Leaked DLSS 5 File Found Inside NBA 2K27 Cuts Frame Rates Nearly in Half in Modder Test

A sports game accidentally leaked Nvidia's next big feature

On August 26, Brazilian hardware enthusiast Renan Maniero was digging through the PC early-access build of NBA 2K27 and found a file that had no business being there: nvngx_dlssnr.dll, 158 megabytes, identified by Windows as NVIDIA DLSSNR version 310.8.0.0 and dated that same day.

The name matched an internal naming convention for DLSS-NR, or Neural Rendering, that had surfaced earlier this year in hidden profiles buried in driver 610.47 and viewable through GeForce Profile Inspector, according to Tech Times. Until Maniero's discovery, no one outside Nvidia had ever gotten their hands on a runtime copy.

Within about 24 hours, members of the RenoDX modding community had extracted the file and injected it into a game it was never built for: Remedy Entertainment's 2019 title Control. Tom's Hardware independently confirmed the resulting numbers, per Tech Times: on an RTX 5070 Ti at 4K, frame rates on character models dropped from 71 FPS to 35 FPS with Neural Rendering switched on. That's a 51% hit.

This deserves the caveats that come with it. This was an unofficial DLL injection into a game with zero developer support for the feature, not an optimized, Nvidia-sanctioned integration. The tester didn't disclose the specific settings or internal rendering resolution used, according to Tech Times, which matters a lot when you're talking about a feature that adjusts its own workload dynamically.

The file's size is itself informative. At 158 MB, it's roughly triple the size of a typical DLSS 4 Super Resolution build (20 to 50 MB) and more than double the DLSS 4.5 Ray Reconstruction DLL, which TechPowerUp measured at 72 MB. Tech Times reports that analysis points to roughly 148 million FP8 parameters accounting for most of that bulk, and TechSpot's DLL analysis suggested this is an unfinished build. FP8, an 8-bit floating-point format, is built for the fifth-generation Tensor Cores on Blackwell GPUs, which is why this only runs on RTX 50-series cards.

None of this tells you what DLSS 5 will actually look and perform like when Nvidia ships it for real, in a game built to use it, with a finished DLL. It does tell you Nvidia's neural rendering approach is computationally heavy enough that a half-baked build cut frame rates in half in a stress case picked by modders, not by Nvidia's marketing team. Whether that gap closes before launch is the open question, and it's Nvidia's to answer, not the modding community's.

Google DeepMind tries to fix a trust problem in AI benchmarks

In an unrelated development in AI testing, Google DeepMind said on August 27 that it ran what it calls the industry's first double-blind evaluation of a commercial frontier-class AI model, according to Google DeepMind's own blog post and confirmed by BigGo Finance, The Decoder, and Gigazine.

The problem DeepMind is trying to solve: AI benchmark scores are only meaningful if the model hasn't already seen the test questions. Once test material ends up in a model's training data, a high score just measures memorization, and there's no reliable way to catch that after the fact, according to resultsense.

The old way of handling this required a trust-based tradeoff. Either the outside evaluator handed its test questions to the AI company, which then had the answer key, or the AI company handed over its model weights, exposing the core intellectual property the business is built on, per The Decoder. The Decoder cites a case involving Anthropic's Fable 5 model, where an ARC-AGI benchmark evaluation was delayed because Anthropic enforces a 30-day data retention policy on its strongest models.

For the pilot, DeepMind tested a Gemini 2.5 Flash-Lite model against confidential benchmarks from MLCommons, run inside Google Cloud's Confidential Space, a confidential-computing product that produces cryptographic proof that neither side accessed the other's material, according to BigGo Finance. Partners included MLCommons, AVERI, OpenMined, and Singapore's AI Safety Institute.

Under the setup, the outside evaluator's test prompts stay encrypted from Google, and Google's model weights stay encrypted from the evaluator, according to Gigazine's account of DeepMind's blog post. Neither party has to simply be trusted. The system produces cryptographic attestation instead.

BigGo Finance reports that a February 2026 NIST report flagged the underlying problem this is meant to address: common benchmark practices often fail to account for uncertainty or prevent AI models from overfitting to known test sets. Gigazine adds that AI agents have in some cases been caught actively hunting for outside information sources during tests specifically to inflate their scores.

DeepMind says the approach matters most for the highest-stakes evaluations, cybersecurity testing and government-run assessments in particular, where a lab's word that it didn't peek at the questions is worth the least and where evaluators may be legally barred from sharing sensitive material at all.

This is still a small-scale pilot: one lightweight model, four outside partners. Nothing about how frontier AI systems get tested has changed yet. Whether it becomes standard practice depends on whether AI labs are willing to submit their largest, most capable models, the ones where a compromised benchmark would matter most, to the same conditions.