Leaked code for Nvidia’s next-generation upscaler is producing some of the ugliest frame-rate charts of the year, and this time the damage isn’t landing on old budget cards. It’s landing on the RTX 3060, the RTX 3070, the RTX 3080 and the RTX 3090, GPUs that still make up a huge share of the graphics hardware running on Steam today. Multiple hardware outlets, including eTeknix and OC3D, published tests on August 30 and 31, 2026 showing DLSS 5’s Neural Rendering mode dropping some RTX 30-series games from over 100 frames per second down to single digits.
None of this is an official Nvidia benchmark. It’s leaked, pre-release AI model code, forced onto hardware it was never built for by community modders. But the numbers are real, they’re independently reproduced across several outlets, and they say something specific about how Nvidia is likely to gate its next big graphics feature. Here’s what actually happened, why Ampere cards choke on it, and what it means if you’re still running a three- or four-year-old GeForce card.
What leaked, and why RTX 30-series cards are even in this story
DLSS 5 is Nvidia’s next major revision of its AI upscaling and frame-generation stack, and the company has already teased it publicly on its own DLSS technology page, describing a Neural Rendering pipeline that “runs in real time at up to 4K resolution for smooth, interactive gameplay.” That official page does not say a word about RTX 30-series compatibility, because Nvidia hasn’t shipped DLSS 5 yet.
What did ship, unintentionally, was a build of the DLSS 5 Neural Rendering DLL embedded inside an early-access version of a 2026 sports title. Modders pulled the file out, and the first round of coverage in prior weeks focused on getting it running on RTX 40-series “Ada Lovelace” cards, since the original binaries were compiled for Blackwell (RTX 50) only. That earlier mod story showed Ada hardware handling the workload reasonably well. RTX 30-series “Ampere” cards are a different animal entirely, and the newest wave of testing makes that painfully clear.
The test results: DLSS 5 Neural Rendering on Ampere hardware
According to reports gathered by hardware outlets this week, community testers loaded the patched DLL on a handful of RTX 30-series cards, across a spread of games with very different rendering loads. The pattern was consistent: frame rates that were entirely normal without DLSS 5 Neural Rendering collapsed once it was switched on. Some titles landed in a marginally usable range with settings tuned down. Most did not.
| Game | GPU tested | Without DLSS 5 Neural Rendering | With DLSS 5 Neural Rendering | Reported by |
|---|---|---|---|---|
| Deep Rock Galactic | RTX 3080 | ~130 FPS | ~4 FPS | Community testing, cited by hardware outlets |
| Kingdom Come: Deliverance II | RTX 3070 (laptop) | ~29ms render latency | 3,300ms+ render latency (well under 1 FPS) | Community testing, cited by hardware outlets |
| Control / FiveM / KCD2 | RTX 3050, RTX 3060 Ti | Playable | ~1 FPS | Community testing, cited by hardware outlets |
| Hogwarts Legacy | RTX 3090 | N/A | ~30 FPS (tuned settings) | Community testing, cited by hardware outlets |
| Mortal Shell 2 | RTX 3080 Ti | N/A | ~34 FPS (tuned settings) | Community testing, cited by hardware outlets |
| Ready or Not | RTX 3080 | N/A | ~30 FPS (tuned settings) | Community testing, cited by hardware outlets |
| Neverness to Everness | RTX 3080 | N/A | ~29-30 FPS (tuned settings) | Community testing, cited by hardware outlets |
| NBA 2K27 | RTX 3070 (laptop) | N/A | ~25 FPS (tuned settings) | Community testing, cited by hardware outlets |
| Satisfactory | RTX 3080 Ti | N/A | ~30 FPS, described as “still far from playable” | eTeknix, Aug. 31, 2026 |
Look at the pattern in that table and two things stand out. First, even the “good” results on RTX 30-series parts cap out around 25 to 34 frames per second with settings pulled way down, nowhere near what those same cards deliver in those games normally. Second, the worst cases aren’t just slow, they’re unusable: a Deep Rock Galactic session running at roughly 130 frames per second without the feature falls to around 4 frames per second with it switched on. On the RTX 3070 laptop GPU, render latency in Kingdom Come: Deliverance II went from an already-modest 29 milliseconds to more than 3,300 milliseconds, a latency spike so severe it functions as a slideshow rather than a game.
Why performance falls off a cliff on RTX 30-series cards
The likely technical explanation traces back to how the DLSS 5 Neural Rendering model is built. Prior reporting on the leaked DLL, drawn from outlets covering the initial RTX 40-series mod, described the network as running primarily in FP8 precision, a data format that Ada Lovelace’s tensor cores execute natively. Ampere, the architecture underneath every RTX 30-series card, does not have that same native FP8 execution path. When a model built around FP8 math gets forced onto hardware that has to approximate or emulate that precision instead of running it directly, the performance cost is not small. It’s the difference between a GPU doing the work it was designed to do and a GPU faking its way through math it was never wired for.
That distinction matters because it means the RTX 30-series numbers aren’t really a story about Nvidia shipping a broken feature. They’re a story about what happens when experimental, pre-release AI code gets pulled out of a game build and dropped onto silicon three architecture generations removed from its intended target, with none of the driver-level optimization work Nvidia would normally do before a public launch.
This is not an official Nvidia benchmark, and that distinction matters
It’s worth being precise about what has and hasn’t been confirmed here. Nvidia has not published an RTX 30-series DLSS 5 benchmark, has not claimed the feature works on Ampere, and has not commented on the leaked file at all as of this writing. Everything in the numbers above comes from modders running unofficial, unsupported software on hardware Nvidia never validated it against. That’s a meaningfully different thing than a vendor shipping a feature and having it perform badly.
What is confirmed, and what makes this worth covering as news rather than dismissing as internet noise, is that the same performance collapse has now been reported independently by more than one hardware outlet, testing separate games on separate cards, and arriving at broadly the same conclusion: DLSS 5 Neural Rendering, in its current leaked form, does not run acceptably on Ampere.
From an early-access game build to a community modding trail
The DLSS 5 Neural Rendering file first surfaced through an early-access build of a 2026 sports game, where it sat unused as part of Nvidia’s forward-looking integration work. Modders extracted it, and the initial round of experimentation focused on getting the Blackwell-only binaries running on RTX 40-series cards, a project that produced workable, if unofficial, results in prior coverage. The jump to RTX 30-series cards came next, and that’s where this week’s numbers come from. It’s a chain of unofficial work built on unofficial work, three steps removed from anything Nvidia has actually shipped.
Nvidia’s silence and a “this fall” launch window
Reporting from eTeknix on August 31, 2026 notes that Nvidia currently plans to officially launch DLSS 5 this fall, but has not disclosed which GPU generations will receive full support. That silence is doing a lot of work in this story. Nvidia’s own DLSS 5 announcement page talks up 4K real-time performance without a single reference to Ampere, Ada, or any specific architecture beyond the implication that the feature is built for its newest silicon. Until Nvidia says otherwise, RTX 30-series owners have no official signal on whether DLSS 5 will run on their cards at all, let alone run well.
Generational gatekeeping has happened before
Nvidia has a documented history of tying new DLSS capabilities to specific hardware generations, and it’s worth putting the current situation in that context rather than treating it as a one-off.
The Optical Flow Accelerator precedent
DLSS 3’s frame generation, launched alongside the RTX 40 series, relied on a new Optical Flow Accelerator built into Ada Lovelace silicon. Nvidia’s own DLSS 3 documentation states plainly that RTX 30-series cards lack that Optical Flow Accelerator, which is why frame generation never came to Ampere despite years of community requests. DLSS 4’s Multi Frame Generation followed the same logic, building on Ada and Blackwell hardware features rather than extending backward.
Seen against that backdrop, an Ampere generation that can’t run FP8-native neural rendering at acceptable speed fits a pattern rather than breaking one. The table below lines up what’s confirmed about each recent Nvidia GPU generation against what the leaked DLSS 5 testing has shown so far.
| GPU generation | Architecture | Native DLSS 3 frame generation | Nvidia’s official DLSS 5 stance (as of Aug. 31, 2026) | Leaked DLSS 5 test outcome |
|---|---|---|---|---|
| RTX 50 series | Blackwell | Yes | Original target of the leaked build | Not independently benchmarked in this round of leaks |
| RTX 40 series | Ada Lovelace | Yes | Not officially confirmed | Runs in modded form, reported as functional in earlier coverage |
| RTX 30 series | Ampere | No (no Optical Flow Accelerator) | Not officially confirmed | Drops to roughly 1-4 FPS in several titles this week |
| RTX 20 series | Turing | No | Not officially confirmed | Similarly poor results reported by outlets covering the same leak |
Who actually owns an RTX 30-series card in 2026
This wouldn’t be much of a market story if RTX 30-series cards were a rounding error, but they aren’t. According to XDA Developers, citing Steam’s own hardware survey data published May 31, 2026, RTX 30-series cards accounted for 32.70% of all RTX-equipped systems on Steam, with RTX 40-series close behind at 34.93%. Combined, those two generations make up more than two-thirds of every RTX card in active use. Cards like the RTX 3060, RTX 3070 and RTX 3080 are approaching six years old and remain some of the most common GPUs on the platform, largely because DLSS upscaling has kept them viable well past what raw silicon would suggest.
That’s the tension at the center of this story. Nvidia’s own software has been the reason millions of RTX 30-series owners felt comfortable holding onto their cards this long. If DLSS 5 becomes the next feature that quietly skips Ampere, a third of the RTX install base loses the exact upgrade path that’s kept it competitive, and that pressure lands at a moment when Nvidia just posted a record $96.2 billion quarter built heavily on newer hardware.
Competitive pressure from AMD FSR and Intel XeSS
Nvidia’s rivals have built their upscaling roadmaps around a different bet: broader hardware compatibility over generation-locked silicon features. AMD’s FSR and Intel’s XeSS have both leaned on approaches that run across a wider span of GPU generations, including older Nvidia hardware, rather than requiring a specific tensor core generation or accelerator block to function at all. That doesn’t mean their image quality or frame-generation results match Nvidia’s newest tech, and Nvidia’s upscaler has generally led head-to-head quality comparisons. But if DLSS 5 launches as another RTX 50-exclusive feature, it hands AMD and Intel a straightforward marketing argument: their upscalers keep working on the GPU you already own, while Nvidia’s newest feature increasingly does not.
What this means for the used GPU and upgrade market
Every time Nvidia has drawn a hardware line under a new DLSS capability, it has nudged the upgrade cycle for cards sitting just below that line. RTX 30-series pricing on the secondhand market has already softened as RTX 40 and RTX 50-series supply has grown, even amid the memory shortage that pushed Nvidia’s own AI server prices up more than 15% this year. A confirmed DLSS 5 generation gap would add another reason for buyers to skip Ampere entirely when shopping used, and another argument for current RTX 30-series owners to start budgeting for a replacement rather than assuming software updates will keep stretching the card’s relevance. None of that is confirmed policy from Nvidia yet, but the leaked performance data gives it a lot more weight than speculation alone would.
The broader pattern: leaked AI features outrunning official rollouts
DLSS 5 isn’t the first piece of unreleased AI graphics tech to leak out of a game build and get reverse-engineered by modders before its official debut, and it won’t be the last. As Nvidia, AMD and Intel all race to bake more of their rendering pipelines into proprietary neural networks tied to specific silicon, the window between “code exists somewhere” and “code ships officially” has become a gray zone where community testers do the vendor’s compatibility testing for free, often on hardware combinations the vendor never intended to support. It’s a useful preview for enthusiasts and a genuine headache for companies trying to control the narrative around a feature they haven’t announced yet.
Five predictions for where DLSS 5 support goes next
- Nvidia’s official fall 2026 DLSS 5 launch will most plausibly ship as Blackwell-first, mirroring how DLSS 3 launched as an RTX 40-series exclusive tied to a specific hardware block.
- Ada Lovelace support is more likely than Ampere support, given that modders have already demonstrated a workable (if unofficial) path on RTX 40-series cards, unlike the collapse seen on RTX 30-series parts.
- Expect Nvidia to address the leak indirectly, through driver updates or terms-of-service language discouraging use of pre-release DLLs, rather than a direct statement about RTX 30-series performance.
- AMD and Intel will likely lean harder into “runs on your existing GPU” messaging around FSR and XeSS updates in the months around DLSS 5’s official launch.
- If RTX 30-series exclusion becomes official, expect a visible bump in used RTX 40-series pricing as Ampere owners look for the cheapest path to full feature support rather than waiting for next-generation cards.
What RTX 30-series owners should actually do right now
There is no action item today, and that’s worth saying plainly. The leaked DLSS 5 build is not a consumer feature, isn’t distributed through Nvidia’s official channels, and running it means using unsupported code extracted from a game file, not a supported driver package. RTX 30-series owners don’t need to do anything differently based on this week’s tests, because nothing about their card’s official feature set has changed. The more useful takeaway is expectation-setting: don’t count on DLSS 5 being a free performance upgrade for a three-to-six-year-old GPU, and treat any official Nvidia statement on the matter, whenever it arrives this fall, as the only information that should actually shape a buying decision.
Frequently asked questions
Is DLSS 5 officially released yet?
No. As of August 31, 2026, DLSS 5 has not had a full public launch. Nvidia has published a technology page describing the feature and reporting indicates a planned launch this fall, but the leaked Neural Rendering DLL currently being tested by modders is pre-release code, not a shipped consumer feature.
Does DLSS 5 work on RTX 30-series GPUs?
Not well, based on this week’s leaked testing. Community reports show frame rates dropping to roughly 1 to 4 FPS in several games when the leaked DLSS 5 Neural Rendering DLL is forced onto RTX 30-series “Ampere” cards, with a handful of titles reaching a marginal 25-34 FPS under tuned settings. Nvidia has not confirmed official RTX 30-series support.
Why does DLSS 5 perform so badly on RTX 30-series cards?
The leaked model reportedly runs in FP8 precision, a format Ada Lovelace (RTX 40) and Blackwell (RTX 50) tensor cores handle natively. Ampere, the architecture behind RTX 30-series cards, lacks that native FP8 execution path, which appears to force slow workarounds instead of direct hardware acceleration.
Which GPUs will officially support DLSS 5?
Nvidia has not disclosed which generations will get official DLSS 5 support ahead of its planned fall 2026 launch. Given the leaked build’s Blackwell-only original compilation and DLSS 3’s precedent of an RTX 40-exclusive Optical Flow Accelerator requirement, RTX 50-series support looks most likely, with Ada Lovelace as a secondary possibility.
Is it safe to try the leaked DLSS 5 files myself?
It isn’t recommended. The files are unofficial, extracted from a game build without Nvidia’s authorization, and unsupported by any driver package. Performance is inconsistent at best and unusable at worst, and there’s no guarantee of stability or compatibility with a given system.
How many people still use RTX 30-series GPUs?
RTX 30-series cards made up 32.70% of all RTX-equipped systems on Steam as of data published in May 2026, according to XDA Developers’ analysis of Steam’s hardware survey, with RTX 40-series close behind at 34.93%. Together those two generations account for more than two-thirds of active RTX hardware.
Has Nvidia commented on the leak?
No official statement from Nvidia addresses the leaked DLSS 5 file or its performance on RTX 30-series cards as of this writing. The company’s public DLSS 5 page discusses the feature’s real-time 4K capability without referencing RTX 30-series compatibility one way or another.
Is this the same as the earlier DLSS 5 RTX 4090 mod story?
It’s related but distinct. Earlier coverage focused on modders getting the leaked DLSS 5 DLL functional on RTX 40-series “Ada Lovelace” cards, which share more architectural similarities with the original Blackwell target. This week’s reports cover a separate round of testing on RTX 30-series “Ampere” cards, where results are dramatically worse.
