Switch 2 Found to Use Desktop-Grade DLSS 4: Input Features Reduced by 75%, Presets H/I as Two Low-Load Modes
IT之家 November 17 news, developer Eren Doğan posted on X yesterday (November 16), disclosing that he extracted the 英伟达 DLSS models from 任天堂 Switch 2 and found that they include the first-generation Transformer model introduced with DLSS 4, with model weights identical to the PC desktop version, but no DLSS 4.5 models were found.

According to the mined code information, Switch 2 has two exclusive presets named H and I. According to Doğan's analysis, they use a convolutional neural network, i.e. a CNN model, with a model size the same as the regular desktop version, but the input feature grid size is only one quarter of the latter, thus reducing the computational load and speeding up processing; the actual speedup still awaits testing and verification. He believes this may be the lightweight solution previously referred to as "DLSS Lite" or "Tiny DLSS."


IT之家 found in public materials that the input feature grid size refers to how large an information grid the AI needs to analyze when processing a frame. The larger the grid, the more information typically needs to be processed; shrinking the grid can potentially reduce the computational load.
Related Articles
- AMD FSR 4.1 upscaling + 4th-gen frame generation technology lands on the Vulkan engine for the first time: DOOM: The Dark Ages game update 5 goes live
- 9-year-old boy spends CAD 28,000 on in-game purchases, parents' refund request rejected by Roblox
- After a cyberattack, 11 Sega Japan arcade games gradually restore online features
- GDC survey: 40% of game industry professionals use AI, 55% hold a critical attitude toward it
- Huawei HarmonyOS 7 game instant-start toggle goes live, supports preloading the two most popular games on the device
- Prediction market platform Kalshi recruits a GTA 6 researcher, requiring a full livestream of gameplay and analysis of in-game and out-of-game economies