Source-led independent guide

DLSS 5 benchmark evidence center

A useful benchmark page starts by naming the measurement. This page has not run a frame-rate test in any game; this site has no DLSS 5 game runtime. This center separates vendor-published specifications from our own image-pipeline timing and leaves unsupported frame-rate claims blank.

An evidence center for DLSS 5 claims: NVIDIA-published facts, our measured image pipeline snapshot, reproducible fields, and limits on what we can conclude.

What we have measured

One Studio batch snapshot: a 3-second 720p public-domain clip, 36 frames, an RTX 4090 reference machine, 87.4 seconds (2.43 seconds per frame), recorded on 2026-09-25; batch processing time, not real-time performance.

What we have not measured

Frame rates in any game; in-game DLSS 5 versus DLSS 4.5 comparisons; DLSS 5 performance on RTX 40 / 30 / 20 Series; any multiplier conclusions.

What cannot be inferred

NVIDIA’s up to 1.4X statement is scoped to NBA 2K27 + RTX 50 Series + RTX Spark laptops + GeForce NOW; its 10X, −24% and −20% statements also have specific scopes, are not our observations, and cannot predict other games or GPUs.

Why a single number is not a universal benchmark

DLSS 5 is a game-rendering technology with engine, driver, GPU and game-build dependencies. An image conversion pipeline has a different input, model path and output contract. Putting those numbers in one leaderboard would make the result look more precise than the evidence allows. This page therefore uses an evidence ledger: the source URL and date for a vendor statement, or the complete run fields for a local measurement. A missing field stays unknown rather than being filled from a review, a screenshot or a guessed conversion.

“Ten times faster in half a year”: the Chinese coverage carries three numbers

On 2026-10-19 Leikeji, published through Sina Finance, wrote that over roughly half a year the DLSS 5 model had become “a full ten times” faster, citing the 13 October update: 24% fewer parameters, 20% less VRAM, and a further doubling over the September version (source: https://cj.sina.com.cn/articles/view/5061312402/12dad7f9201902mi9g , 2026-10-19). Two days later ZAKER aggregated community tests with different numbers: The Witcher 3 up to 90% faster when comparing the October model (“preset B”) against the September one (“preset A”), about 40% in other games, and roughly 25% in NBA 2K27 at 1440p in PC Gamer’s test (1080p from 143 fps to 168 fps, still under the 227 fps baseline with DLSS off), noting those comparisons came from injecting DLLs through RenoDX and OptiScaler (source: https://www.sohu.com/a/1090833276_120333371 , 2026-10-21). They describe different things: ten times is the model against its March version; 90% and 25% are frame-rate changes for named games under a named comparison method. This site has no game runtime and publishes no frame rates; it does not merge the three into one “how much faster is DLSS 5” claim. For your game, only that game’s own measurement counts, not this news.

The measured local snapshot

Our published local observation is deliberately narrow: a three-second, 720p public-domain clip was processed as 36 frames in 87.4 seconds on an RTX 4090 reference machine, or 2.43 seconds per frame. It describes one batch run through the local Studio workflow at the recorded time. It does not say the pipeline is real-time, does not compare another GPU, and does not establish an in-game DLSS 5 frame rate. Repeating it requires the same source clip, output target, software version and hardware record.

What a future comparison must include

A credible run should preserve the original input and output, state the requested dimensions, name the model or preset, record the GPU and driver, include queue and retry time, and identify whether the result came from a game frame, a browser image workflow or local Studio. Review quality and speed separately. If there is no direct measurement, link the published source and mark the field unavailable. This keeps the center useful for AI readers without pretending that an illustrative image proves a performance claim.

Why there are no in-game frame rates here

We do not have the game runtime and controlled comparison scenes, so we do not publish unmeasured frame rates. Browser image results, the local Studio batch snapshot and in-game DLSS 5 are separate evidence tracks; none validates the others. Every number must be labelled as NVIDIA-published or measured by our own pipeline. For comparison, preserve source media, dimensions, model/preset, hardware, driver, software version, date, queue/retry time and output; keep unmeasured FPS and multipliers blank. The four quotations in the following table come from NVIDIA’s October update article: https://www.nvidia.com/en-us/geforce/news/dlss-5-faster-ai-model-hogwarts-legacy-rtx-spark-laptops/ Checked 2026-10-26. These are NVIDIA’s statements, not observations made by this site.

At-a-glance comparison

Official wording (verbatim)What it saysSource and check date
DLSS 5’s 3D-Guided Neural Rendering model is now 10X faster overall compared to when it was announced in MarchNVIDIA says the model overall is faster than its March version; this does not mean games run ten times faster.NVIDIA · 2026-10-13 · checked 2026-10-26
The DLSS 5 October update shrinks the footprint by 24% to roughly 118 million parameters while reducing 4K VRAM usage by 20%NVIDIA reports a 24% smaller footprint, roughly 118 million parameters and 20% lower 4K VRAM usage.NVIDIA · 2026-10-13 · checked 2026-10-26
The result is half the GPU overhead of the September version while maintaining comparable visual fidelityNVIDIA reports half the GPU overhead relative to September, with comparable visual fidelity.NVIDIA · 2026-10-13 · checked 2026-10-26
boosting frame rates by up to 1.4XNVIDIA’s up to 1.4X frame-rate statement applies to NBA 2K27 + RTX 50 Series + RTX Spark laptops + GeForce NOW, not other games or GPUs.NVIDIA · 2026-10-13 · checked 2026-10-26

Turn the guide into a reviewable test

Inspect a cached example before uploading your own image; official technology claims and this site’s independent conversion results stay clearly separated.

Sources and boundaries

This is an independent reference. It is not an NVIDIA official page or partnership.

Frequently asked questions

Is this an FPS leaderboard?

No. The page does not publish unmeasured game FPS, multipliers or cross-GPU rankings.

What local measurement is available?

One RTX 4090 reference run processed 36 frames from a 3-second 720p clip in 87.4 seconds; it is batch context.

Can I compare a browser image result with DLSS 5 in a game?

No. They have different inputs and runtime contracts, so the page keeps them as separate evidence tracks.

What should I record for a new run?

Record input, output dimensions, model or preset, GPU, driver, software version, queue/retry time and date.

Why are some fields blank?

A blank or unknown field means no source or run record was available; it is not an invitation to infer a value.

Where are the source links?

Vendor references and the local workflow boundary are linked below; each future result should add its own evidence record.