英伟达把 DLSS 5 的模型瘦了 24%:显存省两成,帧生成最多快 35%

📡 AI News 2026-10-18 4 min read

Nvidia slimmed down the DLSS 5 model by 24%: 20% less VRAM usage, frame generation up to 35% faster. Over the past two days, Nvidia dropped three sets of numbers. The DLSS 5 neural rendering model's size was cut by 24%, shrinking to about 118 million parameters, VRAM usage at 4K dropped by 20%, and the GPU compute it consumes is only half of the September version. On speed, the same model is twice as fast as September's, and compared to the first tech demo in March

💡 What You Will Learn

Nvidia slimmed down the DLSS 5 model by 24%: 20% less VRAM usage, frame generation up to 35% faster. Over the past two days, Nvidia dropped three sets of numbers. The DLSS 5 neural rendering model's s

📜 Table of Contents

Nvidia Slimmed Down the DLSS 5 Model by 24%: 20% Less VRAM, Frame Generation Up to 35% Faster

Nvidia dropped three sets of numbers over the past couple of days. The DLSS 5 neural rendering model has been cut down by 24% in size to roughly 118 million parameters, VRAM usage at 4K is down by 20%, and the GPU compute it consumes is now only half of what the September version needed. On speed, the same model is twice as fast as September's, and 10x faster cumulatively compared to the first tech demo back in March.

Launching alongside these numbers is Preset C, a new performance mode for frame generation that's up to 35% faster on the RTX 5060, and it works on both RTX 40 and 50 series cards. Trimming DLSS 5 on one hand and extending the life of existing cards on the other — those two things together are what this update is really about.

First, the slimming: a smaller model saves VRAM and compute

Item This update's change
Model size 24% smaller, ~118M parameters
4K VRAM usage 20% lower
GPU overhead Roughly halved
Model speed 2x the September version
Cumulative speedup 10x the March first demo
Image fidelity Nvidia says it's on par with the September version

On image quality, Nvidia only says "on par" and didn't release any frame-by-frame comparisons — we'll wait for third parties to verify.

35% is the peak, and there are caveats

Frame gen multiplier Old mode Preset C Gain
Frame gen off 83 fps N/A None
2x 120 fps 132 fps 10%
4x 197 fps 244 fps 23.9%
6x 255 fps 348 fps 36.5%

The test rig was an RTX 5060 8GB paired with a Ryzen 7 9800X3D, running Cyberpunk 2077 at 1440p, medium ray tracing, DLSS Super Resolution on Balanced. The higher the multiplier, the bigger the payoff. At 2x you only gain 10%; you need 6x to hit that 35%, so in practice you'll want 4x or above to feel a real difference.

The details are on the rendering side. With Preset C enabled, the baseline frame rate for regular rendering goes from 42.5 fps to 58 fps, and only then does the interpolated number jump from 255 fps to 348 fps. What's being saved is the compute cost of generating each frame — rendering itself didn't get faster. Nvidia also acknowledges the trade-off: image quality takes a hit, because what got swapped out is the optical flow model that handles the game UI and ray-traced effects.

What 40 series owners actually get

Feature RTX 40 series RTX 50 series
Frame gen Preset C Supported Supported
DLSS 5 neural rendering Not supported Supported, including RTX Spark laptops and GeForce NOW

A lot of coverage lumped both things into one headline, but the benefits are clearly split. 40 series owners get the frame generation speedup; DLSS 5, which can alter lighting and materials, remains 50 series exclusive. If you're on a 40 series card and normally run frame generation, you get some free frames out of this — but if you want to see the kind of visual changes DLSS 5 brings, you'll still need a new card.

This slimming matters more than the image quality controversy

DLSS 5 first showed up at GTC in March, and the harshest criticism was about its "one-click beauty filter" look — characters' faces got smoothed until they glowed, and social media even coined a term for it. The tech's real Achilles' heel at the time was its overhead. The debut demo ran on two RTX 5090s; after the September launch, a single 50 series card could run it, but at the cost of a 50% to 60% performance drop with neural rendering on, per Nvidia's own figures. Domestic media tested it on an RTX 5080 running NBA 2K27 at 4K max settings: 120 fps with DLSS 5 off, only around 45 fps with it on.

Cutting the overhead in half and saving 20% of VRAM solves exactly that problem, which is why this update is being pushed with NBA 2K27. Nvidia says the new model lets it hit up to 1.4x its original frame rate, going from 65 fps to over 90 fps at 4K. That figure covers just one game at one resolution — no comparisons for other cards or other resolutions were provided.

When you can use it, and how to turn it on

Date What's happening
October 16 New model rolls out to NBA 2K27, RTX Spark laptops, and GeForce NOW
October 23 DLSS 5 comes to RTX Remix, available to mod authors
November 5 Hogwarts Legacy gets DLSS 5 support; official new comparison shots show facial changes noticeably toned down versus the March demo
Late November Where Winds Meet follows

Frame gen Preset C is already in the latest NVIDIA App beta, so no need to wait. In settings, check "Access beta or experimental features," go to the Graphics page, find DLSS Override - Model Presets, click Custom, and select Preset C from the frame generation dropdown, then apply. You can enable it per game, or set it globally for all supported games at once.

Number Meaning
24% DLSS 5 model size reduction
20% 4K VRAM usage reduction
2x Model speed vs. the September version
10x Cumulative speedup vs. the March first demo
35% Max gain from Preset C on RTX 5060 at 6x frame gen

Which generation of card are you on? This time, you can only pick one: the 40 series' frame gen speedup or the 50 series' neural rendering.

Related Articles
2026-09-02
Nvidia RTX Spark chip benchmark results are out, multi-core is 12% faster than AMD, but Apple M5 Max is 35% faster than it
2026-08-08
Doubao, Qwen and Yuanbao All Shut Down Their AI Agents — From July 15, Your AI Companion Is Gone
2026-07-29
Hermes /learn: Give It a Folder, It Generates a Skill — AI That Learns How to Work
2026-08-10
AI Education Companies 2026: Who's Building the AI Classroom and What Actually Works
2026-08-19
233 Hospitals in Chongqing Adopted the Same AI — and the Happiest Users Are Seniors
2026-08-28
Memory prices are skyrocketing—is buying just one stick okay? Real-world test: a single DDR5 stick costs 8% performance, but X3D users lose only 2%

Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.

💬 Comments (0)

No comments yet. Be the first!

Login to comment