GeForce RTX 5070 DLSS 5 benchmark: launch data

GeForce RTX 5070 DLSS 5 benchmark: what the launch number actually measures

The entire verified launch set for the desktop GeForce RTX 5070 is one figure: up to 260 FPS in NBA 2K27 at 2560×1440, Ultra, with ray tracing and the full DLSS stack, including DLSS 5. No other FPS result, percentile, or 1% low has been published for that exact combination.

DLSS 5 is scheduled to launch for the GeForce RTX 50 Series on September 3, 2026 at 21:00 Pacific Time. Until then, treat the 260 FPS figure as launch-gated and keep every other performance claim labelled as an inference or an internal measurement.

What the DLSS 5 runtime contains

  • Neural Rendering: a guided neural path replaces parts of rasterized or ray-traced shading with a learned approximation using scene features, motion vectors, and prior history.
  • Ray Reconstruction: a denoiser trained on ground-truth ray-traced samples cleans low-sample reflections, ambient occlusion, global illumination, and shadows.
  • Super Resolution and Frame Generation: reconstruction produces the final high-resolution image, while interpolation adds frames between rendered outputs.

DLSS 5 is presented as the first release in the family to combine these three neural components in one runtime feature rather than expose them only as separate toggles. All three run on the RTX 5070’s dedicated Tensor cores. The game engine still has to provide the correct inputs, so the driver alone cannot turn an incomplete integration into the published result. Teams should read NVIDIA’s launch material before assuming an unintegrated renderer will behave like the NBA 2K27 demonstration.

Read every part of the 260 FPS claim

“Up to” identifies a peak

The qualifier does not describe an average, median, or 1% low. It marks the highest rate observed during NVIDIA’s validation run under the documented conditions. A percentile result would be lower, and the gap can vary widely by scene.

Ultra defines the workload

NBA 2K27’s Ultra preset controls shadow-map resolution, reflection method, crowd density, arena materials, and the active ray-traced effects. Dropping to High or disabling ray tracing creates a different test whose result is not part of the launch claim.

2560×1440 fixes the output resolution

The RTX 5070 is positioned as a 1440p-class card. The same workload should run faster at 1920×1080 and slower at 3840×2160, but neither value has been published. The card can drive 4K in less demanding games; that does not turn the launch result into a 4K measurement.

The full stack means every named component is active

DLSS 5, Ray Reconstruction, Super Resolution, and Frame Generation are enabled together. Generated frames account for a large share of the delivered output at this rate. Disabling Frame Generation or Neural Rendering would lower the result, but the size of either drop is unknown.

The measurement belongs to one desktop SKU

The launch setup uses a desktop card, one GPU, and a consumer driver. It does not apply to laptop, Max-Q, Ti, SUPER, or other mobile and desktop variants. Board power, CUDA and Tensor counts, and memory configuration differ by SKU, and none of those internal figures can be inferred from the 260 FPS headline.

SKU Position in RTX 50 Series Launch DLSS 5 figure published? Workload used in launch figure
GeForce RTX 5070 Mainstream 1440p class Yes, up to 260 FPS in NBA 2K27 at 1440p Ultra with ray tracing and full DLSS stack NBA 2K27, Ultra preset, 2560×1440, ray tracing on, full DLSS stack including DLSS 5
GeForce RTX 5070 Ti Upper-mainstream 1440p to entry 4K Not part of this launch figure set No published DLSS 5 launch number
GeForce RTX 5080 High-end 4K class Not part of this launch figure set No published DLSS 5 launch number
GeForce RTX 5090 Flagship 4K and high refresh 4K Not part of this launch figure set No published DLSS 5 launch number

Only the RTX 5070 has a figure in this stated launch set. The RTX 50 Series reference provides family positioning, but no relative scaling between these SKUs should be derived from the 5070 result.

Runtime order matters

  1. The game renders color, depth, motion vectors, and prior history at an internal resolution chosen by driver policy or the title profile.
  2. Ray Reconstruction consumes low-sample ray-traced buffers and produces denoised outputs learned from ground truth.
  3. Super Resolution reconstructs the internal image at the display resolution.
  4. Frame Generation interpolates between reconstructed outputs using motion vectors and other features.
  5. DLSS 5 Neural Rendering integrates its learned approximation into shading and material evaluation, separating it from earlier releases.

The 260 FPS figure is the final output of this pipeline. It cannot be assigned to one component without a separate trace.

Why one figure is not a benchmark suite

  • DLSS 5 requires engine integration, and NBA 2K27 is the demonstration title within a small launch set.
  • Neural-path driver overhead depends on the engine integration, so cross-engine scaling is outside the launch claim.
  • Frame Generation depends on base rate, motion-vector quality, and scene complexity, making one game and preset the cleanest launch comparison.
  • The published number uses the words “up to” because it is a peak rather than a mean or percentile.

Decide whether the result applies to your project

Confirm engine support

The feature requires a supported engine branch or a confirmed custom integration. An older engine build will not expose Neural Rendering simply because the computer has an RTX 5070.

Stay in the same resolution band

A 1080p project should expect a higher result and a 4K project a lower one, but both need measurement. Variable-resolution rendering creates another curve, with 1440p serving as one point rather than a representative average.

Compare the scene with NBA 2K27

The launch title combines arena lighting, large crowds, player skin and hair, court reflections, and its own ray-tracing profile. A long-view first-person shooter and a top-down strategy game stress different parts of the renderer.

Check which DLSS components are enabled

A title with only Super Resolution and Frame Generation cannot reproduce a full-stack result that also uses Neural Rendering and Ray Reconstruction.

Integration checks for studios

  • Engine version: verify the shipping branch because an older integration may not expose Neural Rendering.
  • Driver policy: the driver and game profile must both report the feature flags, so a pre-launch driver will not enable the path.
  • Motion vectors: noisy or incorrect vectors reduce the value of Frame Generation and Neural Rendering before they affect headline performance.
  • Ray-traced effects: two Ultra presets can have very different costs depending on active effects and sample counts.
  • Internal resolution: changing it shifts work between Super Resolution, Neural Rendering, and the base renderer.

Expected direction of scaling, not predicted FPS

Workload Resolution Preset DLSS stack Frame rate expectation
NBA 2K27 Ultra with ray tracing, full DLSS stack 2560×1440 Ultra Full stack including DLSS 5 Up to 260 FPS (verified launch figure)
Same game and preset, full DLSS stack 1920×1080 Ultra Full stack including DLSS 5 Higher than 260 FPS, value not published
Same game and preset, full DLSS stack 3840×2160 Ultra Full stack including DLSS 5 Lower than 260 FPS, value not published
Same game and preset, Frame Generation disabled 2560×1440 Ultra DLSS 5, Ray Reconstruction, Super Resolution, no Frame Generation Lower than 260 FPS, value not published
Same game and preset, DLSS 5 neural rendering disabled 2560×1440 Ultra Ray Reconstruction, Super Resolution, Frame Generation only Lower than 260 FPS, value not published
Same game at High preset, full DLSS stack 2560×1440 High Full stack including DLSS 5 Higher than 260 FPS, value not published
Same game at Ultra preset, ray tracing disabled 2560×1440 Ultra Full stack including DLSS 5 Different from 260 FPS, value not published

The launch data supports the direction shown here, not an exact value for the unpublished rows.

What the launch claim excludes

  • 1% lows and 99th-percentile frame times.
  • Measured 1080p and 4K performance.
  • Performance with Frame Generation disabled.
  • Results from games other than NBA 2K27.
  • Thermally throttled, power-limited, or overclocked behavior.
  • Comparison figures for the RTX 5070 Ti, 5080, or 5090.
  • Results from competing GPU vendors.
  • Per-component attribution for the 260 FPS output.
  • Power draw, acoustics, and temperature during the run.

A reproducible project test

  1. Build a representative NBA 2K27 scene that exercises ray-traced effects, complex materials, and motion-vector sources similar to the project.
  2. Capture a fixed camera path at 2560×1440, Ultra, ray tracing on, and the full DLSS stack.
  3. Record peak, average, and 1% low for an internal reference.
  4. Run the same path in the studio’s title with matching feature flags and compare the shape of the results rather than the absolute FPS.
  5. Repeat with Frame Generation disabled to isolate its contribution.
  6. Repeat with Neural Rendering disabled to isolate that contribution.
  7. Repeat on the same hardware at 1920×1080 and 3840×2160 to map the resolution curve.

The purpose is a stable regression reference, not a requirement to reproduce 260 FPS.

Image quality still needs human review

Static cameras, low-frequency motion, and stable lighting tend to reconstruct cleanly. Fast camera movement, thin geometry, and sharp specular highlights expose errors first. Ray Reconstruction can smear when the sample count is too low, while generated frames can introduce their own visual faults. At very high output rates those differences become harder to see, but they do not disappear.

Latency-sensitive competitive first-person shooters deserve extra caution. NBA 2K27 has a different input profile, which is one reason it suits the demonstration.

Driver, profile, and engine must agree

The driver must post-date the September 3, 2026 release time, the game profile must expose DLSS 5, and the engine branch must support Neural Rendering. If any one is missing, the title will fall back to an earlier DLSS release or no DLSS path. A third-party engine also makes the studio dependent on its vendor’s integration schedule.

Genres will not scale alike

For a studio planning a 2026 or 2027 title, sports simulations, racing games, and slower RPGs are close to the launch profile because their camera motion and lighting suit the known failure cases. Fast shooters, fighting games, and competitive multiplayer need more visual and latency QA. Mobile, Max-Q, Ti, SUPER, and D variants require separate measurements.

Frequently asked questions

What is the verified RTX 5070 DLSS 5 launch result?

Up to 260 FPS in NBA 2K27 at 2560×1440, Ultra, with ray tracing and the full DLSS stack active. No other FPS value belongs to the verified launch set for that configuration.

Why does the claim say “up to”?

It is a peak observed during validation, not an average, median, or 1% low.

Does the RTX 5070 support DLSS 5?

It is officially supported as an RTX 50 Series product. Performance data remains gated until the September 3, 2026 release time.

Does 260 FPS apply to laptop or Max-Q cards?

No. It applies only to the desktop GeForce RTX 5070.

Does the result include Frame Generation?

Yes. Disabling it lowers the delivered rate, but the reduced value has not been published.

Will other DLSS 5 games reach 260 FPS?

Not necessarily. Every title has different engine, scene, and integration costs.

Does the result describe frame pacing?

No. Frame pacing, 1% lows, and 99th-percentile frame times require separate measurements.

How should developers use the number?

Use it as a reference for one known workload, then build a local scene and capture a full frame-time trace on the same card.

Can it predict RTX 5070 Ti, 5080, or 5090 performance?

No. Those cards have no published figure in this launch set.

Where is the official description?

NVIDIA’s DLSS 5 research page describes the launch workload. The RTX 50 Series reference records the 5070’s position in the lineup.


One response to “GeForce RTX 5070 DLSS 5 benchmark: launch data”

  1. […] GeForce RTX 5070 DLSS 5 benchmark: launch data […]

Leave a Reply

Your email address will not be published. Required fields are marked *