Skip to content
VSPT

Performance guide

VSPT is limited by the graphics card. A frame is the trace (the rays that find the light), then a denoiser and an upscaler that turn those few rays into a clean picture, then the game’s own work. Almost every setting that changes the frame rate changes the trace, so the trace is what is measured here, in milliseconds. A frame at 60 a second has 16.7 ms in all; at 30, 33 ms.

The trace is done at a lower size than your screen when an upscaler is on. At 1440p with Performance upscaling it is 1280x720, which is what the figures below use.

The trace kit is a benchmark that runs VSPT’s renderer, without the game, on a scene captured in the game. It runs the same scene several times, each time with one part taken out, so the difference is what that part costs. This is one scene on four cards, with version 3 of the kit (V3):

PartNVIDIA GeForce RTX 5080, msAMD Radeon RX 9070 XT (a tester's card), msAMD Radeon RX 7900 XTX (a tester's card), msAMD Radeon RX 9060 XT (a tester's card), msRX 9070 XT against RTX 5080RX 7900 XTX against RTX 5080RX 9060 XT against RTX 5080
Whole trace, default settings5.9923.315.137.53.9 times2.5 times6.3 times
Whole trace, light reuse off4.0510.58.1318.82.6 times2.0 times4.6 times
Whole trace, the game case (worst-case fog)12.449.535.677.64.0 times2.9 times6.3 times
Whole trace, as advised for AMD (light reuse and fog off)6.1013.710.624.22.2 times1.7 times4.0 times
Light reuse1.9412.86.9218.76.6 times3.6 times9.7 times
Foliage (leaves, grass, other cut-outs)2.535.775.019.652.3 times2.0 times3.8 times
Fog (a worst case)4.4920.714.631.74.6 times3.3 times7.1 times
Classic clouds2.973.884.305.131.3 times1.4 times1.7 times
NRD's split signal0.260.311.050.05

Measured 2026-11-09. Trace kit V3 (rt_bench, the renderer without the game). The same tester's dusk scene as the other kit sets: grassland with trees, pools and a stone spire; 1280x720 trace (what 1440p traces at Performance upscaling). Settings: Three bounces at three quarters of the pixels, light reuse through RTXDI, radiance cache, 4 light candidates; the game case adds NRD's split signal, the thickest fog (density 0.05, traced exactly) and classic clouds, and the costs of those three are taken from it; the AMD row is the game case with light reuse and fog off. 240 frames, median of 3 rounds. No upscaler or denoiser in the timing. Build: RTX 5080: rt_bench from b6bdb183d, V3 as first built, run on 8 November. RX 9070 XT, RX 7900 XTX and RX 9060 XT: V3 rebuilt with the fix for the AMD shader compile crash (640fc43da), as sent to testers, the same build on all three. NVIDIA GeForce RTX 5080: driver 32.0.16.2101 (Windows driver version string). AMD Radeon RX 9070 XT (a tester's card): driver AMD Software 26.10.1. AMD Radeon RX 7900 XTX (a tester's card): driver AMD Software 26.10.1. AMD Radeon RX 9060 XT (a tester's card): driver AMD Software 26.10.1. A cost is the whole trace's time with the part, less the time without it. NRD's split signal, 27 ms on the RX 9070 XT in V2, now costs every card little; its row has no comparison columns, since at a millisecond or less they would say little. Fog and clouds also cost the Radeons far less than in V2, with no change of their own (see the V2 against V3 table). Light reuse is now the largest cost that only the Radeons pay: in the main path tracing pass its code costs the RX 9070 XT and RX 7900 XTX 4 to 10 ms, against about 0.1 on the 5080, whose light reuse is RTXDI's own passes. The fog cost is a worst case: the kit traces the thickest fog exactly, at a moonlit dusk that this fog all but blocks, where the game uses a faster method for fog and usually lighter fog. The RX 9060 XT (RDNA4, 32 compute units to the RX 9070 XT's 64) takes about 1.6 times the 9070 XT's time.

What it says of an RTX card still holds: fog is the largest single cost of these, and light reuse and foliage are about 2 to 2.5 ms each.

The Radeon columns were taken through the ray tracing pipeline, which AMD cards no longer use (below), so they overstate what each part costs an AMD card now, and they do not say how the cards compare now. On the pipeline the RX 7900 XTX was faster than the RX 9070 XT in every case: the pipeline’s cost on AMD cards is memory traffic, and the 7900 XTX has the most memory bandwidth of the three. How the cards compare as they trace now is in the next section.

The fog costs in this table are a worst case, and overstate what fog costs in the game. The kit traced the game’s thickest fog, exactly, at a moonlit dusk that this fog all but blocks. The game uses a faster method for fog, and its fog is usually lighter. Version 4 of the kit and later trace fog as the game does.

A tester’s analysis of the pipeline that AMD’s Windows driver builds for VSPT found that the driver compiles ray tracing pipelines with continuations. Whatever a path keeps from one ray to the next is saved to memory and loaded back at every ray, so each feature’s extra state costs an AMD card far more than an NVIDIA one.

So AMD cards on Windows trace inline: the rays are ray queries in a compute shader, which keeps the whole path in the card’s registers, and nothing is saved and loaded at each ray. NVIDIA cards stay on the ray tracing pipeline, which is faster on them. In version 5 of the kit (V5), on a tester’s RX 9070 XT, the main path tracing pass was 2.8 to 4.2 times as fast inline in a cave and in a torch-lit scene, and 1.2 to 1.5 times on the grassland of the scene above (2 times with Clear water on), where the grass and leaves gain least. The game case’s whole trace went from 14.62 ms to 5.93 in the cave, and from 22.38 to 16.74 on the grassland. Inline, Clear water costs that card 0.14 ms of the main pass where it cost 6.3, and light reuse adds nothing to the main pass, which leaves its own passes: about 1.5 ms in the cave and 3 on the grassland. These are from the trace kit, not the game.

Version 7 of the kit (V7) puts the cards side by side as VSPT 4.0 traces on each: the RTX 5080 on its pipeline, the Radeons inline.

PartNVIDIA GeForce RTX 5080, msAMD Radeon RX 9070 XT (a tester's card), msAMD Radeon RX 9060 XT (a tester's card), msRX 9070 XT against RTX 5080RX 9060 XT against RTX 5080
Default settings, in the cave3.724.638.301.2 times2.2 times
Default settings, on the grassland5.8010.119.51.7 times3.4 times
The game case, in the cave4.255.3110.01.2 times2.4 times
The game case, on the grassland8.4415.129.21.8 times3.5 times

Measured 2026-11-22. Trace kit V7 (rt_bench, the renderer without the game), the same packed kit on every card. The RX 9070 XT ran it on 12 November, the RTX 5080 and the RX 9060 XT on 22 November. The kit's two scenes: a torch-lit cave, and a tester's open grassland at dusk; 1280x720 trace (what 1440p traces at Performance upscaling). Settings: Each card on its own path, as in the game: the RTX 5080 on the ray tracing pipeline, the Radeons inline (AMD's Windows driver). The whole trace; 240 frames, median of 3 rounds, every run kept. Build: The V7 kit, packed with the code of the test build 0.4.0-dev.8 (6b4c3e61b). NVIDIA GeForce RTX 5080: driver 621.01 (Windows driver version string 32.0.16.2101). AMD Radeon RX 9070 XT (a tester's card): driver AMD Software 26.10.1 (Windows driver version 32.0.32015.7029). AMD Radeon RX 9060 XT (a tester's card): driver AMD Software 26.10.1 (Windows driver version 32.0.32015.7029). Over the two scenes’ game cases the RX 9070 XT takes 1.5 times as long as the RTX 5080 and the RX 9060 XT 2.9 times. The RX 9060 XT takes 1.8 to 2.1 times as long as the RX 9070 XT in every case of the kit. Traced inline, as the Radeons trace, the RTX 5080 is no faster than the RX 9070 XT on the grassland (15.42 ms against 15.06): on its own pipeline it reorders its rays, which a ray query cannot. No RX 7900 XTX has run this kit.

Over the two scenes the RX 9070 XT takes about 1.5 times as long as the RTX 5080, and the RX 9060 XT about 2.9 times. The RX 9060 XT takes about twice as long as the RX 9070 XT in every case, more than the 1.6 times it took on the pipeline. No RX 7900 XTX has run this version yet.

The defaults are made for these cards: DLSS Ray Reconstruction with Performance upscaling, and the High preset. When the frame rate is low, in order of what helps:

  1. Quality preset. Medium and Low turn off fog and caustics, put fewer plants in bounced light and trace fewer bounces.
  2. Upscaling quality. Ultra Performance traces the fewest pixels.
  3. Fog. The most expensive single part in the first table above: 4.5 ms on an RTX 5080, though that is the kit’s worst case, and fog usually costs less in the game. Over open water it can still be the largest part left: in one view of the sea at 4K with Performance upscaling, the game’s everyday fog was 3.8 ms of a 20.5 ms trace.
  4. View distance. Everything loaded is traced, and on a card with 8 to 12 GB it also fills the video memory. On a 16 GB card at 4K, about 800 blocks is the ceiling in a built-up area: 13.2 GB was in use there.
  5. Clouds. Classic clouds cost the least, then Volumetric at Low, Medium and High:
PartNVIDIA GeForce RTX 5080, ms
Clouds off1.83
Classic clouds2.25
Volumetric, Low2.63
Volumetric, Medium2.85
Volumetric, High3.75

Measured 2026-10-27. rt_bench, trace stage. terrain3, over the sea (rise 40, look down -3), made-up fair sky; 3840x2160 from 1280x720, DLSS Ray Reconstruction Ultra Performance. Settings: Four rounds of 240 frames, spread within 0.1 ms. Build: 0.2.8 development build. NVIDIA GeForce RTX 5080: driver not recorded. These are the trace stage's totals, including the sea and sky. The cost of the clouds is the difference from the first row.

Light reuse costs 1.9 ms on an RTX 5080 and is the part most worth keeping: it is what makes torch-lit places clean.

A feature you have turned off costs little, because the renderer builds its shaders for the features in use. The first time a new combination is used, the driver compiles a shader for it, which can take up to about 30 seconds, and the frames until then run on a stand-in that is slightly slower.

Clear water costs one more ray where you look into water, and nothing where no water is in view: about 3% of the trace at a pond and 7 to 8% over the sea. These figures are from 0.3.0, where it was off by default:

PartNVIDIA GeForce RTX 5080, msNote
Pond, from the eyes0.303.9 % of the trace
Open sea0.698 % of the trace

Measured 2026-11-08. rt_bench, trace stage, the setting on against off. terrain3's pond from the eyes, and the open sea; 2560x1440, DLSS Ray Reconstruction. Settings: Clear water on against off. Build: development build. NVIDIA GeForce RTX 5080: driver not recorded. One more ray where the camera looks into water, so it costs nothing elsewhere.

In VSPT 4.0 Clear water is on by default, and water is smooth near you and stepped far away (the settings Near and Far water wave resolution). At 4K with Performance upscaling:

PartNVIDIA GeForce RTX 5080, msNote
Clear water, pond from the eyes0.55
Clear water, open sea to the horizon1.10
Clear water, open sea looked down on1.58
Light in the water, pond from the eyes0.722.07 in 0.3.0
Light in the water, open sea to the horizon1.423.65 in 0.3.0
Light in the water, open sea looked down on2.305.75 in 0.3.0
Near and Far water wave resolutionunder 0.2nothing that can be measured; from a second run, of five rounds

Measured 2026-11-12. rt_bench, trace stage, one build switched at run time, the median of three rounds. a pond from the eyes, and the open sea; 3840x2160, DLSS Ray Reconstruction, Performance (a 1920x1080 trace). Settings: Each part on against off. Build: 0.4.0-dev.7 test build. NVIDIA GeForce RTX 5080: driver 621.01 (Windows driver version string 32.0.16.2101). Clear water is one more ray where the camera looks into water, so it costs nothing elsewhere. The trace here is at 1920x1080, 2.25 times the pixels of the guide's other tables.

The trace is 0.7 to 1.3 ms longer where water fills the view with Clear water on, its default, than with it off, so Clear water is the first thing to turn off if the frame rate drops over open water. The two wave resolution settings cost nothing that can be measured: leave them. Water itself is cheaper to look at than in 0.3.0: at the same settings the trace is 15% shorter looking down at the sea, 13% towards the horizon and 8% at a pond, and with Clear water on it is still shorter than 0.3.0’s was with it off.

One water setting is new in VSPT 4.0 and off by default: Refracted underwater sun rays. While you are under water it bends the sun’s rays and the shadows on the bottom the way the surface bends the light. It is off for what it costs: about 3 to 11% more trace while you are under water (1.3 to 4.6 ms at 4K with Performance upscaling on an RTX 5080), and nothing while you are out of it.

At 0.3.0’s settings VSPT 4.0 runs the same. Flying over a generated village, on an RTX 5080 at 4K with Performance upscaling:

SettingNVIDIA GeForce RTX 5080, fpsNote
0.3.0, standing59.2
Release candidate at 0.3.0's water settings, standing59.0
Release candidate at its defaults, standing58.4
0.3.0, flying57.1two rounds
Release candidate at 0.3.0's water settings, flying57.4two rounds, one of them the slower first round
Release candidate at its defaults, flying56.6two rounds, one of them the slower first round
0.3.0, settled65.0
Release candidate at 0.3.0's water settings, settled64.6
Release candidate at its defaults, settled64.2

Measured 2026-11-15. The game's frame log, flying over a generated village; traced frames a second, the median of up to three rounds flown in turn, the versions minutes apart in each. A phase in which shaders were still compiling is left out. A generated village, standing, flying over it, and settled after the flight; 3840x2160, DLSS Ray Reconstruction, Performance (a 1920x1080 trace). Settings: The High preset's defaults, view distance 832, Classic clouds. "At 0.3.0's water settings" is VSPT 4.0 with Clear water off and water stepped everywhere, as 0.3.0 has it; "at its defaults" has Clear water on and water smooth near you. Build: the code of the release candidate 0.4.0-rc.1, and VSPT 0.3.0. NVIDIA GeForce RTX 5080: driver 621.01 (Windows driver version string 32.0.16.2101). At the same settings the two versions came within 0.6 of a frame a second of each other standing in every round, and within 1.5 flying and settled, neither ahead throughout: they read as alike. The first round was flown while something else used the graphics card and reads about 3 frames a second lower for both. The whole evening reads about 2 frames a second below the flights of 12 November for both versions, so compare the rows of this table with each other and not with that one. With a Ryzen 7 5800X3D.

Each version was flown three times in turn, minutes apart. At the same settings they stayed within about a frame a second of each other, neither ahead throughout, so read them as alike. VSPT 4.0’s own defaults, with Clear water on and smooth water near you, cost about half a frame a second here, where little water is in view. The figures are from its release candidate’s code.

Two hitches are gone in VSPT 4.0: one twice a second with volumetric clouds and a long Cloud distance, and a small one every five seconds. With the first, the slowest one in a hundred of the game’s own frames (its work on the processor, before the picture is traced) went from 11 to 13 ms down to 7 to 8. In the flights above, with Classic clouds, that slowest one in a hundred was a little shorter than 0.3.0’s as well: 10.8 ms against 11.6 standing, and 8.3 against 9.8 flying, on one evening.

The first start after an install or update

Section titled “The first start after an install or update”

After that first world, VSPT remembers which shaders your last session used and compiles those first when it has to compile again, where 0.3.0 compiled all of them after every update. A start with nothing to compile shows the traced picture within ten seconds of the world loading.

VSPT 4.0 supports AMD Radeon RX 7000 and newer, on Windows and on Linux. Expect lower performance than on an NVIDIA card. There is no DLSS: the picture is cleaned up by NVIDIA’s NRD denoiser (REBLUR), which is good but not as good as DLSS Ray Reconstruction, and upscaled by AMD FSR 4.1 on Windows. On the first launch on a card, a dialog recommends these settings. FSR’s own surfaces take about 224 MB of video memory at 4K with Performance upscaling. The RX 6000 series has known problems and isn’t supported yet (FAQ).

On Linux, where AMD’s software support is poor, VSPT uses FSR 3.1 built from AMD’s source and has no frame generation. Testers have run it on an RX 7900 XTX and an RX 9070 XT with Mesa’s RADV driver, where it traces on the ray tracing pipeline, not inline: on a tester’s RX 7900 XTX the pipeline ran at 52.8 traced frames a second where inline ran at 11.8 (the figures). Everything else on this page was measured on Windows.

What to set, the most important first:

  1. View distance 256 or less. A 16 GB AMD card filled its video memory at the game’s 1536, which made it crawl at 1 to 2 frames a second. At 128 to 256 it had 9 to 11 GB free. AMD’s acceleration structures are much larger than NVIDIA’s, and the renderer’s memory budget does not yet allow for it. On Linux a full card refuses memory rather than slowing down; the view is then made again one upscaling mode down, and chat says so.
  2. Performance upscaling, with NRD REBLUR and FSR 4.1 where the settings offer it.
  3. Light reuse: On, unless the frame rate needs it. Inline, its own passes cost a tester’s RX 9070 XT about 1.5 to 3 ms in the trace kit (above). Off, lamp-lit places are noisier, since each pixel then samples the lamps alone; daylight barely changes.
  4. Clear water and Fog: On, and tell us how it runs. Inline, Clear water cost that card almost nothing in the trace kit, and the testers’ cards below ran in the game with both on.

Testers’ Radeon cards have sent figures from the game while tracing inline, with light reuse, fog and Clear water on: a tester’s RX 9070 XT at 1440p with Performance upscaling ran at 88 traced frames a second standing still and at 48 moving among plants swaying in the wind, and an RX 9060 XT at about 1080p with Balanced upscaling at 41, with frame generation doubling what it showed. Each was in its tester’s own world, so they say what those cards ran at there, not how one compares with another. They are on the measurements page, with the trace kit’s comparison of inline against the ray tracing pipeline. The figures from before AMD cards traced inline, there too, overstate what Light reuse and fog cost an AMD card now.

NRD’s picture is softer than DLSS Ray Reconstruction’s, and it can leave a trail behind light that moves. Its settings are the player’s, on a page of their own, Settings → Path Tracing → Upscaling: 22 rows that take effect on the next frame (which to try first).

Most of the rows change how NRD blends, not how much work it does. One adds work: Detail resolve, which puts the relief of normal maps back into lamp light and bounced light. What NRD and Detail resolve cost was timed with version 7 of the trace kit, on an RTX 5080 and on a tester’s RX 9070 XT:

PartNVIDIA GeForce RTX 5080, msAMD Radeon RX 9070 XT (a tester's card), msNote
NRD, in the cave1.110.94its own pass; the whole frame is 1.12 ms longer on the RTX 5080 and 0.89 on the RX 9070 XT
NRD, on the grassland1.150.94its own pass; the whole frame is 1.13 ms longer on the RTX 5080 and 0.94 on the RX 9070 XT
Detail resolve on top, in the cave0.350.28the whole frame; 0.33 of it in NRD's pass on the RTX 5080, 0.22 on the RX 9070 XT
Detail resolve on top, on the grassland0.330.27the whole frame on the RX 9070 XT, 0.21 of it in NRD's pass. On the RTX 5080, NRD's pass alone: its trace there varies from round to round by more than the setting adds

Measured 2026-11-12. trace kit V7's NRD cases: each case against the case before it, the median of three rounds of 240 frames. The RTX 5080 ran them from the development tree, the RX 9070 XT from the packed kit. a torch-lit cave, and open grassland; 1280x720 trace (1440p with Performance upscaling), no upscaler after NRD. Settings: The game's defaults, traced inline as AMD cards trace; NRD REBLUR. Build: the code of the test build 0.4.0-dev.8. NVIDIA GeForce RTX 5080: driver 621.01 (Windows driver version string 32.0.16.2101). AMD Radeon RX 9070 XT (a tester's card): driver AMD Software 26.10.1 (Windows driver version 32.0.32015.7029). NRD's own textures take 61 MB at this size, and 96 MB with Detail resolve on, on both cards. In the game an RTX card traces through the ray tracing pipeline; these cases trace inline, as AMD cards do, so they say what NRD adds on each card and not how fast one card is beside the other. The RX 9070 XT's cave figures are read from the kit's logs: its report leaves out a run whose slowest frames take over 1.5 times the median, as they did in every round of the cave on that card, with and without NRD (about twice). Those runs' medians agree to 0.03 ms across the three rounds.

On the RX 9070 XT, NRD adds about 0.9 ms to a frame and Detail resolve about 0.3 ms more, a little less than on the RTX 5080. Detail resolve is off by default; its cost is small, so turn it on if you like how it looks. One AMD card has run these cases so far, at this one trace size. Cavity shadow, a row that darkens the hollows in block textures and needs Detail resolve on, adds next to nothing more.

On that RTX 5080 in the game, at 4K with Performance upscaling, NRD with FSR 3.1 ran a little ahead of Ray Reconstruction, in the village above:

SettingNVIDIA GeForce RTX 5080, fpsNote
Ray Reconstruction, standing59.5
NRD with FSR 3.1, standing60.6
Ray Reconstruction, flying59.0from the next round: shaders were still compiling in this one
NRD with FSR 3.1, flying61.0
Ray Reconstruction, settled64.6
NRD with FSR 3.1, settled67.0

Measured 2026-11-15. The game's frame log, flying over a generated village; traced frames a second, one round, the two flown minutes apart. A generated village, standing, flying over it, and settled after the flight; 3840x2160, Performance upscaling (a 1920x1080 trace). Settings: The High preset's defaults with Clear water off and water stepped everywhere, view distance 832, Classic clouds; DLSS Ray Reconstruction, or NRD REBLUR at its defaults with FSR 3.1. Build: the code of the release candidate 0.4.0-rc.1. NVIDIA GeForce RTX 5080: driver 621.01 (Windows driver version string 32.0.16.2101). NRD and FSR 3.1 together took about 2.4 ms of a frame here and Ray Reconstruction about 3.4, and the trace was about 0.6 ms longer with NRD. This is NRD on an NVIDIA card, whose rays go through the ray tracing pipeline. An AMD card traces inline and uses FSR 4.1 where it is offered, so this says nothing of a Radeon card's frame rate. With a Ryzen 7 5800X3D.

NRD and FSR 3.1 together took about 2.4 ms of a frame there, a millisecond less than Ray Reconstruction, and the trace was about 0.6 ms longer with NRD. That says nothing of a Radeon card, which traces inline and uses FSR 4.1 where it is offered.

Still to come on AMD:

  • Faster tracing of open land with grass and leaves, where the inline trace gains least.
  • A video memory budget that fits AMD’s structures.
  • The cause of the RX 6000 series’ problems.
  • Figures from the game on more Radeon cards: so far an RX 9070 XT and an RX 9060 XT on Windows and an RX 7900 XTX on Linux have sent them (measurements).
Setting RTX card Radeon card
Denoiser DLSS Ray Reconstruction NRD REBLUR
Upscaler DLSS FSR 4.1 on Windows, FSR 3.1 on Linux
Upscaling quality Performance, or Ultra Performance on a smaller card Performance
Light reuse Leave on On; off if the frame rate needs it (why)
Fog On at High, Off on a smaller card On
Clear water On by default: the first thing to turn off if the frame rate drops over open water, with no effect elsewhere (what it costs) On, and tell us how it runs
Near and Far water wave resolution Leave: they cost nothing Leave
Refracted underwater sun rays Off by default. On, 3 to 11% more trace while you are under water, and nothing out of it (what it costs) Off by default; not measured on a Radeon card
Detail resolve Not used with Ray Reconstruction Off by default; try it for sharper materials. About 0.3 ms a frame on an RX 9070 XT (what it costs)
Frame generation As your card allows (which one) AMD FSR 3 on Windows; not offered on Linux
View distance As far as the video memory allows: about 800 on a 16 GB card at 4K in a built-up area 256 or less

Every figure is in one data file on this site, with its date, card, driver, test, scene and settings, and the measurements page shows all of it. When new numbers arrive they are added there, and this guide is built from them. Two kinds of test are used:

  • The trace kit, a benchmark VSPT’s developers send to testers privately (it contains the game’s art). It gives the same measurement on every card, so cards and parts can be compared. It is not the game: it has no entities, no particles and no upscaler, and no denoiser but in version 7’s two NRD cases.
  • The game’s own frame log, from play: this is what a player gets, and varies with the world and where you stand.

If you have an AMD card and would like to test, on Windows or on Linux with Mesa’s RADV driver, say so in the VSPT channel on the Vintage Story Discord.