Nvidia GeForce RTX 4090 review: a new level in graphics power

We've covered Nvidia DLSS 3 frame generation in prior content quite extensively and we were the first outlet to gain access to the new technology. This is disruptive stuff that caused a number of headaches when we first went hands on with the RTX 4090 and preview versions of the games. How could we measure performance in these titles? How could we actually show the DLSS 3 frame generation effect in action at its intended high frame-rates in a world where 4K 120Hz capture cards do not exist?

We overcame this by using OBS - the streaming tool - to record the display output of the graphics card. The new Ada Lovelace architecture comes with an exceptionally performant media engine that does allow for 4K video to be encoded at 120fps (NVShare tops out at 8K 60fps instead). However, we were effectively running the games in concert with 4K120 capture, and OBS itself would not always sync to the game update. Our solution? To ensure that game performance was locked to 120fps, then to capture footage in bulk, using frame-rate analysis on the captured clips to root out the areas of the captures where frames were dropped in the encoding process.

Performance analysis? Most outlets use PresetMon - an internal tool that offers a wealth of data - and Nvidia has created its own version for FrameView that does record frame-times with DLSS 3. However, we found that DLSS 3 plumbed straight into our video capture performance tools. We did have access to new titles but owing to lack of time, we've stuck to using our preview data here, where we can now show everything.

Marvel's Spider-Man: The Fisk Construction Mission

I used the Fisk Construction mission to get this data, equalising the video feeds by essentially sticking to the quicktime events in the game. You'll immediately notice that something looks off: there's little gain from DLSS 2 vs native rendering, with both RTX 3090 Ti and RTX 4090 delivering very similar performance levels. What's happening here is that we're hitting the CPU limit imposed by the Core i9 12900K in our new test platform. This is what makes the DLSS 3 frame amplification technology so interesting: as it is rendering two frames, then interpolating an 'inbetweener' frame, there is no additional CPU overhead. Therefore, DLSS 3 takes game fluidity in a new direction that would not have previously been possible. And yes, the game is indeed running at circa 235 frames per second.

Marvel's Spider-Man: The Times Square Benchmark

The gameplay benchmark is quite dynamic, so I also tested the game in a static scene in Times Square, home of most of the best Marvel's Spider-Man benchmarks. As the GPU load is consistently higher than the gameplay bench, you'll see some difference between the native resolution rendering and DLSS 2 performance mode scores, but really, we're still leaving a lot of GPU power on the table. RTX 4090 should be using DLSS quality mode here, where it would achieve the same result. Still, this is about frame generation - and once again you can see the end result here - a doubling of frame-rate with DLSS 3. If the frame-rate looks a little unsteady, remember that the difference between 180fps and 200fps is just 0.5ms. The higher frame-rate goes, the less relevant the metric really is and measuring by actual time taken to render a frame is more informative.

Cyberpunk 2077: The Cherry Blossom Marketplace

To help show DLSS 3 deployed on a high-end ray traced gaming experience, we were given an early build of Cyberpunk 2077 with the new technology integrated. However, the build was far from completion and unfortunately did not feature the RT Overdrive additions that effectively deploy full path-tracing to the game's lighting. Still, psycho RT is a punishing workload and a great test for the technology. The accelerant in providing an end-to-end 4x frame-rate multiplier comes from two stages of DLSS. Native 4K RT rending is exceptionally tough, even on the RTX 4090, so at the first stage, DLSS 2 provides a huge performance uplift - essentially from 34fps to 88fps. Frame generation then kicks in, taking us to an average 134fps. In this graph, you'll see where RTX 3090 Ti tops out via DLSS 2, then the performance increase moving to RTX 4090 at the same level, then with DLSS 3 enabled - frame generation is not a feature available to RTX 3000 GPUs.

Cyberpunk 2077: Night City Driving

Driving through Night City is often perceived as one of the most challenging areas of Cyberpunk 2077 performance. In actual fact, the Cherry Blossom Marketplace is more demanding - a lower base frame-rate delivers higher frame-rate multipliers. Therefore, percentage increases may be lower here, but as you'll see from the frame-rate numbers, the game runs more smoothly.

There's a lot more to discuss with DLSS 3 and we have content in the pipeline to discuss its image quality challenges and where we go into more depth on latency. In the meantime, our existing DLSS 3 first look contains a lot of useful information for you to check out. A short note though on the latency tests for Cyperpunk 2077 - in the first look, DLSS lag came in at 54ms. In the build sent out to reviewers, this dropped substantially to circa 35ms.

Nvidia GeForce RTX 4090 analysis

• Introduction, hardware and power analysis
• Game benchmarks: Control, Cyberpunk 2077, Doom Eternal
• Game benchmarks: F1 22, Forza Horizon 5, Gears 5
• Game benchmarks: Hitman 3, Red Dead Redemption 2, Shadow of the Tomb Raider
• RT benchmarks: Control, Dying Light 2, Cyberpunk 2077, F1 22
• RT benchmarks: Hitman 3, Metro Exodus Enhanced Edition, Marvel's Spider-Man
• DLSS vs FSR 2.0: Dying Light 2, Marvel's Spider-Man
• DLSS 3: Marvel's Spider-Man, Cyberpunk 2077 [This Page]
• Nvidia GeForce RTX 4090: the Digital Foundry verdict