DLSS model row: Transformer 2 (M, L) costs 3-4 ms of GPU a frame on an RTX 3060 Ti (56 / 52 fps against 62 with K); the hints say "a little more GPU time" #2836

Open

opened 2026-11-12 16:54:31 +00:00 by github-sync · 0 comments

github-sync commented

2026-11-12 16:54:31 +00:00

Owner

@klikso opened on GitHub · 2026-11-12 15:58 UTC · original

What happened

On an RTX 3060 Ti the DLSS model row's two "Transformer 2" choices cost 3-4 ms of GPU time a frame over the default, which is 6-10 fps at 60. The row's hints call M "sharper, at a little more GPU time" and say nothing of L's cost, so a player who picks the newest model (as the owner of this PC had) loses a sixth of the frame rate without a way to tell why: the FPS counter's GPU time just reads higher.

Measured

Windows 11, RTX 3060 Ti 8 GB (driver 32.0.16.2114), Ryzen 5 5600X, output 2560x1440, the game rendering 1152x648 and DLSS upscaling to 1536x864 (mode 2 in the log: the handheld profile). main ba005ca4 with #2711, #2731 (at bce3d17e, where it is still on by default), #2685, #2709, #2825 and #2826 merged locally; the same build for every arm, the model set with RX_VK_DLSS_PRESET. Hyrule Field child, 30 s of the camera turning after a settle, RX_VK_FRAMEGEN=0, HOS_FPS_MAX=240. Arms alternated, two runs each (K and L four).

DLSS model row	preset	fps mean	GPU a frame
Transformer (the default)	K	63.9 / 61.5 / 62.2 / 61.6	8.1 ms
NVIDIA default	0	62.1 / 62.2	8.2 ms
Transformer 2 (M)	M	55.9 / 56.1	11.1 ms
Transformer 2 (L)	L	52.2 / 52.3 / 52.5 / 53.0	12.2 ms
(CNN, not in the menu)	E	61.3 / 62.8	7.7 ms
the game's own upscale (`RX_VK_DLSS=0`)	-	62.8 / 61.8	7.3 ms

So on this card K costs ~0.8 ms over the game's own upscale, M ~3.8 ms and L ~4.9 ms, at a 1536x864 output. The frame here is the GPU's time plus the replay's, so each of those ms is frame time.

Suggestion

Not a code change I can make well (the hints are translated per language in menu_i18n.h):

say the cost in the hints of M and L, something like "... sharper, at 3-5 ms more GPU time a frame on an RTX 30 card; less on newer ones", or
have the FPS counter's Breakdown name DLSS's share, so the cost is visible where the player looks.

Numbers from RTX 40/50 cards would say whether the cost is this high there; I only have this one.

> [@klikso](https://github.com/klikso) opened on GitHub · 2026-11-12 15:58 UTC · [original](https://github.com/NotRexed/Hookshot/issues/2836) ### What happened On an RTX 3060 Ti the DLSS model row's two "Transformer 2" choices cost 3-4 ms of GPU time a frame over the default, which is 6-10 fps at 60. The row's hints call M "sharper, at a little more GPU time" and say nothing of L's cost, so a player who picks the newest model (as the owner of this PC had) loses a sixth of the frame rate without a way to tell why: the FPS counter's GPU time just reads higher. ### Measured Windows 11, RTX 3060 Ti 8 GB (driver 32.0.16.2114), Ryzen 5 5600X, output 2560x1440, the game rendering 1152x648 and DLSS upscaling to 1536x864 (`mode 2` in the log: the handheld profile). `main` ba005ca4 with #2711, #2731 (at bce3d17e, where it is still on by default), #2685, #2709, #2825 and #2826 merged locally; the same build for every arm, the model set with `RX_VK_DLSS_PRESET`. Hyrule Field child, 30 s of the camera turning after a settle, `RX_VK_FRAMEGEN=0`, `HOS_FPS_MAX=240`. Arms alternated, two runs each (K and L four). | DLSS model row | preset | fps mean | GPU a frame | |---|---|---|---| | Transformer (the default) | K | 63.9 / 61.5 / 62.2 / 61.6 | 8.1 ms | | NVIDIA default | 0 | 62.1 / 62.2 | 8.2 ms | | Transformer 2 (M) | M | 55.9 / 56.1 | 11.1 ms | | Transformer 2 (L) | L | 52.2 / 52.3 / 52.5 / 53.0 | 12.2 ms | | (CNN, not in the menu) | E | 61.3 / 62.8 | 7.7 ms | | the game's own upscale (`RX_VK_DLSS=0`) | - | 62.8 / 61.8 | 7.3 ms | So on this card K costs ~0.8 ms over the game's own upscale, M ~3.8 ms and L ~4.9 ms, at a 1536x864 output. The frame here is the GPU's time plus the replay's, so each of those ms is frame time. ### Suggestion Not a code change I can make well (the hints are translated per language in `menu_i18n.h`): - say the cost in the hints of M and L, something like "... sharper, at 3-5 ms more GPU time a frame on an RTX 30 card; less on newer ones", or - have the FPS counter's Breakdown name DLSS's share, so the cost is visible where the player looks. Numbers from RTX 40/50 cards would say whether the cost is this high there; I only have this one.

github-sync referenced this issue

2026-11-12 16:55:43 +00:00

perf(vk): import game memory in 4 MB chunks, not 16, so VRAM mirrors fit the game's pools #2685

github-sync referenced this issue

2026-11-12 16:55:51 +00:00

fix(vk): make the blocks at the start in the pool the textures are made in (#2690) #2711

github-sync referenced this issue

2026-11-12 16:55:53 +00:00

perf(vk): RX_VK_WB_MAX_KB writes large GPU-written bindings back only where the CPU reads or writes them (opt-in) #2731

github-sync referenced this issue

2026-11-12 16:56:27 +00:00

perf(vk): an import mostly outside every game pool is trimmed back to its live parts #2825