. V-Ray, meanwhile, is mostly interesting for the RTX results, which show a boost; the CUDA trial doesn’t run well on the “Blackwell”-based RTX 50 series yet. (Nvidia says this is a known issue.)

One Peek at AI Performance

The field of local AI processing is as wide as the sea is deep, and testing that kind of workload is almost as new as the RTX 50 series itself. We’re working up another AI test or two as we write this–of course, locally run AI modeling and training could well be the biggest draw of this card–but for now, we have a measure from our friends at UL of local AI text-generation performance with four of the most common LLM models out there of the moment: Phi, Mistral, and two flavors of Llama. Time to first token, and average token rate per second, are the typical measures for tasks like this; you’ll see those measures on the second two tabs below. The overall UL Procyon score is generated by the software and mostly useful only in relation to other equivalent scores.

We’ll need more context for more cards, but we suspect we won’t soon be seeing anything higher than the scores afforded here by the RTX 5090 among consumer-grade GPUs.

We’re working toward adding a local-model image-generation benchmark to the suite (at the moment, it may only be useful for comparing Nvidia cards to Nvidia cards, and high-end cards at that), so more on that soon.

A Helping of AAA Games

Next up: Five of the AAA games we use for testing. Now, the RTX 4090 and RTX 5090, despite the lust of most gamers for one, are really not gaming cards, first and foremost. Some well-heeled shoppers will buy them for that, to be sure. But the RTX 5090’s $1,999 starting price is certainly going to give pause to many a shopper, who will look further down Nvidia’s stack when the lower-end RTX 50 cards debut.

Several games here (Returnal, Call of Duty: Modern Warfare III) we tested strictly with “classic” FPS measures at our three common resolutions of 1080p, 1440p, and 4K. We also tested a couple (F1 2024, Cyberpunk 2077) with a version of DLSS on, and DLSS plus frame generation. (A third DLSS title, Black Myth Wukong, which runs as a stand-alone benchmark, we ran with DLSS on and with DLSS plus frame generation on. There’s no DLSS-off function with this test.) We’ve outlined the base settings we used with each game; we’ll get into more detail in the final review of this card.

Here’s the skinny on the first two games…

and here’s the latter three…

Note: Here, we did not factor in DLSS 4, supported in Cyberpunk 2077, quite yet. Yes, that is one of the prime new features touted by Nvidia with the launch of the RTX 50 series. DLSS 4, which has one of its key traits Multi-Frame Generation (MFG), is showing up in Cyberpunk 2077 as one of about 75 games that will support this new version of DLSS at launch. It’s a big part of Nvidia’s big claims about the RTX 50 series (especially the much ballyhooed “$549 RTX 5070 that will perform like an RTX 4090!” in Nvidia’s CES 2025 keynote, which is largely down to specific DLSS 4 scenarios). And DLSS 4 will be a big factor in future titles, we are sure. But DLSS 4 support will be lean enough from the start that it’s not the primary focus of our testing here on day one. (Again, see our companion story for more on DLSS 4, coming up shortly.)

As for what we saw? The numbers speak for themselves. Some of these games at lower resolutions are clearly limited even by the tip-top Ryzen 9 9950X we used, but at the higher resolutions, you will see some very healthy bumps depending on the game. We’re also reporting here “1% lows,” which is a popular measure of the average 1% lowest of frames in a given benchmark run. This reflects relative smoothness and the prevalence of stutter; the closer the “1% low” to the overall average frame rate, the smoother the experience. This metric is not always entirely dependent on the GPU but can be down to the game and the CPU, too; we’ll be looking at how this metric shakes out with subsequent cards and if the relative gulfs are consistent from card to card in a given game. (Note the wide swings with Call of Duty, say, versus Returnal. Also note that we had some technical issues collecting 1% lows from the RTX 4090 for Cyberpunk 2077. We’re working to fix that.)

How About a Couple of Legacy Games?

Finally, we’ve revised our bank of what we call “legacy” games (that is, older titles) to ones that aren’t quite as antique as the golden oldies (Bioshock Infinite, Tomb Raider, Sleeping Dogs) that we were using until recently. These two, Shadow of the Tomb Raider (2018) and Total War: Three Kingdoms (2019), are classics that we’ll keep around to see how a card deals with older games that may not be optimized right out of the gate. Note: Total War is especially sensitive to CPU power and cores, so the Ryzen 9 in our testbed will give it an extra boost here and also show some interesting results…

Some solid increases across the board here suggest that with at least these two titles (admittedly, a tiny sample size), the RTX 5090 may not see too much of a cliff with older games. Nvidia tends to be good this way out of the gate.

More Testing and Numbers to Come…

This is just a tease of the RTX 5090’s capabilities versus the former flagship RTX 4090, the card that most folks who are looking at the RTX 5090’s capabilities are likely to weigh it against. We haven’t gotten in hand–yet–a sample of the RTX 5080 (which will launch alongside the RTX 5090, on Jan. 30), and the announced GeForce RTX 5070 Ti and RTX 5070 are coming a little later.

A fuller assessment of the RTX 5090 may require the RTX 5080 in hand, or at least a few more cards down the stack, like the RTX 4080 or AMD’s closest competitor, the Radeon RX 7900 XTX. But we’ll be testing it and a host of other key competing cards all through the winter–possibly into the spring!–to see how the new graphics landscape in 2025 shakes out.

One thing’s clear, though: The path to consumer-card royalty is looking like it runs straight through the RTX 4090 to the RTX 5090.

Get Our Best Stories!

Sign up for What’s New Now to get our top stories delivered to your inbox every morning.

This newsletter may contain advertising, deals, or affiliate links.
By clicking the button, you confirm you are 16+ and agree to our
Terms of Use and
Privacy Policy.
You may unsubscribe from the newsletters at any time.

About John Burek

Executive Editor and PC Labs Director

I have been a technology journalist for 30-plus years and have covered just about every kind of computer gear—from the 386SX to 64-core processors—in my long tenure as an editor, a writer, and an advice columnist. For almost a quarter-century, I worked on the seminal, gigantic Computer Shopper magazine (and later, its digital counterpart), aka the phone book for PC buyers, and the nemesis of every postal delivery person. I was Computer Shopper’s editor in chief for its final nine years, after which much of its digital content was folded into PCMag.com. I also served, briefly, as the editor in chief of the well-known hard-core tech site Tom’s Hardware.

During that time, I’ve built and torn down enough desktop PCs to equip a city block’s worth of internet cafes. Under race conditions, I’ve built PCs from bare-board to bootup in under 5 minutes.

In my early career, I worked as an editor of scholarly science books, and as an editor of “Dummies”-style computer guidebooks for Brady Books (now, BradyGames). I’m a lifetime New Yorker, a graduate of New York University’s journalism program, and a member of Phi Beta Kappa.

Read John’s full bio

Read the latest from John Burek