DLSS Preset M / K are broken #17

New issue

Open

opened 2026-11-12 23:24:35 +00:00 by Schiggy · 5 comments

Schiggy commented 2026-11-12 23:24:35 +00:00

• Copy link
• Copy Markdown

Longshot version

2.1.0

What kind of problem

Low fps or stutter

What happened

Preset L and M seem broken.
Once you switch from Preset K to L the VRAM increases for a moment extremly but then comes down to a normal level that is still under 11500 GB.
GPU Power Draw increases by 36%
FPS Drop by 20%

This doesnt happen for me at Ultra Performance Mode there everything works very normal.
Preset L / M normaly use 200-300mb more VRAM and decrease performance by 4% if you have no GPU headroom left.
I have GPU headroom left and my VRAM doesnt seem full except for the short moment in the beginning.

Where in the game

Kokiri Forest Balcony

How often

Every time

Graphics card and driver version

RTX 5070 NVIDIA 617.42

CPU, RAM and operating system

5700X3D 32GB Win 11

Changes from the defaults

No response

Other software hooked into the game

No, nothing hooked in

longshot_log.txt

hookshot_log.1

Screenshots or video

Preset K DLAA

Preset L DLAA

Preset K DLSS Ultraperformance

Preset L DLSS Ultraperformance

Before you send it

• I'm on the newest release, or the problem is still there on it.
• I searched the open issues for this problem.
• I'm not attaching game files, keys, longshot.exe, the .bin caches or saves.

### Longshot version 2.1.0 ### What kind of problem Low fps or stutter ### What happened Preset L and M seem broken. Once you switch from Preset K to L the VRAM increases for a moment extremly but then comes down to a normal level that is still under 11500 GB. GPU Power Draw increases by 36% FPS Drop by 20% This doesnt happen for me at Ultra Performance Mode there everything works very normal. Preset L / M normaly use 200-300mb more VRAM and decrease performance by 4% if you have no GPU headroom left. I have GPU headroom left and my VRAM doesnt seem full except for the short moment in the beginning. ### Where in the game Kokiri Forest Balcony ### How often Every time ### Graphics card and driver version RTX 5070 NVIDIA 617.42 ### CPU, RAM and operating system 5700X3D 32GB Win 11 ### Changes from the defaults _No response_ ### Other software hooked into the game No, nothing hooked in ### longshot_log.txt [hookshot_log.1](/attachments/11333970-e2eb-4fda-8b6d-41e5aea4cb99) ### Screenshots or video Preset K DLAA  Preset L DLAA  Preset K DLSS Ultraperformance  Preset L DLSS Ultraperformance  ### Before you send it - [x] I'm on the newest release, or the problem is still there on it. - [x] I searched the open issues for this problem. - [x] I'm not attaching game files, keys, longshot.exe, the .bin caches or saves.

hookshot_log.1.txt

5.5 MiB

Screenshot 2026-11-13 001605.png

3.1 MiB

Screenshot 2026-11-13 001651.png

3.2 MiB

Screenshot 2026-11-13 001724.png

2.9 MiB

Screenshot 2026-11-13 001742.png

3 MiB

Schiggy added the

bug

label 2026-11-12 23:24:35 +00:00

llun commented 2026-11-12 23:32:49 +00:00

Contributor

• Copy link
• Copy Markdown

Taking a look at this (llun's agent; RTX 5080, Windows 11).

From your log so far, at 3440x1440 DLAA: GPU time is about 10.9 ms a frame with K and 15.3 ms with L, and DLSS's own memory goes from 189 MB to 631 MB, with a peak of 1419 MB right after the switch, while the model row passes M on its way to L (a feature is made for each, 11 frames apart). I'll check which part of that is how Longshot drives DLSS and which is the model's own cost (as measured on an RTX 3060 Ti in https://github.com/NotRexed/Hookshot/issues/2836), and reproduce it on an RTX 5080.

Taking a look at this (llun's agent; RTX 5080, Windows 11). From your log so far, at 3440x1440 DLAA: GPU time is about 10.9 ms a frame with K and 15.3 ms with L, and DLSS's own memory goes from 189 MB to 631 MB, with a peak of 1419 MB right after the switch, while the model row passes M on its way to L (a feature is made for each, 11 frames apart). I'll check which part of that is how Longshot drives DLSS and which is the model's own cost (as measured on an RTX 3060 Ti in https://github.com/NotRexed/Hookshot/issues/2836), and reproduce it on an RTX 5080.

Schiggy commented 2026-11-12 23:37:14 +00:00

Author

• Copy link
• Copy Markdown

@llun wrote in #17 (comment):

Taking a look at this (llun's agent; RTX 5080, Windows 11).

From your log so far, at 3440x1440 DLAA: GPU time is about 10.9 ms a frame with K and 15.3 ms with L, and DLSS's own memory goes from 189 MB to 631 MB, with a peak of 1419 MB right after the switch, while the model row passes M on its way to L (a feature is made for each, 11 frames apart). I'll check which part of that is how Longshot drives DLSS and which is the model's own cost (as measured on an RTX 3060 Ti in https://github.com/NotRexed/Hookshot/issues/2836), and reproduce it on an RTX 5080.

Its too late I dont know if you said that or the post said that but a 3060 runs Preset L and M worse than a 5080 will.
3000 series scales worse on M and L so my comment of it being a 4% cost compared to Preset K only applies to 4000 and 5000 series RTX.

@llun wrote in https://git.magicbean.dev/MagicBeanDev/Longshot/issues/17#issuecomment-11825: > Taking a look at this (llun's agent; RTX 5080, Windows 11). > > From your log so far, at 3440x1440 DLAA: GPU time is about 10.9 ms a frame with K and 15.3 ms with L, and DLSS's own memory goes from 189 MB to 631 MB, with a peak of 1419 MB right after the switch, while the model row passes M on its way to L (a feature is made for each, 11 frames apart). I'll check which part of that is how Longshot drives DLSS and which is the model's own cost (as measured on an RTX 3060 Ti in https://github.com/NotRexed/Hookshot/issues/2836), and reproduce it on an RTX 5080. Its too late I dont know if you said that or the post said that but a 3060 runs Preset L and M worse than a 5080 will. 3000 series scales worse on M and L so my comment of it being a 4% cost compared to Preset K only applies to 4000 and 5000 series RTX.

llun commented 2026-11-12 23:42:32 +00:00

Contributor

• Copy link
• Copy Markdown

@Schiggy fair point, the 3060 Ti was only for context: RTX 30 cards run L and M worse, so here are your own RTX 5070's numbers from your log instead (same spot, GPU time a frame):

Mode	input	K	L	L over K
Ultra Performance	1147x480 (0.55 MP)	8.4 ms	8.6–8.9 ms	about +0.4 ms
DLAA	3440x1440 (4.95 MP)	10.8–11.0 ms	15.2–15.4 ms	about +4.4 ms

L's extra cost grows with the pixels DLSS gets, about 0.8–0.9 ms per million in both rows, so DLAA at 3440x1440 gives it nine times the input of Ultra Performance. The few-percent figures you've seen are probably from upscaling modes, where the input is a quarter of the output or less. That points at the model's own cost at DLAA rather than something Longshot passes wrong: the preset, the sizes and the flags in your log all look right. I'll check that on an RTX 5080 at DLAA with K, M and L before saying it for sure.

What does look like ours is the memory jump right after switching. The model row goes K → M → L, so picking L made an M feature on the way, and NVIDIA's runtime holds a dropped feature's memory for a few seconds: 1419 MB at once against 631 MB for L alone, on a card that was already close to its limit. I'm testing a fix that waits until the row has stopped on a model before making it, and I'll post before and after numbers.

@Schiggy fair point, the 3060 Ti was only for context: RTX 30 cards run L and M worse, so here are your own RTX 5070's numbers from your log instead (same spot, GPU time a frame): | Mode | input | K | L | L over K | |---|---|---|---|---| | Ultra Performance | 1147x480 (0.55 MP) | 8.4 ms | 8.6–8.9 ms | about +0.4 ms | | DLAA | 3440x1440 (4.95 MP) | 10.8–11.0 ms | 15.2–15.4 ms | about +4.4 ms | L's extra cost grows with the pixels DLSS gets, about 0.8–0.9 ms per million in both rows, so DLAA at 3440x1440 gives it nine times the input of Ultra Performance. The few-percent figures you've seen are probably from upscaling modes, where the input is a quarter of the output or less. That points at the model's own cost at DLAA rather than something Longshot passes wrong: the preset, the sizes and the flags in your log all look right. I'll check that on an RTX 5080 at DLAA with K, M and L before saying it for sure. What does look like ours is the memory jump right after switching. The model row goes K → M → L, so picking L made an M feature on the way, and NVIDIA's runtime holds a dropped feature's memory for a few seconds: 1419 MB at once against 631 MB for L alone, on a card that was already close to its limit. I'm testing a fix that waits until the row has stopped on a model before making it, and I'll post before and after numbers.

llun referenced this issue 2026-11-12 23:51:43 +00:00

fix(dlss): the DLSS model row's preset is taken once it holds it, not a feature for each one stepped past #39

Schiggy commented 2026-11-12 23:53:33 +00:00

Author

• Copy link
• Copy Markdown

@llun wrote in #17 (comment):

@Schiggy fair point, the 3060 Ti was only for context: RTX 30 cards run L and M worse, so here are your own RTX 5070's numbers from your log instead (same spot, GPU time a frame):
Mode input K L L over K
Ultra Performance 1147x480 (0.55 MP) 8.4 ms 8.6–8.9 ms about +0.4 ms
DLAA 3440x1440 (4.95 MP) 10.8–11.0 ms 15.2–15.4 ms about +4.4 ms

L's extra cost grows with the pixels DLSS gets, about 0.8–0.9 ms per million in both rows, so DLAA at 3440x1440 gives it nine times the input of Ultra Performance. The few-percent figures you've seen are probably from upscaling modes, where the input is a quarter of the output or less. That points at the model's own cost at DLAA rather than something Longshot passes wrong: the preset, the sizes and the flags in your log all look right. I'll check that on an RTX 5080 at DLAA with K, M and L before saying it for sure.

What does look like ours is the memory jump right after switching. The model row goes K → M → L, so picking L made an M feature on the way, and NVIDIA's runtime holds a dropped feature's memory for a few seconds: 1419 MB at once against 631 MB for L alone, on a card that was already close to its limit. I'm testing a fix that waits until the row has stopped on a model before making it, and I'll post before and after numbers.

I see but is GPU utilization not showing the full picture here ?
Because my GPU is not nearly 100% maxed out and yet I lose FPS majorly.
I thought my GPU overhead would catch that increased cost.

@llun wrote in https://git.magicbean.dev/MagicBeanDev/Longshot/issues/17#issuecomment-11860: > @Schiggy fair point, the 3060 Ti was only for context: RTX 30 cards run L and M worse, so here are your own RTX 5070's numbers from your log instead (same spot, GPU time a frame): > Mode input K L L over K > Ultra Performance 1147x480 (0.55 MP) 8.4 ms 8.6–8.9 ms about +0.4 ms > DLAA 3440x1440 (4.95 MP) 10.8–11.0 ms 15.2–15.4 ms about +4.4 ms > > L's extra cost grows with the pixels DLSS gets, about 0.8–0.9 ms per million in both rows, so DLAA at 3440x1440 gives it nine times the input of Ultra Performance. The few-percent figures you've seen are probably from upscaling modes, where the input is a quarter of the output or less. That points at the model's own cost at DLAA rather than something Longshot passes wrong: the preset, the sizes and the flags in your log all look right. I'll check that on an RTX 5080 at DLAA with K, M and L before saying it for sure. > > What does look like ours is the memory jump right after switching. The model row goes K → M → L, so picking L made an M feature on the way, and NVIDIA's runtime holds a dropped feature's memory for a few seconds: 1419 MB at once against 631 MB for L alone, on a card that was already close to its limit. I'm testing a fix that waits until the row has stopped on a model before making it, and I'll post before and after numbers. I see but is GPU utilization not showing the full picture here ? Because my GPU is not nearly 100% maxed out and yet I lose FPS majorly. I thought my GPU overhead would catch that increased cost.

llun commented 2026-11-13 00:43:23 +00:00

Contributor

• Copy link
• Copy Markdown

@Schiggy measured on an RTX 5080 (Windows 11, driver 621.24) at your settings: 3440x1440, DLAA, Kokiri Forest, 4 runs.

Model	GPU a frame	fps	power	GPU utilization
K	7.4–7.6 ms	80–84	140–145 W	74%
M	9.7–9.9 ms	70–72	186–188 W	79%
L	10.6–10.7 ms	66–67	198–202 W	81%

So the same shape as yours: L costs about 20% of the fps and about 40% more power than K at DLAA, on a 5080 too. That part is the model's own work at DLAA's input size, and it is the same with the fix below.

On utilization: you're right that it doesn't show the full picture. Here it only went from 74% to 81%, yet the frame got about 3 ms longer, the same as DLSS's extra GPU time. In Longshot a frame's GPU work isn't overlapped with the next frame's preparation: the frame is presented only after DLSS has run, and the next one isn't started until then. So any extra GPU time lands on the frame time even though the GPU sits idle at other moments in the frame. Utilization counts those idle moments; fps doesn't. With a game that overlaps frames, the headroom you see would absorb it. That overlap is being worked on separately (frame overlap), not in this issue.

The memory jump at the switch was ours, and it's fixed in #39: stepping K → M → L no longer makes M's feature on the way. VRAM rises about 0.64 GB instead of about 1.2 GB for the first seconds after the switch (DLSS's own memory 813 MB instead of 1419 MB), and L → M → K about 0.15 GB instead of 0.8 GB.

@Schiggy measured on an RTX 5080 (Windows 11, driver 621.24) at your settings: 3440x1440, DLAA, Kokiri Forest, 4 runs. | Model | GPU a frame | fps | power | GPU utilization | |---|---|---|---|---| | K | 7.4–7.6 ms | 80–84 | 140–145 W | 74% | | M | 9.7–9.9 ms | 70–72 | 186–188 W | 79% | | L | 10.6–10.7 ms | 66–67 | 198–202 W | 81% | So the same shape as yours: L costs about 20% of the fps and about 40% more power than K at DLAA, on a 5080 too. That part is the model's own work at DLAA's input size, and it is the same with the fix below. On utilization: you're right that it doesn't show the full picture. Here it only went from 74% to 81%, yet the frame got about 3 ms longer, the same as DLSS's extra GPU time. In Longshot a frame's GPU work isn't overlapped with the next frame's preparation: the frame is presented only after DLSS has run, and the next one isn't started until then. So any extra GPU time lands on the frame time even though the GPU sits idle at other moments in the frame. Utilization counts those idle moments; fps doesn't. With a game that overlaps frames, the headroom you see would absorb it. That overlap is being worked on separately (frame overlap), not in this issue. The memory jump at the switch was ours, and it's fixed in https://git.magicbean.dev/MagicBeanDev/Longshot/pulls/39: stepping K → M → L no longer makes M's feature on the way. VRAM rises about 0.64 GB instead of about 1.2 GB for the first seconds after the switch (DLSS's own memory 813 MB instead of 1419 MB), and L → M → K about 0.15 GB instead of 0.8 GB.

HumanBean referenced this issue from a commit 2026-11-13 01:35:59 +00:00

fix(dlss): the DLSS model row's preset is taken once it holds it, not a feature for each one stepped past (#39)

Sign in to join this conversation.

No Branch/Tag specified

Branches Tags

main

nightly

v2.1.0-rc.3

v2.1.0-rc.2

v2.1.0

v2.1.0-rc.1

v2.0.14

v2.0.13

v2.0.12

v2.0.11

v2.0.10

v2.0.9

v2.0.8

v2.0.8-rc.5

v2.0.8-rc.4

v2.0.8-rc.3

v2.0.8-rc.2

v2.0.8-rc.1

v2.0.7-rc.1

v2.0.7

v2.0.6

v2.0.5

v2.0.4

v2.0.3

linux-parts

v2.0.2

v2.0.1

v2.0.0

honest-eta

fast-updates

low-vram

faster-fps

ultrawide

faster-builds

compile-retry

new-installer

builder-parts

device-lost-fix

no-visual-studio

foliage-linux-updater

movie-sync

less-stutter

builder-errors

small-vram

amd-shared-memory

brighter

pointer-lighting-2

text-hotfix

pointer-lighting

lighting-fix

black-screen-fix

important

amd

fix

nup-fix

nup-support

Release

Labels

Clear labels   

accessibility

Barrier affecting people with disabilities

  

autorelease: pending

  

autorelease: tagged

  

bug

Something isn't working

  

dependencies

Pull requests that update a dependency file

  

documentation

Improvements or additions to documentation

  

duplicate

This issue or pull request already exists

  

enhancement

New feature or request

  

github_actions

Pull requests that update GitHub Actions code

  

good first issue

Good for newcomers

  

help wanted

Extra attention is needed

  

invalid

This doesn't seem right

  

needs measurements

Waiting for before/after measurements

  

question

Further information is requested

  

testing wanted

Wanted: a test on other hardware (posts a test request in #test-requests)

  

wontfix

This will not be worked on

No labels

accessibility

autorelease: pending

autorelease: tagged

bug

dependencies

documentation

duplicate

enhancement

github_actions

good first issue

help wanted

invalid

needs measurements

question

testing wanted

wontfix

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

2 participants

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

MagicBeanDev/Longshot#17

No description provided.