MagicBeanDev/Longshot

Watch

Fork 98

Game: DLSS Quality significantly slower than Game's Own at 100% and 67% render resolution (RTX 5090, v2.2.0-rc.1) #263

New issue

Open

opened 2026-11-13 18:22:10 +00:00 by batbatb · 0 comments

batbatb commented

2026-11-13 18:22:10 +00:00

Longshot version

Hookshot v2.2.0-rc.1

Channel

Release candidate

What kind of problem

Low fps or stutter

What happened

[COMPILED AND HEAVILY ASSISTED WITH THE HELP OF CHATGPT]

In Hookshot v2.2.0-rc.1, NVIDIA DLSS Quality consistently delivers substantially lower performance than the game's own upscaler, even when both use the same nominal 67% internal render resolution.

More notably, Game's Own at 100% render resolution maintains approximately 80 FPS, while DLSS Quality at 67% averages only 64 FPS. AMD FSR 3.1 Quality also provides virtually no average-FPS improvement over Game's Own at 100% in the tested scene.

My initial observation was that Game's Own at 100% also appears noticeably sharper than the lower-resolution alternatives, raising a question about their performance-to-image-quality tradeoff.

I performed repeated CapFrameX benchmarks and separate Vulkan GPU pass-timing captures to investigate.

This report focuses on performance and static reconstruction quality. Previously reported ghosting and motion artifacts are tracked separately in #2511.

System information

Component	Specification
Hookshot	v2.2.0-rc.1
Source revision	`006e7dfe`
GPU	NVIDIA GeForce RTX 5090
Driver	NVIDIA 621.24
CPU	AMD Ryzen 7 9800X3D
RAM	32 GB DDR5-6400
OS	Windows 11 Pro, build 26200
Renderer	Vulkan 1.4
Output resolution	3840×2160
Benchmark software	CapFrameX 1.9.2.4

Benchmark methodology

Fixed gameplay location and stationary camera, after initial shader/pipeline preparation.
Three approximately 60-second runs per configuration for the initial comparison.
One additional 60-second run per configuration for the matched-resolution experiment.
Output resolution maintained at 3840×2160.
Frame generation disabled, V-Sync disabled, and frame-rate cap set to Uncapped.
Sharpness set to 100%.
Same graphics configuration and scene across tests.
Average FPS calculated from the recorded MsBetweenPresents frame-time data.
1% Low FPS calculated from the average of the slowest 1% of frame times.

Test A — Game's Own 100% vs. DLSS/FSR Quality

Configuration	Average FPS	1% Low FPS	Average GPU power
Game's Own — 100%	80.20	67.92	213.4 W
NVIDIA DLSS — Quality (67%)	64.07	54.14	298.6 W
AMD FSR 3.1 — Quality (67%)	80.40	66.90	198.6 W

Each result is the arithmetic mean of three independent captures.

Observations:

DLSS Quality is approximately 20.1% slower than Game's Own at 100%, despite rendering from a nominal 2560×1440 input instead of 3840×2160.
DLSS also reports approximately 85 W higher GPU power consumption.
FSR Quality delivers virtually identical average FPS to Game's Own at 100%, with slightly lower 1% lows and GPU power consumption.
All configurations were highly consistent across their three runs.

Test B — Matched nominal 67% rendering resolution

Configuration	Average FPS	1% Low FPS	Average GPU power
Game's Own — 67%	85.86	71.22	183.0 W
NVIDIA DLSS — Quality	63.98	55.47	300.6 W
AMD FSR 3.1 — Quality	80.03	67.04	201.7 W

These are single-run measurements per configuration.

Observations:

DLSS Quality performs approximately 25.5% slower than Game's Own at the same nominal 67% render scale.
FSR Quality performs approximately 6.8% slower than Game's Own at 67%.
Reducing Game's Own from 100% to 67% improves its FPS by approximately 7.1%, suggesting that the initial 80 FPS result was not simply a fixed 80 FPS ceiling.

The DLSS and FSR diagnostic logs explicitly confirm 2560×1440 input and 3840×2160 output. The Game's Own 67% figure is based on the selected in-game setting.

Vulkan GPU profiling investigation

I performed three additional sessions using:

RX_VK_PASS_TIMING=1

For comparison, the following table uses the last 12 profiling reports from each configuration, representing 7,200 presented frames per mode.

GPU profiling metric	Game's Own 67%	FSR Quality	DLSS Quality
Total instrumented GPU time/frame	9.25 ms	9.71 ms	12.93 ms
Rendering command-buffer time	7.23 ms	7.85 ms	10.93 ms
Time within named rendering spans	6.84 ms	7.10 ms	6.76 ms
Time between named rendering spans	0.39 ms	0.76 ms	4.18 ms
Upload command-buffer time	2.02 ms	1.86 ms	2.00 ms

Notable profiling result

The most significant difference appears in GPU time between named rendering spans.

DLSS reports approximately 4.18 ms/frame in this category, compared with 0.39 ms for Game's Own and 0.76 ms for FSR.

FSR's reconstruction work is explicitly categorized by the profiler, accounting for approximately 0.40 ms/frame across seven named FSR GPU spans.

The DLSS-specific time is not similarly isolated. It appears principally in the unclassified interval between named spans.

Additionally, the normal replay statistics report CPU-side memory sync time around 5 ms/frame under DLSS, considerably above the approximately 1–2 ms/frame seen with the other configurations.

These results suggest that the DLSS/NGX rendering and synchronization path deserves investigation. However, the unclassified GPU time cannot be attributed exclusively to DLSS inference or any specific defect without further profiling.

The pass-timing instrumentation itself introduces additional overhead and synchronization. Therefore, these values are used to identify possible bottlenecks, not as direct substitutes for the uninstrumented CapFrameX frame-time measurements.

Static image-quality comparison

[Image Quality Comparison]

Screenshots will compare:

Game's Own at 100% render resolution.
NVIDIA DLSS at Quality / 67%.
AMD FSR 3.1 at Quality / 67%.

All images are captured at 3840×2160 output resolution, from the same camera position, using the same sharpening configuration.

Initial subjective observation: Game's Own at 100% appears sharper than DLSS/FSR Quality. The screenshots are intended to allow independent evaluation of static detail, edge reconstruction, and softening.

I am not presenting these static screenshots as evidence of temporal ghosting or motion artifacts, as those concerns are already covered by existing reports.

Expected vs. observed behavior

Expected: Lowering rendering resolution should reduce the underlying rendering workload. DLSS and FSR introduce their own reconstruction costs, so higher FPS is not guaranteed; however, their relative performance and resulting image quality should provide a useful tradeoff.

Observed: In the tested scene, DLSS Quality produces substantially less FPS and higher GPU power consumption than Game's Own at 100% rendering. FSR Quality offers virtually no average-FPS advantage over Game's Own at 100%. Native reconstruction also remains faster in the matched-resolution experiment.

It would be useful to determine whether this behavior is an expected consequence of reconstruction costs, an integration/synchronization limitation, or an optimization opportunity.

Questions for maintainers

Is the substantially higher DLSS GPU execution time expected for the current NGX integration on an RTX 5090?
Could the approximately 4 ms/frame interval between named GPU spans under DLSS include external NGX work, synchronization, or other uninstrumented commands?
Is the elevated CPU-side memory sync time a downstream consequence of longer GPU work, or potentially a separate bottleneck?
Is there a known reason why FSR Quality provides nearly identical FPS to Game's Own at 100% rendering in this scenario?

Supporting evidence

Hookshot_upscaler_benchmark_evidence.zip — original CapFrameX JSON captures, calculated per-run metrics, and the three corrected GPU profiling logs.
[Image Quality Comparison]

Related report: #2511 — Upscaler availability, ghosting, lines, and graphical artifacts.

The intention of this report is to provide reproducible performance measurements and profiling evidence, rather than assume the upscaler implementations themselves are defective.

Where in the game

Kokiri Forest, Standing

How often

Every time

Graphics card and driver version

NVIDIA GeForce RTX 5090, NVIDIA 621.24

CPU, RAM and operating system

AMD Ryzen 7 9800X3D, 32 GB DDR5-6400, Windows 11 Pro, build 26200

Changes from the defaults

Sampler cache: On (default: Off)
Robust buffers: Off (default: Auto)
VRAM budget: 11 GB (default: Auto)
V-Sync: Off (default: Auto)
Fullscreen: On (default: Off)
Game frame rate cap: Uncapped (default: 60 (the console's))
Frame pacing: On (default: Off)
Sharpness: 100% (default: 0%)
Output resolution: 3840x2160 (native) (default: 2048x1152 (game's))
Aspect ratio: 16:9 (monitor's) (default: 4:3)

longshot_log.txt

Screenshots or video

No response

Before you send it

The problem is still there on the newest release or nightly.
I searched the open issues for it.
I'm not attaching game files, keys, the game's executable, its .bin caches or saves.

### Longshot version Hookshot v2.2.0-rc.1 ### Channel Release candidate ### What kind of problem Low fps or stutter ### What happened **[COMPILED AND HEAVILY ASSISTED WITH THE HELP OF CHATGPT]** In Hookshot v2.2.0-rc.1, NVIDIA DLSS Quality consistently delivers substantially lower performance than the game's own upscaler, even when both use the same nominal 67% internal render resolution. More notably, **Game's Own at 100% render resolution maintains approximately 80 FPS, while DLSS Quality at 67% averages only 64 FPS**. AMD FSR 3.1 Quality also provides virtually no average-FPS improvement over Game's Own at 100% in the tested scene. My initial observation was that Game's Own at 100% also appears noticeably sharper than the lower-resolution alternatives, raising a question about their performance-to-image-quality tradeoff. I performed repeated CapFrameX benchmarks and separate Vulkan GPU pass-timing captures to investigate. This report focuses on performance and static reconstruction quality. Previously reported ghosting and motion artifacts are tracked separately in #2511. ## System information | Component | Specification | |---|---| | Hookshot | v2.2.0-rc.1 | | Source revision | `006e7dfe` | | GPU | NVIDIA GeForce RTX 5090 | | Driver | NVIDIA 621.24 | | CPU | AMD Ryzen 7 9800X3D | | RAM | 32 GB DDR5-6400 | | OS | Windows 11 Pro, build 26200 | | Renderer | Vulkan 1.4 | | Output resolution | 3840×2160 | | Benchmark software | CapFrameX 1.9.2.4 | ## Benchmark methodology - Fixed gameplay location and stationary camera, after initial shader/pipeline preparation. - Three approximately 60-second runs per configuration for the initial comparison. - One additional 60-second run per configuration for the matched-resolution experiment. - Output resolution maintained at 3840×2160. - Frame generation disabled, V-Sync disabled, and frame-rate cap set to Uncapped. - Sharpness set to 100%. - Same graphics configuration and scene across tests. - Average FPS calculated from the recorded `MsBetweenPresents` frame-time data. - 1% Low FPS calculated from the average of the slowest 1% of frame times. ### Test A — Game's Own 100% vs. DLSS/FSR Quality | Configuration | Average FPS | 1% Low FPS | Average GPU power | |---|---:|---:|---:| | Game's Own — 100% | **80.20** | **67.92** | 213.4 W | | NVIDIA DLSS — Quality (67%) | 64.07 | 54.14 | 298.6 W | | AMD FSR 3.1 — Quality (67%) | 80.40 | 66.90 | 198.6 W | Each result is the arithmetic mean of three independent captures. **Observations:** - DLSS Quality is approximately **20.1% slower** than Game's Own at 100%, despite rendering from a nominal 2560×1440 input instead of 3840×2160. - DLSS also reports approximately 85 W higher GPU power consumption. - FSR Quality delivers virtually identical average FPS to Game's Own at 100%, with slightly lower 1% lows and GPU power consumption. - All configurations were highly consistent across their three runs. ### Test B — Matched nominal 67% rendering resolution | Configuration | Average FPS | 1% Low FPS | Average GPU power | |---|---:|---:|---:| | Game's Own — 67% | **85.86** | **71.22** | 183.0 W | | NVIDIA DLSS — Quality | 63.98 | 55.47 | 300.6 W | | AMD FSR 3.1 — Quality | 80.03 | 67.04 | 201.7 W | These are single-run measurements per configuration. **Observations:** - DLSS Quality performs approximately **25.5% slower** than Game's Own at the same nominal 67% render scale. - FSR Quality performs approximately **6.8% slower** than Game's Own at 67%. - Reducing Game's Own from 100% to 67% improves its FPS by approximately 7.1%, suggesting that the initial 80 FPS result was not simply a fixed 80 FPS ceiling. The DLSS and FSR diagnostic logs explicitly confirm 2560×1440 input and 3840×2160 output. The Game's Own 67% figure is based on the selected in-game setting. ## Vulkan GPU profiling investigation I performed three additional sessions using: `RX_VK_PASS_TIMING=1` For comparison, the following table uses the **last 12 profiling reports from each configuration**, representing 7,200 presented frames per mode. | GPU profiling metric | Game's Own 67% | FSR Quality | DLSS Quality | |---|---:|---:|---:| | Total instrumented GPU time/frame | 9.25 ms | 9.71 ms | **12.93 ms** | | Rendering command-buffer time | 7.23 ms | 7.85 ms | **10.93 ms** | | Time within named rendering spans | 6.84 ms | 7.10 ms | 6.76 ms | | Time between named rendering spans | **0.39 ms** | 0.76 ms | **4.18 ms** | | Upload command-buffer time | 2.02 ms | 1.86 ms | 2.00 ms | ### Notable profiling result The most significant difference appears in **GPU time between named rendering spans**. DLSS reports approximately 4.18 ms/frame in this category, compared with 0.39 ms for Game's Own and 0.76 ms for FSR. FSR's reconstruction work is explicitly categorized by the profiler, accounting for approximately 0.40 ms/frame across seven named FSR GPU spans. The DLSS-specific time is not similarly isolated. It appears principally in the unclassified interval between named spans. Additionally, the normal replay statistics report CPU-side `memory sync` time around 5 ms/frame under DLSS, considerably above the approximately 1–2 ms/frame seen with the other configurations. These results suggest that the DLSS/NGX rendering and synchronization path deserves investigation. However, **the unclassified GPU time cannot be attributed exclusively to DLSS inference or any specific defect without further profiling**. The pass-timing instrumentation itself introduces additional overhead and synchronization. Therefore, these values are used to identify possible bottlenecks, not as direct substitutes for the uninstrumented CapFrameX frame-time measurements. ## Static image-quality comparison [[Image Quality Comparison]](https://imgico.com/a/eyQmE82) Screenshots will compare: - Game's Own at 100% render resolution. - NVIDIA DLSS at Quality / 67%. - AMD FSR 3.1 at Quality / 67%. All images are captured at 3840×2160 output resolution, from the same camera position, using the same sharpening configuration. **Initial subjective observation:** Game's Own at 100% appears sharper than DLSS/FSR Quality. The screenshots are intended to allow independent evaluation of static detail, edge reconstruction, and softening. I am not presenting these static screenshots as evidence of temporal ghosting or motion artifacts, as those concerns are already covered by existing reports. ## Expected vs. observed behavior **Expected:** Lowering rendering resolution should reduce the underlying rendering workload. DLSS and FSR introduce their own reconstruction costs, so higher FPS is not guaranteed; however, their relative performance and resulting image quality should provide a useful tradeoff. **Observed:** In the tested scene, DLSS Quality produces substantially less FPS and higher GPU power consumption than Game's Own at 100% rendering. FSR Quality offers virtually no average-FPS advantage over Game's Own at 100%. Native reconstruction also remains faster in the matched-resolution experiment. It would be useful to determine whether this behavior is an expected consequence of reconstruction costs, an integration/synchronization limitation, or an optimization opportunity. ## Questions for maintainers 1. Is the substantially higher DLSS GPU execution time expected for the current NGX integration on an RTX 5090? 2. Could the approximately 4 ms/frame interval between named GPU spans under DLSS include external NGX work, synchronization, or other uninstrumented commands? 3. Is the elevated CPU-side `memory sync` time a downstream consequence of longer GPU work, or potentially a separate bottleneck? 4. Is there a known reason why FSR Quality provides nearly identical FPS to Game's Own at 100% rendering in this scenario? ## Supporting evidence - `Hookshot_upscaler_benchmark_evidence.zip` — original CapFrameX JSON captures, calculated per-run metrics, and the three corrected GPU profiling logs. - [[Image Quality Comparison]](https://imgico.com/a/eyQmE82) Related report: #2511 — Upscaler availability, ghosting, lines, and graphical artifacts. The intention of this report is to provide reproducible performance measurements and profiling evidence, rather than assume the upscaler implementations themselves are defective. ### Where in the game Kokiri Forest, Standing ### How often Every time ### Graphics card and driver version NVIDIA GeForce RTX 5090, NVIDIA 621.24 ### CPU, RAM and operating system AMD Ryzen 7 9800X3D, 32 GB DDR5-6400, Windows 11 Pro, build 26200 ### Changes from the defaults Sampler cache: On (default: Off) Robust buffers: Off (default: Auto) VRAM budget: 11 GB (default: Auto) V-Sync: Off (default: Auto) Fullscreen: On (default: Off) Game frame rate cap: Uncapped (default: 60 (the console's)) Frame pacing: On (default: Off) Sharpness: 100% (default: 0%) Output resolution: 3840x2160 (native) (default: 2048x1152 (game's)) Aspect ratio: 16:9 (monitor's) (default: 4:3) ### longshot_log.txt . ### Screenshots or video _No response_ ### Before you send it - [x] The problem is still there on the newest release or nightly. - [x] I searched the open issues for it. - [x] I'm not attaching game files, keys, the game's executable, its .bin caches or saves.

Hookshot_upscaler_benchmark_evidence.zip

1.4 MiB

hookshot_log.txt

2 MiB

batbatb added the

kind

bug

label

2026-11-13 18:22:10 +00:00