SCSKiller: The 347ms Worst Shader Stall Becomes 8ms Before You Play

Updated October 18, 2026: SCSKiller, a free open-source Windows tool that compiles a game's shaders into the GPU driver's cache before you play, has published its own measured results, and the headline row is Silent Hill Townfall on an NVIDIA RTX 5090: 24 shader stutters and a worst stall of 347 ms from a cold driver cache, against 0 stutters and an 8 ms worst stall after SCSKiller compiles the shaders ahead of time. The repository behind it says it was created September 29, 2026 and stood at 2,993 GitHub stars under the GPL-3.0 licence when we checked on October 18.

By Indie Kings | October 18, 2026

Shader stutter is not a hardware problem, it is a scheduling problem, and SCSKiller's measured table puts a number on that: on the AMD Ryzen AI Max+ 395, the 2025 integrated chip people call Strix Halo, Final Fantasy VII Rebirth waited on 162 shader compiles over about five minutes of play from a cold driver cache, with a worst stall of 291 ms, and after SCSKiller precompiled the shaders in its own process those readings dropped to 0 stutters and a 7 ms worst stall. That single row, spelled out in the project's own README, is the honest version of the pitch: move the compile work out of the frame and most of the stutter leaves with it.

The tool arrives in a week where this site has already covered Microsoft's Windows 11 responsiveness work and the Smart HDR Manager tray utility, making SCSKiller the third entry in our performance-tool coverage and the only one that attacks compilation stalls rather than memory or display settings. It is the work of a developer who goes by BlueHeisenberg, it is written in C#, and it is free and open source for Windows, per the official repository README. Everything measurable that follows comes from that README, which we fetched in full on October 18, 2026.

What shader compilation stutter actually is

The README explains the mechanism plainly: a game that meets a new effect has the GPU driver compile its shader on the spot, and the frame waits. That wait is shader compilation stutter, and its size depends on how long the driver needs to build every pipeline a scene calls for. On a driver cache that has never seen the game, the first encounters are slow; on a cache that has already seen them, they are not.

This is why the problem is a scheduling problem rather than a raw power problem. The RTX 5090 in the README's table is one of the most powerful consumer GPUs money can buy, and Silent Hill Townfall still logged 24 stutters of 20 ms or more on it in about five minutes of play, one of which held the game for 347 ms. No GPU class removes the stall; only moving the compile work out of frame count does.

For honesty, one boundary matters before any numbers appear. SCSKiller's author writes that the tool fixes shader compilation stutter only, not traversal or streaming stutter, and that a shader it could not find or record can still compile mid-game, per the README. Anyone comparing these figures to their own play should keep both limits in mind.

What SCSKiller does about it

Per the README, SCSKiller reads the shaders a game ships with, works out the pipelines the game will build, and compiles them into the GPU driver's cache ahead of time. The compile happens in SCSKiller's own process with the game closed, so the compilation cost lands outside play rather than inside it. The README also states that the tool compiles through DirectX, as the game would, and never touches the driver or driver settings.

Two engineering details stand out. First, the tool supports NVIDIA and AMD, covers DirectX 12 games, and covers DirectX 11 games on NVIDIA, per the README. Second, it recompiles after a driver update, because clearing the shader cache means clearing the work, which is exactly the kind of detail that decides whether a precompiler stays useful from one driver version to the next.

For games whose pipeline information their files do not expose, the project offers an optional recorder: play a few minutes with it on, then compile from what was recorded. Per the README, the recorder can also work alongside a mod that replaces shaders, and with the recorder on, a game's page graphs your last session's frame times and marks shader stutters apart from other hitches, with those recordings kept on your PC.

The README's own data: cold cache against compiled cache

The table below reproduces the README's results section in full, as published on October 18, 2026. The measurement definition is the one the README prints above its own table, and it travels with every row: about 5 minutes of play per game, sampled from a cold driver cache and again after compiling with SCSKiller, where a stutter is a shader compile the game waited 20 ms or more for. Numbers were captured with the project's own recorder, and the author notes that your numbers depend on the game, GPU and driver.

About 5 minutes of play per game, from a cold driver cache and after compiling with SCSKiller; a stutter is a shader compile the game waited 20 ms or more for. That definition is the piece's honesty spine, and it is worth stating twice, because the two Hogwarts Legacy rows in the table change how it reads. The tool recovered most of the loss there but did not reach zero: 87 stutters from 1,014 on the RTX 5090, with the worst stall down from 2,400 ms to 100 ms, and 206 from 27,743 on Strix Halo, per the README. Those leftover counts are exactly what the author's caveat predicts, where a shader the files or recorder could not capture still compiles in game.

The row to remember: 291 ms becomes 7 ms on Strix Halo

The angle line of this piece is the honest version of one row: Final Fantasy VII Rebirth on the Ryzen AI Max+ 395 went from 162 stutters of 20 ms or more, worst stall 291 ms, to 0 stutters and a 7 ms worst stall, per the README's own five-minute measurement. Strix Halo is a 2025 handheld-class integrated chip, so a many-hundreds-of-milliseconds compile stall landing in a frame is far more painful there than on a desktop GPU with more cooling and pipeline headroom.

The second-repeated row is Silent Hill Townfall on the RTX 5090, where the worst single interview with the driver's compiler engine sat at 347 ms from a cold cache and at 8 ms after compiling, per the README, again under the same 5-minute, 20 ms-threshold definition. The RTX 5090 numbers answer the skeptic who says stutter is a weak-hardware complaint: the strongest consumer GPU on the market still waited a third of a second on one shader compile. Strix Halo answers the reader with a handheld who thinks the problem belongs only to high-end rigs.

What these rows do not claim is universal coverage, and the README says so directly. The measurements are the project's own, taken with its recorder, across a small game list at the time of writing. Treat every number above as the author's demonstration on specific hardware rather than a guarantee for your library, which is exactly the posture the README itself takes.

The library view: every game, its cache, its status

The screenshot the project ships in its README shows the library screen, where games are grouped by store with shader and pipeline counts, cache size, compile time, a status such as "Needs rebuilding" or "Warmed", and Play buttons, per the README's own image description. That layout answers the operational question a precompiler lives or dies on: which games are covered right now, which lost their cache to a driver update, and which still need work before play.

The store grouping matters less for decoration than for coverage. Per the README, the tool finds games in Steam, Epic Games, EA app, GOG, Ubisoft Connect, Xbox (PC), Battle.net, PURPLE, HoYoPlay, Gaijin and the Riot Client, plus games you add yourself, and it puts games known to stutter on top with the reason. It also marks games whose engine prepares its shaders itself, a useful signal because those titles give a precompiler little to do.

Supports and limits: NVIDIA and AMD, DirectX 12, no Intel yet

The stated platform envelope is specific. Per the README, SCSKiller targets Windows 10 version 2004 or Windows 11 on 64-bit, runs against NVIDIA or AMD GPUs, and covers DirectX 12 games plus DirectX 11 games on NVIDIA. Intel GPUs are not supported yet, and the author states the reason honestly: he does not have one to test on, and invites sponsorship of an Intel Arc GPU to change that.

The boundary behaviour around anti-cheat is written with unusual care for a hobby tool, and it deserves a plain summary. Per the README, outside an allowed session the tool never launches anti-cheat games, injects into them or adds its recorder to them; it only reads their files, and it can miss a game, so use is at your own risk and the author is not responsible for bans or lost data. Optional recorder sessions are allowed per game and confirmed at every launch, and the README stresses that games with anti-cheat can treat the recorder as cheating and that no game is known to be safe with it enabled.

Where the data lives is also stated: settings and per-game plans sit in the user's local app data folder, and the compiled pipelines land in the driver's own shader cache rather than a private SCSKiller vault, per the README, which also affirms game files are never modified. The community shader hash database, where other players' recordings cover your games, is a Patreon supporter feature with anonymous, opt-in, hash-only sharing, per the README. The sub-questions around file changes have concrete answers in the README for anyone who wants the whole picture.

Sourcing, official links and the anti-cheat posture

The README carries an official-links warning worth repeating verbatim in substance: the only official downloads are this repository's Releases, and anything else claiming to be SCSKiller is not the project. The listed official channels are the GitHub repository, the scskiller.com website, a Patreon for supporters, a Discord and an X account, per the README. For a tool that explodes in stars within three weeks of being created on September 29, 2026, fake builds under the same name are exactly the scam pattern to expect.

The project also gives readers structured ways to contribute: bug reports and game requests go through issue templates on the repository, per the README. Versioning hygiene is visible too, with architecture, changelog, contributing and security documents present, and security issues routed privately rather than through public issues. The licence is GPL-3.0-or-later with an additional permission for Oodle, the Windows App SDK and graphics driver libraries, and forks are welcome under the condition that they do not present the main project's measurements as their own, per the README.

Our take

We think the strongest thing about SCSKiller is not the table, it is the framing printed above it. Publishing the measurement definition, the 5-minute cold-cache sample and the 20 ms stutter threshold, right next to the results is what made us willing to reproduce the table at all, because it lets a reader disagree with the numbers while still understanding what they mean. We would like to see that same cold-versus-warmed methodology run by someone other than the tool's author, since recorder-measured self-benchmarks are the one load-bearing plank of trust in the whole structure, and the game list is still small. If the methodology holds under independent measurement, this is a scheduling fix Windows gaming has needed for years, and the two near-zero rows on Strix Halo say integrated handhelds benefit from it more than anyone else.

FAQ

What counts as shader compilation stutter in SCSKiller's results?
Per the README, a stutter is a shader compile the game waited 20 ms or more for, with the sample drawn from about 5 minutes of play from a cold driver cache and again after compiling. Tools such as this others generally measure frame waits during play, so the threshold is what separates a compile stall from ordinary hitches.

Is SCSKiller free?
Yes. Per the README, the app is free and open source under GPL-3.0-or-later for Windows, and everything it does on your PC is free, with no account required. Patreon supporters get a community shader hash database and beta builds as extras.

Which GPUs and games does it work with?
NVIDIA and AMD GPUs, Windows 10 version 2004 or Windows 11, DirectX 12 games, and DirectX 11 games on NVIDIA, per the README. Intel GPUs are not supported yet because the author has none to test on. It finds games across eleven storefronts plus manually added titles.

Could running it get me banned in an anti-cheat game?
Per the README, outside confirmed sessions SCSKiller never launches anti-cheat games, injects into them or adds its recorder to them, it only reads their files, but its check can miss a game, so use is at your own risk. Recorder sessions in anti-cheat games are explicitly at your own risk, games can treat the recorder as cheating, and no such game is known to be safe with it enabled.

How much should I trust the numbers in the table?
They are the project's own measurements, taken with its recorder across about 5 minutes of play per game, and the author states your numbers depend on the game, GPU and driver. Treat them as a demonstration on specific hardware, not a universal guarantee, and note the tool fixes shader stutter only rather than traversal or streaming stutter.

Does it modify my game files or drivers?
Per the README, the compiled pipelines go in your driver's own cache and game files are never modified. Only the optional recorder adds two files to a game's folder and removes exactly those when turned off, and it is kept out of anti-cheat games outside confirmed sessions.

Bottom Line

SCSKiller is a free, GPL-3.0 open-source Windows tool that moves shader compilation out of your frames by building the pipelines into the driver's cache in its own process before a game opens, and its repository, created September 29, 2026 and pushed October 18, stood at 2,993 stars when we checked. Under its own published definition, about 5 minutes of play from a cold driver cache with stutter counted at 20 ms or more, its table drives stutters to zero on most rows, and its stubborn Hogwarts Legacy rows are honest about the remainder. The numbers are the author's, the game list is short, and Intel users are out of luck for now, but the methodology is published and the scheduling insight behind it is sound.

Sources: SCSKiller official repository README, fetched in full October 18, 2026 from github.com/BlueHeisenberg/SCSKiller; repository metadata from the GitHub API, October 18, 2026. All measured figures in the table are the project's own, under the README's stated sampling definition.

Image credit: BlueHeisenberg/GitHub