Evidence rule
Every number is labelled as NVIDIA-published or measured by our own pipeline; unmeasured FPS and multipliers are not filled in.
Source-led independent guide
DLSS 5 benchmark evidence center
A useful benchmark page starts by naming the measurement. 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.
Local snapshot
One 3-second 720p public-domain clip was processed as 36 frames in 87.4 seconds on an RTX 4090 reference machine (2.43 seconds per frame).
Not real-time
That snapshot describes a batch workflow, not a real-time promise or a game benchmark.
Reproduction fields
Keep source media, dimensions, model/preset, hardware, driver, date, retries and output beside every result.
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.
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.
At-a-glance comparison
| Evidence type | What is recorded | What it can support |
|---|---|---|
| NVIDIA-published | Source URL, quotation, publication/check date | The vendor’s documented feature or hardware statement |
| Local measured | Input, output, software, hardware, time and date | One reproducible workflow observation |
| Unverified | No matching source or run record | No published conclusion |
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.