An experimental Switch 2 emulator tuned for speed, AMD Ryzen CPUs and NVIDIA RTX 40 / RTX 50 graphics cards. Trudy is a fork of Orchid and keeps all of Orchid's game fixes; it removes what was found slowing Switch 2 games down, and adds hardware-specific builds and defaults on top.
Note
Trudy is an independent fork. Please report problems with Trudy builds here, not to the Orchid, Eglantine or Eden developers.
Why Trudy | Changes | Test results | Settings | Download | Build it yourself | Credits
Ocarina of Time on a Ryzen 7 9800X3D + RTX 4090: 56 FPS in Trudy where Orchid's code runs the same scene at 35 FPS (+61%), with the graphics card 38% busy instead of 55%. Measured with release builds, same game, save and settings (see Test results).
Trudy started by profiling a real game instead of guessing. The biggest costs turned out not to be the CPU instruction set or thread placement, but data moving back and forth between the graphics card and the emulated console's memory. Fixing that helps every Switch 2 game, on any PC; the Ryzen and RTX 40/50 work comes on top.
| Orchid | Trudy | |
|---|---|---|
| GPU results copied back to the game | once the game's CPU has read any page of a buffer the GPU writes, the whole written range is copied back after every write (Ocarina of Time: 8.6 GB/s over PCIe, a third of the 4090's time) | only the pages the game's CPU actually reads (0.7 GB/s); the rest is fetched on demand if it is ever read |
| Ocarina of Time preset | reactive flushing on | reactive flushing off: the game's main core waits on GPU reads 8% of the time instead of 31% |
| Ocarina of Time, Kokiri Forest | the leaves flash with a large white glow (the whole-range copies write stale GPU data over memory the game just wrote) | drawn correctly |
| "Use the recommended settings?" question | No is the default answer | Yes is the default answer, so Enter accepts |
| Switch 2 games without a preset | GPU accuracy High (the GPU thread split stays off), synchronous shaders, reactive flushing on | what all of Orchid's presets share: accuracy Low (split on, ~1.46x draws a second), Balanced fences, asynchronous shaders, reactive flushing off. Only where you have not picked these yourself |
| Ryzen 9000 (Zen 5) | generic x86-64 + PGO builds | also -zen5 builds: -march=znver5 + the same PGO profile |
| Dual-CCD Ryzen (7950X3D, 9950X3D, 9900X3D, 9950X...) | threads go wherever Windows puts them | hot threads kept on one CCD, the 3D V-Cache CCD on X3D parts |
| RTX 40 / RTX 50 | treated as Ampere; the tensor-core path could never switch on (a float16 workaround disabled what it needs) | Ada and Blackwell recognised; the tensor-core path works when asked for (EDEN_NX2_TENSOR=fast) |
| GPU waits | driver sleep, wakes late | 50 µs spin before sleeping on 12+ thread CPUs |
| Profiling tools | GPU-thread CPU zones, sampling profiler | also host GPU time per shader, pipeline, render pass and command buffer (TRUDY_GPU_TIMING=1), a per-command GPU trace (TRUDY_GPU_TRACE=1) and fence/copy counters |
| Game compatibility | the same, with Orchid's newest fixes merged | |
| Settings, keys, firmware, saves | %APPDATA%\eden |
the same folder, so Trudy uses what Orchid already has set up |
git diff upstream/main shows every change against Orchid.
Switch 2 performance (all PCs)
buffer_cache_base.h,memory_tracker_base.h: buffer downloads at fences copy back only the "Preflushable" pages (pages the guest CPU has read after a GPU write) instead of the whole written range. A compute shader's write covers its whole storage binding (up to an 8 MiB window), so one page read by the game sent megabytes back at every fence: in Ocarina of Time ~4,800 copies of ~1.8 MB a second. Pages that were never read get a read trap, exactly as unread ranges already did. As a side effect the full-range copies no longer overwrite data the game's CPU wrote in the rest of the window.TRUDY_PAGE_DOWNLOADS=0restores the old behaviour.game_presets.h: the Ocarina of Time preset now turns reactive flushing off, like every other Switch 2 preset. Orchid's playtest found it faster on, but that was with the whole-range copies; with page copies, off is faster (17.9 ms against 22.9 ms a frame).game_settings.cpp: Switch 2 titles without a preset get the settings all of Orchid's presets share (GPU accuracy Low, Balanced fences, asynchronous shaders, reactive flushing off), only where neither the game's own setting nor your global one was changed.TRUDY_SWITCH2_DEFAULTS=0turns this off.main_window.cpp: the "Use the recommended settings?" question defaults to Yes, so Enter accepts it and Esc declines. In Orchid the default is No, and a game that declines runs on your own settings, by default the slow ones (GPU accuracy High, synchronous shaders).- Experimental, off by default:
TRUDY_FENCE_BATCH=<n>lets up to n fences share one GPU submission. It cut submissions by ~40% but did not change the frame time.
Measuring tools
vk_gpu_timer.cpp(new):TRUDY_GPU_TIMING=1logs, every 10 s, the host GPU time of each compute shader, graphics pipeline, render pass and whole command buffer (Vulkan timestamps), plus counters for fences, submissions and buffer copies.TRUDY_GPU_TRACE=1adds a timestamp after every recorded command and charges the GPU time to the function that recorded it; that is how the copy-back cost above was found.eden-cliwithEDEN_NX2_CRASH_DUMP=<dir>also writes the crash stack with function names toeden-crash-<pid>.txt(stderr is lost when it runs hidden).
AMD Ryzen
src/common/cache_affinity.cpp(new): finds the CPUs that share each L3 cache (Windows:GetLogicalProcessorInformationEx; Linux:/sys/devices/system/cpu/*/cache/index3). With more than one L3 domain it keeps the hot threads on the largest one: guest CPU cores, core timing, GPU front and back threads, GPU fencing, the Vulkan worker and presentation. With a single L3 (9800X3D, 7800X3D, 8-core parts, most Intel CPUs) nothing changes.TRUDY_CCD_AFFINITY=0turns it off.CMakeLists.txt: azen5build preset (-march=znver5 -mtune=znver5, clang 19+ / GCC 14+); MSVC presets mapped to/arch:AVX512and/arch:AVX2. CI buildswindows-x86_64-zen5andlinux-x86_64-zen5.- The dynarmic JIT already emits AVX-512 code when the CPU has it, which Zen 5 runs at full width;
EDEN_NX2_AVX512=0turns it off if a game misbehaves.
NVIDIA RTX 40 / RTX 50
vulkan_device.cpp/.h:Arch_AdaandArch_Blackwell, recognised by name (Vulkan has no architecture id and they expose the same extensions as Ampere). Turing and Ampere cards are never upgraded by mistake.- Eden turns the float16 shader feature off on every NVIDIA Ampere+ card because float16 math is
broken there, and the tensor-core (cooperative matrix) path needs that feature, so it could
never switch on. Trudy keeps the device feature on for RTX 40/50 while ordinary shaders still
avoid float16 math, so
EDEN_NX2_TENSOR=fastnow works. It is not the default: in Ocarina of Time it rejected one of the game's compute shaders and the frame time was the same. EDEN_NX2_GPU_WAIT_SPIN_USdefaults to 50 on CPUs with 12 or more threads.
Name and icon
- New orchid icon (
dist/trudy.svg): white-and-pink orchids on deep teal.trudy.exe, window titles, dialogs, release filestrudy-*, desktop iddev.trudy.trudy; the about box says what Trudy changes and credits Orchid and Eden. The update checker reads this repository's releases.
All measurements: AMD Ryzen 7 9800X3D, NVIDIA GeForce RTX 4090 (driver 621.01), 32 GB, Windows 11, 2026-11-05.
Hidden eden-cli runs on a copy of the save: load the autosave in Kokiri Forest, then 75 s of
gameplay with Link standing outside his house. Docked, 1x, GPU accuracy Low, asynchronous shaders,
BC3, conservative VRAM, FIFO Relaxed. The 30 FPS patch is off so the full rate shows. Frame
time is the mean of the emulator's own present statistics (600 frames per window).
| Build | Settings | Frame time | FPS | 4090 busy |
|---|---|---|---|---|
Trudy's first build (Orchid's emulation code), -zen5 |
Orchid's Ocarina of Time preset (reactive flushing on) | 28.8 ms | 34.7 | 55% |
Trudy 0.1.0, -zen5 |
the same | 22.9 ms | 43.7 | 37% |
Trudy 0.1.0, -zen5 |
Trudy's preset (reactive flushing off) | 17.9 ms | 56.0 | 38% |
What the profiling showed, before and after (debug-symbol build, GPU timers on):
| Before | After | |
|---|---|---|
| Buffer data copied back from the GPU | 8.6 GB/s in ~4,800 copies | 0.7 GB/s |
| Host GPU time in command buffers | 531 ms per second | 273 ms per second |
| ... of which the game's own draws and compute shaders | 110 ms | 166 ms (more frames) |
| Main game core waiting for GPU data | 35% of its time | 8% (reactive flushing off) |
Measured and not kept: fence submission batching (−40% submissions, same frame time; opt-in),
a cache for MaxContinuousRange (no measurable gain; removed), tensor cores by default (same frame
time, one shader rejected; opt-in).
Visual check: with Orchid's whole-range copies, the leaves in Kokiri Forest flash with a large white glow in 4 of 6 sampled frames while Link stands still; with Trudy's page copies they do not, which is how the game looks: the glow is a bug. The whole-range copies wrote stale GPU data over memory the game had just written.
| Test | Result |
|---|---|
GPU name detection, run against the real RefineNvidiaArchitecture code |
✅ RTX 4090 / 4060 Laptop / RTX 6000 Ada / L40S / L4 → Ada; RTX 5090 / 5070 Ti / RTX PRO 6000 Blackwell → Blackwell; RTX 3090 / RTX A4000 stay Ampere; Quadro RTX 4000 stays Turing |
| L3 domain detection on the 9800X3D | ✅ one L3 domain found and logged, thread placement left to Windows, as intended on single-CCD parts |
| CI builds | ✅ Windows MSVC, -pgo, -zen5; Linux x86-64, -pgo, -zen5; Linux ARM64 and -pgo; macOS ARM |
| Unit tests, Windows MSVC | ampere_kirby.cpp:163, ampere_f2i.cpp:821) are in Orchid's shader-lowering tests, which Trudy does not change |
Not tested yet
- Other games: only Ocarina of Time was available for testing. The copy-back fix is generic, but please report FPS for Orchid and Trudy side by side.
- Dual-CCD Ryzen pinning on real hardware (this PC has a single CCD), and Linux on hardware.
- ThinLTO for the
-zen5builds: CMake's IPO check fails with the CI toolchains, so they are built without it for now.
The table below is Orchid's, copied unchanged. It was measured on a different PC (Ryzen 7 5800X3D, Radeon RX 9070 XT) and shows game compatibility, which Trudy shares.
Note
Measured by hand in November 2026 on one PC: AMD Ryzen 7 5800X3D, AMD Radeon RX 9070 XT, 32 GB RAM, CachyOS (Linux 7.x, Mesa RADV from git), 1x resolution scaling. Other hardware will differ.
| Game | FPS | Notes |
|---|---|---|
| Hyrule Warriors: Age of Imprisonment | 60 | after shader compilation |
| Donkey Kong Bananza | 30, 15-25 in heavy areas | 30 in lighter areas |
| Mario Kart World | 30-40 | lots of off-track asset glitches |
| Mario Tennis Fever | 60 | runs flawlessly |
| Yoshi and the Mysterious Book | gameplay renders a black screen | |
| Super Mario Bros. Wonder | 60 | runs flawlessly |
| Star Fox | 30 | crashes very quickly |
| Splatoon Raiders | 30 | in gameplay |
| Pokémon Pokopia | 60 | in gameplay |
| Pokémon Legends: Z-A | 30 | in gameplay, the Switch 1 frame cap |
| Orbitals | does not launch | |
| Nintendo Switch Sports Resort | 60 | broken lighting, broken controller selection |
| Nintendo Switch 2 Welcome Tour | 30 | your own character is fully black |
| Metroid Prime 4: Beyond | 60 | until it crashes early on |
| Kirby Air Riders | crashes while loading a race | |
| Fire Emblem: Fortune's Weave | 30 | in the first fight, large graphics glitches top left |
| Fast Fusion | stuck loading into a race, then the emulator crashes | |
| Drag x Drive | no controller support, not playable |
Trudy reads the same environment variables and nx2_switches.txt file as Orchid. The ones Trudy
adds or changes:
| Variable | Trudy default | Meaning |
|---|---|---|
TRUDY_PAGE_DOWNLOADS |
on | 0 copies whole GPU-written ranges back at fences again, as Orchid does |
TRUDY_SWITCH2_DEFAULTS |
on | 0 leaves Switch 2 games without a preset on Eden's defaults |
TRUDY_CCD_AFFINITY |
on | 0 lets Windows/Linux place the hot threads anywhere |
EDEN_NX2_GPU_WAIT_SPIN_US |
50 on 12+ thread CPUs |
0 turns the spin off |
EDEN_NX2_TENSOR |
software | fast runs tensor instructions on RTX tensor cores (experimental) |
EDEN_NX2_AVX512 |
on | 0 makes the JIT use AVX2 code only |
TRUDY_FENCE_BATCH |
off | <n> lets up to n fences share one GPU submission (experimental) |
TRUDY_GPU_TIMING / TRUDY_GPU_TRACE |
off | 1 logs host GPU time per shader/pass/command (profiling; costs speed) |
Ocarina of Time: accept the recommended settings when Trudy offers them (Yes is the default, so Enter accepts them). They keep the built-in 30 FPS patch, which holds the game at full speed; without it the game runs at ~56 FPS on a 9800X3D + RTX 4090, and below 60 its game time slows down with the frame rate.
Releases are on the releases page. Each one has:
| File | For |
|---|---|
trudy-<version>-windows-x86_64-zen5.zip |
Windows on Ryzen 9000 / 9000X3D (Zen 5) |
trudy-<version>-windows-x86_64-pgo.zip |
Windows, any x86-64 CPU, profile-guided (preferred elsewhere) |
trudy-<version>-windows-x86_64.zip |
Windows, MSVC build |
trudy-<version>-linux-x86_64-zen5.AppImage |
Linux on Ryzen 9000 / 9000X3D |
trudy-<version>-linux-x86_64-pgo.AppImage |
Linux, profile-guided |
trudy-<version>-linux-x86_64.AppImage |
Linux |
trudy-<version>-macos-arm64.zip |
macOS 15+ on Apple Silicon, experimental |
The -zen5 builds will not start on older CPUs (they use AVX-512 and other Zen 5 instructions);
use -pgo there. Windows: extract the zip and run trudy.exe. Linux: chmod +x the AppImage
and run it; it needs glibc 2.39 or newer.
Linux. Install the dependencies once:
# Arch / CachyOS
sudo pacman -Syu --needed base-devel boost catch2 cmake enet ffmpeg fmt git glslang libzip lz4 \
ninja nlohmann-json openssl opus qt6-base qt6-multimedia qt6-charts qt6-wayland sdl3 zlib zstd \
zip unzip vulkan-headers vulkan-utility-libraries libusb spirv-tools spirv-headersUbuntu 26.04, Debian 13 and other distributions: see docs/Deps.md.
Clone, configure and build. -DYUZU_BUILD_PRESET=zen5 targets Ryzen 9000; use native to target
exactly the CPU you build on, or leave it out for a portable build:
git clone https://github.com/joshpack/Trudy.git
cd Trudy
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DYUZU_TESTS=OFF \
-DYUZU_BUILD_PRESET=zen5 \
-DVulkanHeaders_FORCE_BUNDLED=ON -DVulkanUtilityLibraries_FORCE_BUNDLED=ON
cmake --build build
./build/bin/trudyWindows (MSVC or MSYS2 clang) and macOS builds are described in the build guide;
the presets work there too. A self-built binary has no update checker; git pull is the update.
To bring in new Orchid fixes:
git remote add upstream https://github.com/MythosOfs/Orchid.git
git pull upstream main- User handbook
- Build guide and dependencies
- Fire Emblem: Fortune's Weave bring-up notes
- Switch 2 loading notes
- Development tools
- Fork scan: taking fixes from Eglantine forks
Trudy is built on Orchid by MythosOfs and everyone credited below; all of the emulation work is theirs. Orchid's own credits follow.
Orchid takes fixes from Eglantine and its forks; after every release tools/fork-scan.sh lists what they fixed since. Each ported commit names its source and author in its message.
-
zortos293/Eglantine (Zortos), the project Orchid started from: paged scalar guest memory access, SPSCQueue value lifetime, timed Event waits on Windows, GPU teardown order, buffer cache and query fixes, native Ampere nanosecond timestamps, the separate pipeline cache file names. Radu Ursache: no mesh pipelines without
VK_EXT_mesh_shader. Kaj (co-authored with Claude Opus 5.5): skip guest uploads for new render targets fully overwritten by a clear, adapted with framebuffer render-area, transformed-scissor, linear-attachment and zero-layer guards from 27226d2568c99c70cf78aa37adaf98b60457b2e5. Paulo Tumabing: 32-bit globalATOM.CAS, adapted to Orchid's storage lowering and atomic routing. -
Djihads80/Eglantine and its contributors: merged regularly. Idle staging buffers freed (Takoroka), the RangeSet empty-query guard (néo Luigi), test fixes (kraibse).
-
ZnOw01/Primrose (ZnOw01): shader cache invalidation lifetime fix, adapted to renderer-thread retirement.
-
exploitizm/Eglantine (exploitizm): signed IADD32 negative-overflow boundary fix for GLSL and SPIR-V, adapted from c2f9811073bfe617524d4413742a4d4406fd287a.
-
Selevet Eglantine (sylvieistaken): the windowed GPU mapping table diff, hot loop shortcuts, F2I, local spill and texture query lowering, device memory rebuild on layout changes, decoder end of stream, Windows commit retries, and the GPU wait knobs (
EDEN_NX2_GPU_WAIT_SPIN_US,EDEN_NX2_BACK_SPIN,EDEN_NX2_ACQUIRE_QUEUED,EDEN_NX2_SWAP_STABLE,EDEN_NX2_AVX512). -
PokeDinner (Franseven, mirror), partly from Eden-NS2: any cache hint on native LDG, SUST with a partial channel mask,
EDEN_AUTO_BUTTONS. -
Forks of those: Dee1, Greencus and exploitizm (audio manager lifetime, macro binds past the start table), LucyferDev (A64 saturating shift and multiply-accumulate instructions), FZout (page table block ids, game list titles), sagaopc (NUP directory cycles), Quixomatic (eden-cli, Switch 2 module names), YLTai (SDL motion after reconnect), Aidan Kinzett and oledz-i (NUP detection, geometry stream 0), NicholasBly and ominixisboss (dynarmic), kipavy (storage CAS loop), adamhahnHU and arisqz (sparse layered images), Pavel Barabanov and andrewleeog (F2I bits, host1x syncpoint actions), MFS-code (Switch 2 GOB order), CarlosKeese and Dennys Kaluzny Pereira (lowering), failaip12 (memory manager bounds).
-
After 0.6.3: godpow and its contributors nowhere02, xDarkWindx, Duck912, ZnOw01 and CloudP4bbs (buffer, texture and staging cache guards, descriptor growth, GPU split wake-up recovery, NPad update filter), failaip12 via cira0100 (data races, UBSan fixes, audio and SDL locking), monteger8 for adamd (Donkey Kong Bananza buffer cache skips, indirect draw reset by hash), Aidan Kinzett (NV12 video decoder), Jonas via jojo-chaechae (x64 call reach, Switch 2 channel setup), SavageCore (mod root guards), cblock85, christianball123, Khobzpro, larryliu985, Doughston (n-popescu) and Xnos666.
-
After 0.7.0: Zortos (native BRA.U, register-handle image access, uniform shifts, global memory helpers, the H.264 decoder reset, Aftermath dumps, Switch 2 mixer stream audio), zirulo via Zortos (the Ocarina of Time boot fix), Paulo Rossatto (loose base dumps, GPU-only regions, 2D blit formats), Raikaru (ssl:s contexts and socket timeouts), Mcm190 (H.264 packet framing), Yuri Olive (decoder end of stream), rdenno (fence download ranges), kraibse (tessellation pipelines), a55uka (contentless NUP sections), Andreas Heider (ANV barriers), Aidan Kinzett (button glyphs), Jonas (H0_NH1 halves), exploitizm (IMAD, I2F and FSET lowering, memory manager reference counts), christianball123 (device index fallback), kdegonia-star (global atomic add) and Takoroka (ISETP.EX and UIMAD.WIDE lowering).
-
rdenno: the Mario Tennis Fever fixes in 0.5.1.
-
Drizz1le/Lichii (Inc0rporate, with Claude Opus 5.5): count nested shader returns before reordering demotes for AMD drivers.
-
DrSergioRFilho/Lichii (Dr. Sergio, with Claude Fable 5.1): apply LayeredFS mods to NUP titles without reapplying their merged update, from fe29f5c45ef8f8ab826f148827a1b3962c32d7d7.
-
chuppatricky/Lichii (chuppatricky, with Claude Opus 5.5): shadow Ampere RED and SURED data registers, adapted from fc5d6dc31b49ae614d06c05302f4ee2abdf8ce4c.
-
SoyMoikas/Eglantine (SoyMoikas): uniform-base native Ampere local stores and loads, adapted to Orchid's existing vector lowering and register shadowing from 3c30f286d9631581ea3611c0f30eae46e7f60694 and 8c06623a4651e5f31d7297932f6b9887dfea9006.
-
Hydro74000/Edlws (Hydro74000): make
UniqueFunctionown copies of lvalue callables while moving rvalues, from dd8039f6afc92803938d641a484312931acae9ca. -
DrSergioRFilho/Lichii (Dr. Sergio, with Claude Fable 5.1): check nested config array bounds against the innermost array, from 08d3388f0f5496be67dd5f83dfdcddd0fcd59358.
-
Shenopi/Lichii (999sian): preserve active custom settings while reloading global configuration, adapted from the config-only changes in 88ffd1f312f9a38fd45575dc8e1c7d66ef9b4f6e.
-
trianglebraindev/Lichii (trianglebraindev, with Claude Opus 5.5): two-source
F2IP.U8.F32byte packing, adapted to Orchid's register shadowing from ac82c78b5ca4cfbcb11e3043f05a3354f444f027. -
leadaenima/Lichii (Cursor Agent, co-authored with leadaenima): AMDVLK dynamic subgroup shuffle workaround, adapted from 957913cf14ca1e3776a4984151b63434af543217 and 201450e90a87b494a8d1dc34d32558033a715438. Graphics use the 32-lane broadcast fallback only when subgroup size 32 can be enforced for that stage. Compute and unsupported graphics stages retain Shuffle; Qualcomm keeps its existing full-subgroup fallback. Pipeline hash diagnostics are omitted.
GPLv3 or any later version, see LICENSE.txt. Forked from Orchid, which started from Eglantine, which builds on Eden, yuzu and Sudachi.
