DLSS 5 Performance Data: Official Claims and the State of Independent Testing#

Hello everyone, I’m Zixing. This page is about DLSS 5’s performance, but two things must be said up front:

First, every number in this article comes from NVIDIA’s official press release, each tagged with its conditions; second, this site has seen no verifiable independent retest data, and officials have not published per-item overhead percentages for neural rendering — so this article offers no “measured frame rate table”, only “official claims + data limits”. An old article on this site once cited a set of loss figures of unknown provenance under the headline “measured performance collapse”; those figures were verified to have no traceable first-hand source, and this page has been rewritten with them deleted in full.

1. Item-by-Item Verification of the Official Performance Figures#

Below are all performance-related statements NVIDIA has given officially. When citing any one of them, cite it together with its conditions.

The item most likely to be garbled in circulation is the 370 FPS one. Its full condition chain: 4K resolution, ray tracing on, GeForce RTX 5090 D v2 (the China-market model), NBA 2K27, DLSS enabled. Drop any link and 370 FPS drifts off the official line — reading it as “any RTX 50 series card hits 370 FPS” or “all games hit 370 FPS” both over-generalize the official data.

2. The Direction of the Official Line: “Both at Once”, Not a “Trade-off”#

The old article on this site used the framing “trading performance for image quality”, describing neural rendering as an expensive image quality upgrade. That framing is the opposite of the official line. Per NVIDIA’s official statements:

• Combined with neural rendering, super resolution, and multi-frame generation, DLSS 5 can deliver higher frame rates and better image quality at the same time;
• In the 1080p Ultra preset with ray tracing scenario, nearly all RTX 50 series players can have both;
• No data on “how much performance is lost when enabled” has been published officially.

Therefore, this site no longer states any loss conclusion that lacks official data. If you see claims elsewhere that “enabling DLSS 5 costs dozens of percent of performance”, check whether the source is NVIDIA’s official line — at least as of this writing, that data has not been published.

3. The List of Data Not Published Officially#

In performance writing, being clear about “which data doesn’t exist” matters as much as giving data. NVIDIA has not published the following, and this site will not speculate:

4. The State of Independent Testing#

As of this update (2026-10-05), this site has seen no verifiable independent retest data. The standard for “verifiable” here is first-hand measurement with a stated source and test methodology — not secondhand conclusions or screenshots of unknown origin.

5. How the Technology Relates to Performance: The Official Explanation#

How can officials promise higher frame rates and better image quality at the same time? Breaking down DLSS 5’s operation per the official line:

In this pipeline, the key official emphasis is “guidance”: the AI paints using the 3D data the engine supplies, rather than guessing from a 2D image. As for this pipeline’s per-frame cost on specific hardware, no itemized breakdown has been published, and this article will not answer for officials.

5.1 Pipeline Comparison: DLSS 1–4 vs DLSS 5#

To understand why performance accounting cannot be carried across generations, first see how the two generations’ pipelines divide the work under the official line:

The two work on different objects: the former handles “the part of the image that is missing”, the latter takes part in “generating the final image”. So DLSS 4.5’s measurement framework cannot transfer directly to DLSS 5 — one reason this article records only the official line and does not carry over the old measurement framework.

6. Data Continuity with DLSS 4.5#

Introducing DLSS 5, officials cited one historical DLSS 4.5 figure: the DLSS 4.5 technology stack can use AI to draw 23 of every 24 pixels on screen. This is the only comparison point officially given for understanding the relationship between the two generations.

This site previously did a benchmark roundup of the DLSS 4.5 technology stack; see DLSS 4.5 Benchmarked: Transformer Model Gen 2 Image Quality Gains, Ray Tracing Performance Gains, and 6x Multi-Frame Generation Measured Data. A caution: DLSS 4.5’s measured conclusions cannot be applied directly to DLSS 5 — the two generations’ pipeline compositions differ, and cross-generation data citation misleads.

7. Data Usage Advice for Readers#

8. Update Log#

— Zixing, NTGame Hardware Reviews