The RTX 5090 loses half its performance to DLSS 5, and Nvidia's favorite trick won't save it this time

Published Sep 11, 2026, 4:30 PM EDT Abhinav pivoted from a career in banking to pursue his first love in writing. Even while working full-time, he continued contributing as an editor-at-large, a role he has held for more than 7 years. A lifelong tech enthusiast who has built three gaming and productivity powerhouse PCs since 2018, his passion for technology keeps him closely following the semiconductor industry, from NVIDIA and AMD to ARM. His MSc dissertation explored how artificial intelligence will reshape the future of work, reflecting his curiosity about the wider social impact of emerging technologies. The highly anticipated (and highly controversial) DLSS 5 is finally out. We've seen some breathtaking implementations, but unfortunately, the benchmarks aren't looking as good. Nvidia's latest neural rendering technology delivers some impressive visual results when it comes to atmosphere, characters and objects with respect to tone and depth, but getting there comes with a hefty performance cost that could drive enthusiasts away rather than luring them in. As a matter of fact, the performance hit is so massive that even Nvidia's favorite trick in the book, Frame Generation, could potentially not be able to salvage it this time. There's a structural reason for that, and it's part of these benchmarks that should worry Blackwell GPU owners, even the ones with a flagship RTX 5090. The first official DLSS 5 game brought the RTX 5090 to its knees That might just be a brand-new sentence The visual results of DLSS 5 were impressive, and reviewers with hands-on time on early DLSS 5 builds, myself included, have generally come away rather fascinated with what neural rendering can do to a scene. Right now, NBA 2K27 is the only title with official DLSS 5 support, and Nvidia dropped the accompanying driver update on September 3 ahead of the launch. Digital Foundry's coverage of the implementation concedes that it was subtle and tasteful, at least as far as visuals are concerned. On the performance side of things, however, the story becomes a lot less flattering and raises more questions than it answers. In DF's benchmark runs, the incumbent flagship RTX 5090 fell from 128.4 FPS at 4K resolution down to 62.7 FPS with DLSS 5 enabled, measuring a 51.2% drop on the fastest consumer graphics card on sale. What's even more interesting is that the flagship is not the worst case here. Credit: Digital Foundry. The RTX 5080 suffered from a 58.4% performance hit at 4K, landing at 42.6 FPS, sitting uncomfortably below the 60 FPS floor. On the cheapest Blackwell card, the hit was the mildest on the board, however, DF ran it at high rather than ultra preset at 1080p resolution. TechSpot's analysis found a reasonable explanation for this, stating that DLSS 5 likely has a fixed cost that's dependent on the output resolution, which makes sense when you look at the performance data from the RTX 5060 run. But the performance hit isn't the only cost Neural rendering also drinks power like nothing before it For a feature as compute intensive as DLSS 5, you can bet that there are power draw implications to match it as well. And sure enough, Tom's Hardware's testing confirmed as much, demonstrating that enabling neural-rendering pass can make the Blackwell flagship significantly thirstier, even as frame rates take a nose dive heading for the ground. It's worth noting that the DLSS 5 build tested by Tom's Hardware was an OptiScaler injection mod similar to my own experiment with DLSS 5, and as such, the publication notes that a developer-implemented version through Nvidia's Streamline framework may incur less overhead than what was measured in their tests. In Hogwarts Legacy, an MSI Lightning Z RTX 5090 went from 480W with DLSS 5 disabled to 720W with it on, which was a massive 50% jump in consumption. The Founders Edition of the same card, on the other hand, went from 417W to 547W, resulting in a 31% increase. Control, meanwhile, had the Lightning Z card leaping from 691 W to 802W. Cyberpunk 2077 with path tracing enabled had the FE card drawing 482W without DLSS 5 and only 551W with it, which looks like a mild 14% increase until you take into account that the FE cannot exceed 575W due to a hard power limit. In the same test, the Lightning Z jumped from 580W to 723W. Frame Generation can't seem to save DLSS 5 from itself Nvidia's favorite trick may salvage the frame rates, but what can it do for latency? So far, all testing points to the fact that DLSS 5 costs more power, inevitably produces more heat as a result, and has a chance of pushing certain SKUs well past what their rated 12V-2x6 connectors are designed to sustain, all while frame rates halve on the same hardware. Whether the visual uplift alone warrants that cost is yet to be answered. Now, of course, the obvious question that's raised is that, since it's Nvidia, and one of Blackwell's defining features is 6x Multi-Frame Generation, surely it would be able to salvage some of that performance, right? Well, not quite. Digital Foundry's Oliver McKenzie explained that, because of where the model sits, it gets a little more complicated. DLSS 5 runs after upscaling but before frame generation, which means the base frame rate MFG has to interpolate from is the post-DLSS-5 figure, and not the pre-DLSS-5 one. MFG can certainly lift the counter from there, but then the next issue becomes latency. Digital Foundry measured around 50–60 ms of PC latency on the 5090 with 4x MFG active, and over 60–70 ms on the 5080 and 5060. This is quite different from running MFG stacked on a healthy base frame rate, which is precisely where Nvidia pitched that the technology would shine. It's clear that, in its current state, the Blackwell lineup isn't ready for an enjoyable gaming experience with neural rendering. More power consumption, fewer frames, and for what? It's clear that, in its current state, the Blackwell lineup isn't ready for an enjoyable gaming experience with neural rendering. This puts Team Green in a rather awkward position surrounding a feature set that was deeply unpopular to begin with, and when the very same feature asks requires more power and generates fewer frames, it becomes a tough sell. The silver lining, of course, is the fact that it's still very early for DLSS 5, and given enough time and developer effort, it might just become the next big thing. For now, though, I'd stick with DLSS 4.5.