CPU scaling with DLSS — investigating CPU performance in the age of upscaling
In a time of ubiquitous upscaling, your CPU can quickly become a performance bottleneck.
For this investigation, we used four CPUs, representing the mainstream, entry-level option, and the ideal gaming option for both AMD and Intel. The test pool includes the Ryzen 5 9600X, Ryzen 7 9850X3D, Core i5-14400, and Core i7-14700K. For games, we used Spider-Man 2, Cyberpunk 2077, Microsoft Flight Simulator 2024, Doom: The Dark Ages, and The Last of Us Part One. All tests were run with an RTX 4080 Super. We’ll explain our methodology at the end of the article, so feel free to jump ahead if you want to get an idea of how we tested before looking at the data.
The interesting results come from individual games, but let’s set the stage with geomeans at 4K, 1440p, and 1080p. We intentionally created a GPU-bound scenario at 4K Quality in order to see how scaling held up from there. And that was successful, as you can see only a frame and a half separating the four chips in our test pool. Going to Performance mode, some scaling starts to show up. Keep in mind that 4K Performance mode means the game is using a 1080p internal render resolution.
What’s interesting is the lack of much difference at 4K Performance. We can see scaling with the Core i5-14400 compared to the other three chips. However, the Core i7-14700K and Ryzen 7 9850X3D posted nearly identical results. In our Ryzen 7 9850X3D review, our data shows AMD’s latest chip is 26% faster than the Core i7-14700K on average at 1080p. We’re using a much smaller list of games here, of course, but it’s a good illustration of the overhead involved with upscaling. We may have an internal resolution of 1080p, but we’re still bound by the GPU.
We can see clearer scaling at 1440p. The Core i5-14400 is a complete bottleneck, and the Ryzen 5 9700X is nearly a complete bottleneck. The Core i7-14700K also shows little scaling, while the Ryzen 7 9850X3D manages an 11% jump moving from Quality to Performance mode. It’s important to keep in mind how low the internal resolution is here. We’re looking at an internal resolution of 960p at Quality mode and 720p at Performance mode with a 1440p output.
At 1080p, there’s a complete CPU bottleneck for every chip but the Ryzen 7 9850X3D, which itself benefits far less from moving from Quality to Performance mode. Again, this isn’t surprising when you consider the internal resolution of 720p at Quality mode and 540p at Performance mode. The results are almost identical to what we saw at 1440p given the complete CPU bottleneck.
Looking at the results overall, it goes against some basic assumptions about how lower internal render resolutions shift performance to the CPU. At 4K, the results make sense. Given the complete GPU bottleneck at 4K Quality mode, anything lower represents a performance uplift. The 1440p results stand out the most. Not only was there little to gain from going from Quality to Performance mode, but we also saw clear scaling between each chip, regardless of the quality setting. It’s not as dramatic as what you see at native resolution, but it’s still present.
Although these geomeans are good to get a broad view of CPU scaling with DLSS enabled, individual games are what really matter here. Some games lean heavier on the CPU or GPU, sure, but I was surprised to see how that balance showed up in real results.
As we go through individual games, there are two patterns to pay attention to. If you see Performance mode (light colored bars) outperforming Quality mode (dark colored bars), the GPU is the limiting factor. If you see identical performance between Quality and Performance mode for any given chip, we’re bound by the CPU. In addition, when we talk about a CPU- or GPU-bound scenario, that doesn’t mean a complete bottleneck unless noted. We’re just looking at the component that’s mainly limiting performance.
Cyberpunk 2077 — CPU Scaling with DLSS
Cyberpunk 2077 was one of the few games I tested where there was clear performance scaling at 4K Performance mode. Once again, the results aren’t as stark as what you see at native 1080p. The Core i7-14700K and Ryzen 7 9850X3D posted very similar results despite a large gap at native 1080p.
The more important results are for 1440p and 1080p. In both cases, we run into a complete CPU bottleneck across all four chips in our test pool. That difference is quite dramatic when comparing something like the Core i7-14700K to the Ryzen 7 9850X3D. The bottleneck is so severe that there's virtually no performance difference when looking at 1080p results compared to the 1440p results. Below a 1080p internal render solution, Cyberpunk 2077 becomes completely bound by the CPU.
Another factor at play with Cyberpunk 2077 is Nvidia’s Transformer model. Nvidia itself says the Transformer model is more compute-intensive than the former CNN model, which helps explain the small gaps in CPU scaling at 4K Performance.
Doom: The Dark Ages — CPU Scaling with DLSS
Doom: The Dark Ages showed much different performance characteristics than Cyberpunk 2077. At 4K, we’re bound by the GPU completely, even in Performance mode. This game uses the Transformer model, and it has always-on ray tracing, which helps explain the lack of scaling at 4K, despite the DLSS quality mode.
We can see better scaling at 1440p, but Doom: The Dark Ages doesn’t become bound by the CPU, even when the internal resolution slips below 1080p. The main victim here is the Core i5-14400 — the weakest chip in our test pool — but the other chips still show performance improvements when going from Quality to Performance mode. That’s showcasing that we’re not bound by the CPU.
Down at 1080p, all of the chips start to fall in line. There’s a small uplift from lowering the internal render resolution, but the GPU is still the main performance driver here. The one exception is the Ryzen 7 9850X3D in Performance mode. It’s able to get past a CPU bottleneck with an internal 540p resolution, while all the other chips in our test tool topped out lower.
Flight Simulator 2024 — CPU Scaling with DLSS
Flight Simulator 2024 is notoriously heavy on the CPU, and we can see that in action even at 4K. The Core i5-14400 and Ryzen 5 9600X represent a CPU bottleneck, even at 4K Quality mode, and there’s only relief from that bottleneck when looking at the Ryzen 7 9850X3D and the Core i7-14700K in Performance mode. Although the Core i7-14700K posted similar results to the Ryzen 5 9600X at 4K Quality mode, the performance jump with Performance mode shows that the GPU still represents the performance bottleneck with the Core i7-14700K.
Down at 1440p, every chip becomes the performance limiter except the Ryzen 7 9850X3D. Once again, we can use 1080p as a gut check here; with the Core i7-14700K, Ryzen 5 9600X, and Core i5-14400, you can see basically identical results between 1440p and 1080p across quality modes. The Ryzen 7 9850X3D is the only chip that was able to benefit from the lower internal render resolution, and by a surprising margin. Even down at 1080p, it still scaled in performance as the internal render resolution lowered.
Marvel’s Spider-Man 2 — CPU Scaling with DLSS
Marvel’s Spider-Man 2 is another game that's heavy on the CPU. At 4K, the GPU remains a bottleneck for the Core i7-14700K, Ryzen 5 9600X, and Ryzen 7 9850X3D, but the Core i5-14400 imposes a bottleneck even at 4K Quality mode. With any of the other three chips, though, you’d see virtually no performance difference at 4K, regardless of your DLSS quality setting.
The Ryzen 5 9600X and Core i5-14400 become a bottleneck at 1440p and, naturally, 1080p. The Core i7-14700K doesn’t run into a complete bottleneck at 1440p, but the Ryzen 7 9850X3D still comes out ahead at Quality mode. In Performance mode, there’s a 16% jump with the Ryzen 7 9850X3D. That’s less than what you see at native 1080p, but it’s still a notable performance difference. At 1080p, that lead only becomes more exaggerated as the Ryzen 7 9850X3D comes out nearly 32% ahead of the Core i7-14700K in 1080p Performance mode.
The Last of Us Part One — CPU Scaling with DLSS
If there’s any game we tested that abides by the idea that your CPU doesn’t matter at 4K, it’s The Last of Us Part One. We’re completely GPU-bound at 4K, even when going down to DLSS Performance mode with a 1080p internal render resolution. The overhead for DLSS here is really surprising. At native 1080p, the Ryzen 7 9850X3D is 18% faster than the Core i7-14700K in The Last of Us Part One, but that advantage completely disappears with DLSS overhead.
Even at 1440p, the GPU is still the main performance driver for all chips but the Core i5-14400. We only see scaling when moving to Performance mode, and even then, the differences are minor. 1080p shows very clear scaling, which isn’t surprising considering we’re going down as low as a 540p internal render resolution.
Understanding CPU Scaling with DLSS
In some ways, the results here confirm some assumptions about the balance of performance between your CPU and GPU when the internal resolution lowers. In other ways, it flies in the face of those assumptions. Most clearly, there isn’t a direct connection between internal render resolution and performance, especially at 4K. In fact, the performance overhead of DLSS mostly overrides performance advantages, sometimes cutting the scaling we’d expect to see at 1080p in half, and other times eliminating it entirely.
Still, scaling is present, particularly in those games that lean heavier on the CPU in general. That goes against the assumption that your CPU doesn’t matter at 4K; it does. We’re using CPUs in different classes here, but we’re also using the fastest offerings for gaming from Intel and AMD. Although we didn’t test older generations, there’s a good chance the CPU bottlenecks would become more extreme if we did, especially with a modern, high-end GPU like the RTX 4080 Super.
The 1440p results are probably the most useful. Although 1080p remains the most popular resolution — at least according to the Steam hardware survey — 1440p continues to gain momentum and now represents over a fifth of gamers in Valve’s monthly tally. At 1440p, we do see very clear scaling, not only for lower-end chips, but also for the best-in-class offerings from AMD and Intel.
There are certainly unrealistic configurations here, particularly for the low-end chips. I don’t suspect many are pairing a Core i5-14400 with an RTX 4080 Super. The Core i7-14700K and RTX 4080 Super, on the other hand? That’s a configuration that makes sense. And still, at 1440p, the Core i7-14700K runs into a complete CPU bottleneck with DLSS Quality in Cyberpunk 2077, Flight Simulator 2024, and Marvel’s Spider-Man 2. It’s not that we’re running into a slight CPU bottleneck here. It’s a complete blockade when DLSS is enabled, even with something like the Core i7-14700K.
Doom: The Dark Ages and The Last of Us Part One are more representative of that mild bottleneck where we still see CPU scaling, though it’s much less pronounced. They also serve as a good reminder of how big of a variable each individual game is. We could design a test pool that shows virtually no CPU scaling just as easily as one that shows CPU scaling all the way up to 4K. Having these two games helps ground our results.
The takeaway here is that your CPU plays a big role at 1080p and 1440p, and potentially even up to 4K, depending on the games you play and the DLSS quality mode you use. There’s a very easy way to check if you’re running into a bottleneck with DLSS, thankfully. If you see the same performance at Quality and Performance mode, you’re running into a CPU bottleneck, and you should consider upgrading (maybe with one of our best CPU recommendations).
Jake Roach is the Senior CPU Analyst at Tom’s Hardware, writing reviews, news, and features about the latest consumer and workstation processors.