Model-3-Vulkan

Émulateur 0.9.9.8

Réécriture Vulkan de Supermodel pour l'arcade Sega Model 3, avec des binaires macOS.

Model-3-Vulkan est une réécriture autonome en Vulkan de Supermodel, l'émulateur d'arcade Sega Model 3. Le projet publie des archives distinctes pour macOS Apple Silicon et Intel.

Dirt Devils

Dernière mise à jour :

Plateformes

Architecture

Mac OS X - macOS Mac OS X - macOS Apple Silicon Apple Silicon 64 bits 64 bits

Téléchargement et liens

Historique des versions

0.9.9.8
  • Optimisation JIT hybride couvrant plus de 80 % des instructions PowerPC en natif, réduisant drastiquement les appels au mode interpréteur C.
  • Correction majeure sur les GPU intégrés Intel : suppression d’une écriture redondante de profondeur (gl_FragDepth) bloquant l’optimisation Early-Z/Hi-Z, avec des gains de 20 à 50 % sur des jeux comme VF3tb (+60,9 %) ou Daytona 2.
  • Suppression du délai de chargement des sauvegardes (passé de 3 secondes à 0), sans splash-screen intermédiaire, compatible avec 15 jeux testés.
  • Fix du stutter dans Dirt Devils : optimisation de la gestion des pages 4 Ko partagées entre tables de variables et code chaud, réduisant le temps de simulation par frame de 45–66 ms à 0,28–0,78 ms.
  • Prise en charge native des GPU Apple Silicon (M1–M4) via un build arm64 pur, avec MoltenVK intégré et support de macOS 13.3+.
  • Correction des conflits entre résolution présélectionnée et rendu en basse résolution : les deux paramètres synchronisés pour éviter un rendu forcé à 496×384 malgré une présélection en 2x.

Model 3-Vulkan v0.9.9.8 — Release Notes

2026-11-14 · Windows x64 · Linux x86_64 · macOS Intel · macOS Apple Silicon

Model 3-Vulkan v0.9.9.8 is a performance release. Nothing in the emulation model changed: every game still renders the same picture it did in v0.9.9.6, and the same save states still load. What changed is how much work the CPU and the GPU have to do to get there.

The headline is that the hybrid JIT now translates more than 80 % of the guest instruction stream to native x86-64, the Vulkan renderer stopped paying for a redundant depth write that was silently disabling hardware Early-Z / Hi-Z culling, and UMA GPUs finally get zero-copy geometry upload. Together those turn the integrated-GPU experience from "lock it to 496×384" into a picture that can actually be played at native resolution.

📦 Packages

Platform File Requirements
Windows x64 Model3-Vulkan.v0.9.9.8.for.Windows.zip Windows 10/11, Vulkan 1.3
Linux x86_64 Model3-Vulkan.v0.9.9.8.for.Linux.tar.gz glibc 2.38+, Vulkan 1.3
macOS Intel Model3-Vulkan.v0.9.9.8.for.macOS-Intel.tar.gz macOS 13.0+, bundled MoltenVK 1.3
macOS Apple Silicon Model3-Vulkan.v0.9.9.8.for.macOS-AppleSilicon.tar.gz macOS 13.3+, bundled MoltenVK 1.3

SHA256SUMS.txt is attached to the release. Every archive extracts to a relocatable folder; put your ROM zips in its roms/ directory (or point ROMPath at them). The Linux archive is built on Ubuntu 24.04 with g++-14 and needs no newer than GLIBC_2.38 / CXXABI_1.3.15 / GLIBCXX_3.4.32, so it loads on Ubuntu 24.04, SteamOS 3.5+ and other current distributions. The macOS Intel build carries its own libMoltenVK.dylib and targets macOS 13.0; .zst saves need macOS 14+ because the host core links against a newer Swift runtime, and on macOS 13 the emulator falls back to .st automatically. The Apple Silicon archive is the arm64 slice of the same build: the application, the host core and MoltenVK are all pure arm64, so M1–M4 machines run it natively. Intel Macs should take the Intel-only archive instead.


The short version

  • The JIT now covers over 80 % of the guest instruction stream — more PowerPC instruction classes are translated to native x86-64, so the C interpreter fallback runs far less often. In uncapped-frame-rate sessions (-benchmark, or the performance test button) the emulated machine runs noticeably faster than before.
  • Intel integrated GPUs got 20–50 % faster, and the fix is architectural — the fragment stage no longer writes gl_FragDepth, which had been disabling the GPU's Early-Z / Hi-Z depth rejection on every frame. Occluded pixels are discarded before shading again instead of after it. Measured on Intel Arc Graphics (Core Ultra 5 125H) at 2480×1920 2x MSAA: vf3tb +60.9 %, daytona2 +30.9 %, scuddx +21.9 %, dirtdvls +14.0 %. AMD integrated GPUs use the same architecture and should benefit as well. Discrete GPUs gain too, but they were rarely the bottleneck.
  • Dirt Devils no longer stutters — its variable tables and its hot code share one 4 KB page, which the old "flush the whole page on any store" self-modifying-code check thrashed on. Simulation time per frame dropped from 45–66 ms to 0.28–0.78 ms.
  • Save states now load on frame 0 — the 3-second boot guard is gone by default (0.0), with no splash-in-the-middle. Every hardware block the loader needs is in the save image.
  • Smaller CPU overhead in the renderer — HD-texture descriptor bindings are per-frame rather than per-viewport, and opaque geometry is grouped by vertex source, cutting redundant pipeline binds on the Intel Windows driver.

⚡ PowerPC → x86-64 hybrid JIT: over 80 % native

The JIT shipped in v0.9.9.6 already handled integer arithmetic, compares, branches, byte-reversed loads and stores, single/double float memory ops and register moves. This release extends the native encoder to cover substantially more of the instruction stream; the session-end log line now reads:

PPC: translated code is 812345/987654 instruction(s) native (82.25%).

That line is compiled from the live block tables, so it describes the game that just ran rather than a fixed estimate. With more of the stream native, the dispatcher hands off to the C interpreter far less often, which is where the extra throughput comes from in CPU-bound scenes.

Verification is unchanged and still strict: SM3D_PPC_JIT_VERIFY=1 double-runs every block against the interpreter and aborts on the first difference, and SM3D_PPC_JIT_DUMP=1 writes the active blocks out as disassembly. The C interpreter remains selectable (Core → PowerPC Execution Mode) as the reference path for A/B comparisons.

If you want to see the effect yourself, use Common → Performance test (unlocked frame rate) or -benchmark: those unlock the fixed 57.524 Hz logic pacing so the figure is real throughput rather than the present rate.


🅾️ Intel integrated GPUs: 20–50 % faster

The bug: a depth write that switched off Early-Z

The fragment shader used to end with:

gl_FragDepth = ubo.projection[3][2] * gl_FragCoord.w;

That value is exactly what the rasteriser already produces. Every Model 3 projection in this backend comes from Mat4::FrustumRZ, whose third row is (0, 0, 0, near), so the hardware's interpolated gl_FragCoord.z is identically equal to that expression — the sorted-transparency stage (real3d_oit.frag) already relied on this and never wrote depth. But a fragment shader that writes gl_FragDepth at all cannot have its depth test hoisted above it, so the GPU had to run the entire fragment shader — four 64-byte vertex pulls, a bilinear quad solve and a texture decode — for every pixel, including the ones that were about to be thrown away.

Removing the write restores hardware depth rejection. On a 24-EU-class integrated GPU, where the fragment stage is the bottleneck by a wide margin, that is the whole ballgame.

Measured on Intel Arc Graphics (Core Ultra 5 125H)

-benchmark, 6 seconds per run, same save state, AlphaSortMode=0:

Resolution Game Before After Gain
1920×1080 · MSAA 1x daytona2 87.80 FPS 112.83 FPS +28.5 %
1920×1080 · MSAA 1x dirtdvls 91.17 FPS 105.64 FPS +15.9 %
2480×1920 · MSAA 2x vf3tb 45.48 FPS 73.17 FPS +60.9 %
2480×1920 · MSAA 2x daytona2 37.34 FPS 48.86 FPS +30.9 %
2480×1920 · MSAA 2x scuddx 34.13 FPS 41.62 FPS +21.9 %
2480×1920 · MSAA 2x dirtdvls 41.30 FPS 47.10 FPS +14.0 %

vf3tb at 5× native resolution went from "cannot reach the arcade frame rate" to 1.27× arcade speed. At 1080p the lighter games now sit against the desktop compositor's 120 Hz ceiling, which is why their gains look smaller — they were already limited by presentation, not by the GPU.

Picture verification

Same save state, same emulated frame (-Frame 200), interpolated presents disabled (RefreshRateMode=0), 2480×1920, MSAA 2x:

Game Identical pixels Mean absolute error PSNR
daytona2 99.924 % 0.0185 / 255 49.18 dB
scuddx 99.964 % 0.0153 / 255 47.21 dB
vf3tb 99.941 % 0.0123 / 255 51.64 dB
dirtdvls 99.969 % 0.0091 / 255 51.53 dB

The differences are single-sample tie-breaks on coplanar decals, a fraction of one 8-bit step, and the diff visualisation shows them only along triangle edges. This is the same depth value being tested slightly earlier in the pipeline, not a different picture.

UMA geometry upload

On integrated and unified-memory GPUs the 384 MB ROM vertex buffer is now allocated DEVICE_LOCAL | HOST_VISIBLE | HOST_COHERENT when the driver offers it, and new vertices are written straight through the mapping. That removes the staging copy, the vkCmdCopyBuffer and the TRANSFER_WRITE barrier for every batch of newly cached static geometry. Discrete GPUs fall back to the original device-local buffer automatically, and the log says which path ran:

Vulkan: ROM vertex buffer is host-visible (UMA direct write).

Draw submission

Two changes reduce per-frame CPU in the renderer, which matters most on the Intel Windows driver:

  • The HD-texture descriptor set (set 2) is bound once per frame instead of once per viewport node inside DrawBucket().
  • Opaque geometry is grouped by vertex source (ROM vs. polygon RAM) within each viewport node, so a node switches between the two pipelines at most once instead of alternating. Translucent layers keep their original submission order — grouping them would change how two overlapping translucent surfaces composite.

Both changes are pixel-identical: the four games above compare at 0 differing pixels against the build without them.

What it looks like on an Intel iGPU

Same save state (vf3tb.zst1), 1920×1080 output, 2× MSAA, frame pacing unlocked: 143.9 FPS (6.95 ms/frame, 1 % low 104.5) on integrated Intel Arc graphics — about 2.5× the arcade's 57.52 Hz. Under the heavy 2480×1920 2× MSAA load used in the table above, the same scene went from 45.5 to 73.2 FPS (+60.9%).

VF3tb on Intel Arc at 1080p with unlocked frame pacing: 143.9 FPS, 6.95 ms per frame


🏁 Dirt Devils: the 4 KB page stutter is fixed

dirtdvls has an unusual layout: its variable tables at 0x0005A5A8–0x0005A66C and its hot code at 0x0005AB64–0x0005AEA4 share one 4 KB page. The old self-modifying-code check invalidated every translation on that page whenever any store landed in it, so the game re-translated its own hot code roughly 2,300 times per frame.

The invalidation is now three-level:

  1. a 32-byte line gate,
  2. a one-bit-per-guest-byte map,
  3. a 4 KB page generation counter.

A store only bumps the page generation when the address really overlaps bytes that were translated, and it clears that page's line marks and byte map in the process. ramGen is not touched — that is reserved for icbi, state loads and memory reallocation. Blocks are also constrained to a single 4 KB page so a fetch loop cannot straddle two generations.

Result, same save state, 30 seconds:

Before After
Simulation hand-off per frame 45–66 ms 0.28–0.78 ms
Store-triggered flushes ~2,300 / frame 0
Present rate collapsed holds the panel rate

The code cache is 64 MB, and SM3D_PPC_JIT_VERIFY=1 re-verifies the invalidation against the interpreter block by block (26.6 million blocks, 0 failures on the test matrix).

Before the fix the same save was capped near 20 FPS by re-translation — 45–66 ms of simulation hand-off per frame, no matter how fast the GPU was. The card now renders it at 218.6 FPS (4.57 ms/frame, 1 % low 133.0), 2480×1920 with 2× MSAA and frame pacing unlocked.

Dirt Devils on RTX 5060 Ti at 2480×1920, 2× MSAA with unlocked frame pacing: 218.6 FPS, 4.57 ms per frame


💾 Save states load on frame 0

The deferred load was introduced as a workaround: applying a state to a machine still in its boot ROM left VF3 running at about 1/25 speed, because the 53C810's PCI base address had never been written. That root cause was fixed in v0.9.9.6 by adding the missing hardware blocks to the save image, and v0.9.9.8 finally drops the default guard:

constexpr double kDeferredLoadSeconds = 0.0;
  • Manage Saves → Load and -load-state now apply the state on the first frame, with no start-up artwork flashing in between (loading deferred state ... after 0 emulated frame(s)).
  • Verified across 15 ROM sets — daytona2, dayto2pe, scuddx, scud, scudau, srally2, lemans24, dirtdvls, vf3tb, fvipers2, lostwsga, swtrilgy, magtruck, von2a, spikeofe — covering both .zst containers and legacy .st files.
  • SM3D_DEFER_LOAD_SECONDS= remains as a diagnostic override, still counted in emulated frames rather than wall clock.
  • The usual rule still applies: do not save on a load, transition, score or K.O. screen. Those states can be parked in a cross-frame handshake and may hang when restored.

🖥️ Statistics panel on dual-GPU laptops

On a laptop with both an integrated GPU and a discrete NVIDIA card, the statistics overlay used to name (and read telemetry from) the NVIDIA card even when the Vulkan renderer was running on the integrated GPU, because NVML was probed unconditionally. The readout now checks which device the renderer actually selected: when it is not an NVIDIA adapter, the overlay keeps the Vulkan device name and reads load from the Windows GPU Engine counters, so a session on Intel Arc shows Intel(R) Arc(TM) Graphics and the real 3D utilisation.


🔧 Interface fix: resolution preset vs. low-resolution rendering

Advanced → Picture scaling → Low-resolution 3D rendering pins the 3D scene to 496×384 or 992×768 regardless of the window size. Changing the resolution preset on the Common page did not touch it, so picking 2x (992 x 768) after having used the ultra-low preset still rendered at 496×384 and upscaled — the readout said 496x384 -> 992x768, which looked like the setting had been ignored.

The two controls now stay in step: choosing 1x (496 x 384) sets the internal size to 496×384, 2x (992 x 768) sets it to 992×768, and any higher preset turns low-resolution rendering off.


⚙️ Also in this build

  • -benchmark / performance test — unlocks the 57.524 Hz logic pacing, mutes audio, skips a one-second warm-up and reports real throughput (BENCH: ... FPS ... x arcade). Available from Common → Performance test (unlocked frame rate).
  • Mouse device logging — the log now enumerates each mouse and its SDL id, so a two-light-gun setup can be diagnosed without guessing whether Windows exposed both guns as separate devices.
  • Force-feedback diagnostics — the FFB summary line counts FFStop commands, and a device that rejects a constant-force update is logged once instead of failing silently.
  • Dual-GPU stats — see the section above.
  • Defaults on a clean install — hybrid JIT on, 4.0 Quad audio, automatic presentation cadence with motion interpolation, strong chase-camera stabiliser, warm frosted-glass interface, cabinet bezel, loopback networking, asynchronous sound-board thread. Recording, HD textures, statistics overlay, crosshairs, Sinden border, Vulkan validation, exclusive fullscreen, per-pixel OIT, instance interpolation and MSAA/CRT/glow/FXAA remain off.

🔄 Upgrading

  • Settings and save states are compatible: SUPERMODEL_VERSION is untouched, so existing .zst / .st saves and NVRAM load normally.
  • The default save-state delay changed from 3 seconds to 0; this is the one behavioural difference you will notice immediately.
  • If you had selected an integrated GPU from a testing build and the settings window comes up with blocky artefacts, set Video → GPU back to automatic: that combination instructs the Vulkan context to render on one adapter and present to the display attached to another.

⚠️ Known issues

  • scuddx (Scud Race Deluxe) HUD instruments are still incomplete — a long-standing emulation gap that predates MAME; play the Twin/DX sets (scud, scudau, scudplus, scudplusa).
  • Per-instance interpolation remains off by design: enabling it produces single-frame hairline cracks where a snapped instance meets an interpolated one.
  • Not every moment is saveable: states captured on load/transition/score screens may hang.
  • Force feedback, two light guns, the Sinden border and the two-machine link are implemented but still untested on real hardware in this build.

中文摘要

Model 3-Vulkan v0.9.9.8 是一次纯粹的性能更新:模拟本身没有变化,同一存档仍然逐像素输出同一画面。

三个重点:

  1. 混合 JIT 覆盖了超过 80% 的客户机指令流。更多 PowerPC 指令被翻译成 x86-64 原生代码,落到 C 解释器 的次数大幅减少。在解锁帧率的场景(-benchmark 或「性能测试」按钮)下游戏运行速度明显更快。 每局结束日志会打印 PPC: translated code is N/N instruction(s) native (xx.xx%),可直接核对。
  2. Intel 核显专属优化,实测提升 20–50%(AMD 核显大概率同样受益)。片元着色器不再写 gl_FragDepth——对本后端所有 Mat4::FrustumRZ 投影,它恒等于硬件已经插值出的 gl_FragCoord.z,但写了就等于关掉 Early-Z / Hi-Z:被遮挡的像素仍要跑完整片元着色(4 次 64 字节顶点读取 + 双线性四边形求解 + 贴图解码)。移除后硬件深度剔除恢复。 Intel Arc Graphics(Core Ultra 5 125H)实测 2480×1920 / MSAA 2x:vf3tb +60.9%、 daytona2 +30.9%、scuddx +21.9%、dirtdvls +14.0%;1080p 下 daytona2 87.8 → 112.8 FPS(+28.5%)。画面同帧 A/B:99.92–99.97% 像素完全一致,全图平均误差 0.007–0.019 / 255。
  3. 修复「泥地赛车」(Dirt Devils)运行卡顿。这个游戏的变量表(0x5A5A8–0x5A66C)与热代码 (0x5AB64–0x5AEA4)位于同一个 4 KB 页,旧的「页内任何写入就作废整页翻译」逻辑让它每帧 重译约 2300 次。现在改为三层失效:32 字节行门 + 每字节位图 + 4 KB 页 generation,只有真正 覆盖到已翻译指令字节的写入才作废该页。仿真交接时间从 45–66 ms/帧 降到 0.28–0.78 ms/帧, 存储触发的刷新从每帧约 2300 次降到 0。

实测截图(解锁帧率):RTX 5060 Ti 上《Dirt Devils》2480×1920 / 2× MSAA 为 218.6 FPS (4.57 ms/帧,1% low 133.0);Intel Arc 核显上 vf3tb 1080p / 2× MSAA 为 143.9 FPS (6.95 ms/帧,1% low 104.5,约 2.5× 街机满速)。

其它:

  • UMA 核显零拷贝几何上传:ROM 顶点缓冲在统一内存设备上用 DEVICE_LOCAL | HOST_VISIBLE | HOST_COHERENT,新顶点直接由 CPU 写入映射地址,省掉暂存拷贝、vkCmdCopyBuffer 与 TRANSFER_WRITE 屏障;独显自动回退原路径(日志打印 ROM vertex buffer is host-visible)。
  • 驱动调用减负:高清纹理描述符(set 2)改为每帧绑定一次(原先每个视口节点都绑一次); 同一视口内不透明几何按 ROM / 多边形 RAM 归组,管线切换从交替变为最多一次;半透明层保持 原有提交顺序。两项改动逐像素 0 差异。
  • 存档默认第 0 帧读取:kDeferredLoadSeconds 从 3.0 改为 0.0,不再有读档时整屏启动图, 日志显示 after 0 emulated frame(s)。已用 15 个 ROM set 验证(新 .zst 与旧 .st 都覆盖)。
  • 双显卡笔记本统计面板:在核显上跑时不再显示闲置的 NVIDIA 独显名称与遥测,改为显示实际 Vulkan 设备名与 GPU Engine 占用率。
  • 分辨率预设与低分辨率渲染联动:在「常用」页改分辨率预设时会同步更新 InternalResolution(1x → 496×384,2x → 992×768,3x 及以上 → 关闭),不会再出现选了 2x (992 x 768) 却仍按 496×384 渲染的情况。

默认不打开:录音、高清纹理加载/导出、统计浮层、准星、Sinden 白边、Vulkan 校验层、 独占全屏、逐像素 OIT、实例插值、MSAA/CRT/辉光/FXAA、点对点缩放。 默认打开:混合 JIT、4.0 四声道、自动刷新节奏 + 运动插值、强力追尾镜头稳定、暖色毛玻璃界面。

兼容性:SUPERMODEL_VERSION 未改,旧存档与 NVRAM 继续可用。


Model 3-Vulkan is licensed under GPLv3. Upstream Supermodel attribution and third-party notices are preserved in Data/NOTICE.md, Data/LICENSE-ADDITIONAL-TERMS.md and Data/TRADEMARK.md.

0.9.9.6h
  • Capacité de slots de textures HD multipliée par 4 (2 048 → 8 192) pour éviter les limites lors de l'utilisation de packs massifs.
  • Optimisation du chargement dynamique et du décalage des tuiles pour mieux gérer les textures en streaming et les échanges rapides.
  • Logs améliorés pour suivre les conflits de hachage, les invalidations de cache et les surécritures dynamiques de textures.
  • Mode Hybrid JIT instable : risque de baisses de FPS ; basculez sur l'interpréteur C pur en cas de problèmes.

Known Issue / Experimental Notice:

Hybrid JIT might be unstable in some scenes, as PowerPC Execution Mode is still mid-transition between the pure C interpreter and JIT. If you experience heavy FPS drops, switch back to the pure C interpreter for now. If an issue occurs, please exit and share your \Data\Model3Vulkan.log
It’s critical for ongoing development toward 100% JIT.


🎨 Special Build for HD Texture Pack Creators (Windows x64 Only)

  • 4x Bindless HD Texture Slot Capacity (2,048 -> 8,192 slots): Accommodates massive HD texture packs (3,500+ textures, such as Spikeout Final Edition) in full without hitting the session slot cap.
  • Dynamic Tile-Streaming & Offset-Wrap Optimization: Improved sub-texture hash tracking and cache invalidation for games that stream and swap small texture tiles on the fly across areas.
  • Enhanced Log Diagnostics (Data/Model3Vulkan.log): Added detailed HD texture load/dump status, duplicate hash conflict reporting, and dynamic tile-overwrite tracking for pack authors.

Tip for HD Texture Pack Creators: For games like Spikeout that dynamically reuse texture slots across stages, a few original textures in later areas may not have been dumped during your first playthrough. After updating to v0.9.9.6h, simply turn on Dump Textures and walk through those areas again — the emulator will automatically dump any missing original textures for you.

(Regular players who are not creating or testing massive HD texture packs can stay on v0.9.9.6.)

0.9.9.6
  • JIT hybride PowerPC → x86-64 : accélération jusqu’à 4,1× en scènes CPU-intensives, vérifiée contre l’interprète avec 675 millions de blocs sans différence pixel.
  • Stabilisateur de caméra de poursuite : filtre spring-damper pour éliminer les oscillations horizontales, HUD 3D (vitesse, carte) figé, modes Off/Standard/Strong (NFS-style).
  • Retour du feedback force : initialisation SDL3 corrigée sous Windows, reconnaissance des bases direct-drive (SIMAGIC, Fanatec), calibration live des axes/pedales, table adaptative DPI.
  • Audio 4.0 Quad/5.1 Surround : upmixer centre/LFE, interpolation Catmull-Rom 4-tap, enregistrement multi-canaux lossless (WAV/FLAC), décodage DSB 32-bit float.
  • Sauvegardes corrigées : inclusion du 53C810 PCI et du simulateur réseau, reprise immédiate de VF3/Scud Race sans spin DMA.
  • macOS : support natif Apple Silicon (arm64), correction Retina 4K (fenêtre haute densité), module .zst (macOS 14+), rétrocompatibilité .st sur 13.x.

Model 3-Vulkan v0.9.9.6 — Release Notes


Known Issue / Experimental Notice:

Hybrid JIT might be unstable in some scenes, as PowerPC Execution Mode is still mid-transition between the pure C interpreter and JIT. If you experience heavy FPS drops, switch back to the pure C interpreter for now. If an issue occurs, please exit and share your \Data\Model3Vulkan.log
It’s critical for ongoing development toward 100% JIT.


2026-11-07 · Windows x64 · Linux x86_64 · macOS Intel · macOS Apple Silicon

Model 3-Vulkan v0.9.9.6 is the biggest release this fork has shipped. It is not a maintenance build: the emulation core now has a PowerPC → x86-64 hybrid JIT, the audio path can drive a real 4.0/5.1 arcade cabinet and record it losslessly, the chase camera has a proper Need for Speed-style stabilizer, direct-drive wheels finally get real force feedback and live pedal calibration, and the settings window has been rebuilt around frosted glass and three complete themes.

If you are coming from v0.9.8.6a, the JIT, camera, force-feedback, audio, save-state, graphics and interface sections below are new. The platform section also recaps the v0.9.8.6a / v0.9.8.6x hotfixes so a fresh reader sees the whole picture.

📦 Packages

Platform File Requirements
Windows x64 Model3-Vulkan.v0.9.9.6.for.Windows.zip Windows 10/11, Vulkan 1.3
Linux x86_64 Model3-Vulkan.v0.9.9.6.for.Linux.tar.gz glibc 2.38+, Vulkan 1.3
macOS Intel Model3-Vulkan.v0.9.9.6.for.macOS-Intel.tar.gz macOS 13.3+, bundled MoltenVK 1.3
macOS Apple Silicon Model3-Vulkan.v0.9.9.6.for.macOS-AppleSilicon.tar.gz macOS 13.3+, native arm64

SHA256SUMS.txt is attached to the release. Every archive extracts to a relocatable folder; put your ROM zips in its roms/ directory (or point ROMPath at them).


The short version

  • Up to 4.1× faster in CPU-bound scenes — the PowerPC 603e now translates hot basic blocks to native x86-64, verified against the interpreter for 675 million blocks with pixel-identical output.
  • A chase camera that behaves like a modern racing game — spring-damper yaw stabilisation around the car's pivot, with full-screen 3D HUDs (speedometer, course map, turn arrows) staying perfectly still.
  • Force feedback that finally reaches the wheel — SDL3 haptics initialised correctly on Windows, raw direct-drive wheel bases recognised, single-axis constant force fixed, and a live wheel/pedal calibration lab.
  • A soundtrack, not a stereo mix — native 4.0 Quad and 5.1 Surround, a smart center/LFE upmixer, 4-tap Catmull-Rom SCSP voices, 32-bit float DSB decoding, a C¹ soft-clipper, and a lock-free multi-channel WAV/FLAC recorder.
  • Save states that resume instead of stalling — the 53C810 PCI base address and halt flag and the simulated network board are now part of the save, so VF3 and Scud Race pick up where they left off instead of spinning on the DMA chain.
  • A front end that belongs in 2026 — frosted-glass panels, warm light / warm dark / classic themes, adaptive tables that work at any DPI, per-game audio balance, and a much richer diagnostic log.
  • The platform work that makes it shippable — AMD Windows Vulkan driver limits handled, Linux pinned to a Steam Deck-compatible glibc, macOS Retina 4K fixed, GNOME Wayland title bars restored.

🎯 Try these first

Three settings define this release. The screenshots are from the Windows build; the same options exist on Linux and macOS.

Hybrid JIT — Core → PowerPC Execution Mode

Keep Hybrid JIT (recommended) selected and leave PowerPC Frequency at 0 (auto): 0 reads the board's multiplier, so Daytona 2 (2.1×) runs at 166 MHz without manual tuning. The C interpreter is only a reference for A/B debugging.

Core tab: PowerPC Execution Mode set to Hybrid JIT

5.1 Surround + lossless recording — Sound → Channels

Set Channels to 6 (5.1 Surround) and keep Smart 5.1 Surround Upmixer ticked: the emulator builds a real centre channel and a 120 Hz LFE from the quad mix. Tick Record Multi-Channel Lossless Audio to capture the same 5.1 stream; wav_float32 is the default, flac_pcm24 gives smaller lossless files. Recordings land in Data/Recordings/Sound/.

Sound tab: 6-channel 5.1 output and lossless recording

Chase camera stabilizer — FFB / Sensitivity

The new defaults are Stick steering smoothing: Strong, Stick steering sensitivity: 66% and Chase camera stabilizer: Strong (steadiest horizon). This is what removes the left-right horizon swing on a gamepad and keeps Daytona 2's speedometer, course map and turn arrows still. Prefer the raw arcade camera? Set the stabilizer to Off; Standard (NFS-style follow) is the middle ground.

FFB / Sensitivity tab: stick steering and chase camera stabilizer


⚡ PowerPC 603e → x86-64 hybrid JIT

PowerPCMode = 1 is the new default. Hot PowerPC basic blocks are translated once into native x86-64 and then executed directly, with:

  • a 16 MB RWX code cache and PPC_FAST_MEMORY direct-memory windows;
  • self-modifying-code page tracking;
  • idle-loop skipping;
  • a lock-step verifier that runs every translated block against the interpreter.

Verification: 675 million blocks compared across five games, zero mismatches; the same save state and frame is pixel-identical to the interpreter. In a CPU-bound scene the frame rate went from 55.2 Hz to 224.5 Hz — about 4.1×.

The C interpreter remains one menu entry away (PowerPCMode = 0) for reference runs and A/B tests. The JIT lives only in the Vulkan build; the retired OpenGL build keeps the interpreter.


🎥 Chase camera & steering feel

The new Chase Camera Stabilizer (ChaseCameraStabilizer, Off / Light / Standard / Strong, default Strong) is a spring-damper filter applied to the camera's horizontal yaw around the player car's pivot in view space. It removes the “the whole city swings when I breathe on the stick” problem on a gamepad without touching the car's actual heading. Full-screen 3D HUD viewports are excluded, so Daytona USA 2's speedometer, route map and turn arrows stay nailed in place.

Paired with it:

  • Gamepad Steering Smoothing (Off / Light / Medium / Strong, default Strong) and Sensitivity (25–100%, default 66%) on the FFB / Sensitivity page and in the in-game Tab overlay.
  • Selecting Off resets sensitivity to 100% automatically.
  • Binding steering to a physical racing wheel bypasses both filters completely, preserving 1:1 direct response.

🏎️ Force feedback, pedal calibration & wheel lab

Force feedback in the SDL3 build was not merely tuned — it was fixed end to end:

  • Correct haptic initialisation on Windows. SDL_INIT_HAPTIC is now brought up before the joystick/gamepad subsystems, so DirectInput wheel bases actually appear in SDL's haptic list.
  • Raw joystick wheel bases are accepted. The old isGamepad gate rejected SIMAGIC, MOZA, Fanatec and similar direct-drive bases; they are now opened as haptic devices.
  • Constant force sign and axis encoding fixed, with the 20 ms attack envelope retained for a smooth torque build-up instead of a staircase.
  • Vibration routed through SDL_HAPTIC_SINE, with a rumble fallback for devices that do not expose a sine effect.
  • Hardware autocenter is released, so the game's own spring effect is in control.

On the FFB / Sensitivity page there is now a live calibration lab:

  • centred bidirectional axes, full-travel pedals, deadzones and custom min/off/max ranges;
  • columns measured from the current font and DPI, stretching to fill a wide window and switching to a horizontal scrollbar instead of clipping -32768 when the window is narrow;
  • a one-click “Release H-shifter to neutral (from gears 1–4)” binding generator on the Controls tab.

Direct-drive wheel axis calibration: live X/Y/Z/RX/RY/RZ values with deadzone, saturation and min/off/max

The table is live. A direct-drive base reports six axes (X/Y/Z/RX/RY/RZ); each row shows the raw position, the rest/travel mode (centred bidirectional for the wheel, full travel for pedals), the deadzone (3% by default), saturation (100%) and the captured min / off / max range. On a narrow window the table switches to a horizontal scrollbar instead of clipping -32768.

Force Feedback is off on a fresh install by design: most players use a pad or keyboard, and enabling the drive board would make supported pads rumble constantly. Wheel users: tick Force Feedback on the FFB / Sensitivity page (or launch with -force-feedback) once.


🔊 Audio: from stereo fold-down to a real cabinet

Native 4.0 Quad and 5.1 Surround

NbSoundChannels now supports 1 (Mono), 2 (Stereo), 4 (4.0 Quad [FL, FR, RL, RR]), and 6 (5.1 Surround [FL, FR, FC, LFE, RL, RR]). The default is 4.0 Quad, which matches the arcade cabinet and behaves correctly on a stereo desktop endpoint.

The optional 5.1 upmixer (SurroundUpmix51) synthesises:

  • Center (FC) from the front/rear common-mode signal, lifting vocals, navigator calls and lead melodies out of the left/right pair;
  • LFE through a pre-warped 120 Hz second-order Butterworth low-pass, extracting engine rumble and collision weight.

Lossless multi-channel recorder

RecordWav captures the finished 4.0/5.1 mix through a lock-free SPSC ring buffer and a dedicated disk-writer thread — no fwrite, no mutex, no blocking on the real-time audio callback.

  • wav_float32 — WAV, 32-bit IEEE float
  • wav_pcm16 — WAV, 16-bit PCM
  • flac_pcm24 — FLAC, 24-bit PCM
  • flac_pcm16 — FLAC, 16-bit PCM

Files land in Data/Recordings/Sound/___ch.wav|flac, with proper WAVE_FORMAT_EXTENSIBLE / VORBIS_COMMENT speaker masks for Premiere, DaVinci Resolve and Audacity. Recording keeps the full 4/6-channel mix even on a stereo endpoint, so the file is a true discrete surround capture rather than a fold-down.

Recording is off by default. Turn it on in the Sound page or with -record-wav.

SCSP and DSB fidelity

  • 4-tap Catmull-Rom cubic interpolation on non-FM SCSP PCM voices, with FM carriers and modulators kept bit-exact; this removes the mirror-image aliasing on pitch-shifted engine, tyre and voice samples.
  • Slave-SCSP FM ring-buffer cross-talk fixed, Slots 16–17 EFSDL/EFPAN wired to DSP.EXTS[0..1], LFORE phase reset implemented, and dynamic AR/D1R/D2R/RR/KRS envelope updates corrected.
  • 32-bit float DSB MPEG decoding straight into the 64-tap polyphase resampler, with per-sample gain ramping to remove 57.5 Hz zipper noise.
  • C¹-continuous knee soft-clipper on the master bus: bit-exact below -1.41 dBFS, smooth saturation above it, no flat-topped hard clipping when MusicVolume is pushed past 100%.

Restored music and a hidden-egg fix

  • Looped MPEG intros are back. MpegDec now preserves valid negative pre-loop offsets, which restores Daytona USA 2's 9.9-second attract intro and the 12.6-second race-theme intros.
  • Daytona USA 2 no longer plays Track 8 on every course. Single-button view cycling (VRCycle) used to hold the default VR3 button down on every frame from boot, which triggered the arcade's “hold VR3 while a race loads” Easter egg. View cycling now emits a six-frame momentary pulse only when you actually press the button.
  • Per-game audio balance (Scud Race, Daytona 2, Star Wars Trilogy, Spikeout) tames the arcade's own effects/music divider; any explicit value in the ini or on the command line wins.

💾 Save states and the network board

  • 53C810 PCI state block. The SCSI controller's baseAddress and halt flag are now part of the save state. Loading at the first frame used to leave the 0xF9/0xC1 register windows disabled until the game wrote its PCI configuration, which made VF3 and Scud Race spin on the DMA chain. New saves load immediately; old saves fall back to 0xC0.
  • The simulated network board is saved. CSimNetBoard state, shared RAM and buffer are part of the save, with a synthesised single-machine ready state (and the 256 → 848 segment-size repair) for older saves. Attract-mode saves no longer time out into the boot self-test or hang on “Please Wait” at car select.

Also in this release: the drive board's command queue (one command per frame to the Z80), the wheel servo gain and constant-force/vibration cache, a WheelBoardServo save block, and the FFVibrate effect-count fix.


🖼️ Graphics & presentation

  • Per-pixel transparency sorting (AlphaSortMode = 1) — a fixed-slot SoA A-Buffer, four layers per pixel, resolved per priority. It fixes scenes the two-layer scheme cannot: more than two translucent surfaces at one pixel, and translucent polygons submitted out of depth order. At 4K + MSAA 4× it holds the same 120 Hz as the two-layer path; the old linked-list prototype ran at 27–37 Hz and was discarded.
  • Two-subpass on-chip composite for the default path. 4K + 4× MSAA VF3 went from 226 to 240–241 Hz (about +6%), bit-identical at 1×.
  • Hardware MSAA (2/4/8×), exact quad interpolation, derivative-based texture filtering with a long-axis tap mode, and the arcade-exact checkerboard transparency mode as an opt-in experiment.
  • HD texture packs — PNG, WebP, DDS BC1/BC2/BC3/BC5/BC7, and .dds.zst, with bindless sampling, async decode, chunked GPU upload and hot reload. F4 toggles replacement/original, F3 rescans. Exported dumps now write alpha only where the material actually reads it, eliminating the white dots and dashed lines caused by the arcade's 1-bit alpha dither.
  • High-refresh presentation — Original / Auto / Fixed modes, VRR 2×/3×/4×/8× phase stepping, camera and instance interpolation, F8 hot-switch, and a present-phase calibration that cut about 4.6 ms of end-to-end latency on a Fixed 120 Hz setup.
  • macOS Retina 4K — the game window requests a high-density drawable and the render geometry is computed in device pixels, so 4K panels render at native resolution instead of a 1080p upscale.

🎨 Interface

  • The tab row stays put. Common / Core / Video / Sound / Controls / FFB / Sensitivity / Network are now a fixed header; scrolling a long page no longer takes the tabs with it, so switching pages never means scrolling back to the top first.
  • Frosted-glass panels with a real dual Kawase blur of the game artwork behind them, 1px glass highlight edges, and a 120 ms panel fade.
  • Three complete themes — warm eye-care light (default), warm dark, and the original flat interface reproduced value for value.
  • The game list carries 63 WebP box arts, a cover preview, per-game descriptions, filters, remembered selection/scroll position, and an optional “auto-launch on startup”.
  • Adaptive tables — the wheel/pedal calibration table measures columns at the current font and DPI, stretches when there is room, and shows a horizontal scrollbar instead of clipping when there is not.
  • Single config file — Data/Config/Model3-Vulkan.ini. Old Supermodel.ini files are migrated or removed automatically; a stale GameXMLFile path is repaired at startup.
  • Richer log — Data/Model3Vulkan.log starts with the build timestamp, host memory and CPU count, the audio DSP chain, and every non-default axis calibration value.
  • Portable ROM folders — a rom/roms/romset/romsets folder next to the executable is detected automatically and stored as a relative path, so moving the whole folder keeps working.

🧩 Platform work

  • Windows + AMD: per-viewport uniforms moved out of push constants, shrinking R3DPush from 208 B to 96 B. AMD's Windows Vulkan driver reports a 128-byte limit, which used to make the emulator exit silently after the splash screen.
  • Linux: release builds pinned to Ubuntu 24.04 + g++-14, with a CI guard that fails above GLIBC_2.39; the 24.04 build itself needs at most GLIBC_2.38, so it loads on SteamOS 3.5 and Ubuntu 24.04. The bundled wqy-microhei font has a Latin-glyph guard, /Data/ is discovered on Linux, and libdecor is included for GNOME Wayland title bars.
  • macOS: Retina/4K fix, non-Windows mouse/keyboard input in the in-game Tab overlay, and the .zst HostCore module (macOS 14+; .st remains the fallback on macOS 13).
  • SDL3 input: per-device mice via RawInput, late input sampling just before the simulation step, controller hot-plug, SDL haptics, and the Sinden light-gun border.
  • Config/log cleanup: the legacy Supermodel-LegacyOG.exe is retired from the release package; Model3-Vulkan.exe is the product.

🧾 Upstream 0.4 sync (2026-11-07)

The last round before this release pulls the low-risk fixes from upstream Supermodel 0.4:

  • 53C810 PCI base address and halt flag in save states;
  • thread-safety fixes in the DSB FIFO, SCSP MIDI FIFO and the audio mute flag;
  • drive-board command queue and wheel-servo state;
  • non-merged clone ROM sets (srally2dx, swtrilgyp);
  • Lost World light-gun reload / off-screen indicator (lostwsgo → lostwsgp);
  • VF3 startup viewports with subnormal matrices skipped;
  • Real3D config-register zero-initialisation;
  • the fourth force-feedback effect slot.

⚙️ Defaults on a clean install

The release package ships without an ini, so a fresh install starts from these defaults. This is the list worth knowing before your first launch.

Off by default (opt in when you want them):

Option Key Where
Audio recording RecordWav Sound
HD texture replacement LoadHDTextures Core
HD texture dumping DumpHDTextures Core
Raw memory / texture dumps DumpMemory, DumpTextures Network
Statistics overlay ShowStatsOsd Video
Crosshairs Crosshairs Video
Sinden border SindenBorder Video
Vulkan validation VkValidation Video
Exclusive fullscreen / exclusive refresh ExclusiveFullscreen, ExclusiveRefreshRate Video
Per-pixel OIT AlphaSortMode = 1 Video
Instance interpolation InstanceInterpolation no UI (internal A/B switch)
MSAA / CRT / glow / FXAA MSAA = 0, PostCRT = 0, PostGlow = 0, PostFXAA = 0 Video
Integer scaling / widescreen stretch IntegerScale, WideScreen, WideBackground Video
Force feedback ForceFeedback FFB / Sensitivity

On by default, deliberately:

  • Hybrid JIT (PowerPCMode = 1).
  • 4.0 Quad audio (NbSoundChannels = 4). 5.1 is a deliberate choice; the upmixer only runs when six channels are selected. On a stereo endpoint the 4-channel mix is folded by the emulator, not by SDL.
  • Auto refresh pacing + motion interpolation (RefreshRateMode = 1, MotionInterpolation = 1).
  • Chase camera stabilizer Strong and gamepad steering smoothing Strong (both bypass a real wheel).
  • Warm light frosted-glass UI (UITheme = 0, UiFrostedGlass = 1).
  • Cabinet bezel backdrop (BezelMode = 1) — turn it off in Video if you prefer black bars.
  • Network self-link (Network = 1, SimulateNet = 1, ports 1970/1970 on loopback), required for the netboard games to boot.
  • Async sound board thread and high process priority.

🔄 Upgrading from v0.9.8.6a

  • Save states and NVRAM remain compatible: SUPERMODEL_VERSION is untouched.
  • Supermodel.ini is migrated to Data/Config/Model3-Vulkan.ini on first launch; if both exist, the stale file is removed.
  • Existing .st saves still load; new saves prefer the .zst container. Data/Saves/ is kept.
  • HD texture packs in Data/Assets/HDTextures// keep working; DDS and .dds.zst are new options, not replacements.
  • The pipeline cache is rebuilt once after the upgrade; the first launch is slower than subsequent ones.
  • If you extract into a fresh folder, copy Data/Saves, Data/NVRAM, Data/Recordings, Data/Assets/HDTextures and your ROM folder across.

⚠️ Known issues

  • Scud Race Deluxe (scuddx) HUD — the grey CONDITION panel and some wrong-colored rank text are a long-standing Model 3 emulation gap. Upstream Supermodel 0.4 still shows the same thing on the same save state. The Twin/DX versions (scud, scudau, scudplus, scudplusa) display the HUD correctly; use those if it bothers you.
  • .zst on macOS 13 — the HostCore module needs macOS 14+. The emulator logs it once and falls back to .st; saving and loading still work.
  • Force feedback is opt-in — see the note above. Without a wheel it stays off, which is why a fresh install does not rumble a gamepad.
  • 5.1 on a stereo endpoint — if you select 6 channels, SDL folds the stream to stereo with its own coefficient table; the default 4.0 mode uses the emulator's fold and is the recommended setting for stereo speakers.
  • Recording changes the device stream — with RecordWav on, the mixer keeps a 4/6-channel stream even on a stereo endpoint. That is intentional (the file must be discrete), but it means toggling recording can change the monitoring level slightly.

中文摘要

Model 3-Vulkan v0.9.9.6 是这条分支至今最大的一次更新,不是小修补。

重点推荐先试这三项(截图见上方 Try these first):

  1. Core → PowerPC Execution Mode:保持 Hybrid JIT (recommended),频率 0 即自动。
  2. Sound → Channels:选 6 (5.1 Surround) + Smart 5.1 Surround Upmixer;要录音再勾 Record Multi-Channel Lossless Audio(默认关闭)。
  3. FFB / Sensitivity:Stick steering smoothing = Strong、66%、Chase camera stabilizer = Strong。
  • 混合 JIT:PowerPC 603e 热代码块直接翻译成 x86-64,6.75 亿个块与解释器逐块对照零差异、 同帧像素一致;CPU 瓶颈场景从 55.2 Hz 提升到 224.5 Hz(约 4.1 倍)。
  • 极品飞车式追尾镜头:围绕车身支点的弹簧阻尼水平偏航稳定器,手柄微操不再让整条地平线 左右甩;Daytona 2 的码表、路线图、转弯箭头保持不动。默认「强力」,可关/轻/标准/强力。
  • 真正的力反馈:Windows 下 SDL3 触觉正确初始化,直驱方向盘(SIMAGIC / MOZA / Fanatec 等 raw joystick 底座)可被识别,恒力方向与单轴编码修正,振动走 SDL_HAPTIC_SINE;新增实时 方向盘/踏板死区与全行程校准(六轴表截图见上方力反馈章节)、H 挡一键回空挡。力反馈总开关默认关闭, 用方向盘时在「力反馈 / 灵敏度」页打开即可。
  • 音频全面升级:原生 4.0 四声道(默认)与 5.1 环绕,智能中置/120 Hz 低音炮上混; SCSP 非 FM 通道 4-tap 三次样条插值;DSB MPEG 32-bit Float 直通 + 64-tap 重采样; 主总线 C¹ 软限幅;内置无锁多声道 WAV/FLAC 录音(默认关闭)。
  • 读档更可靠:53C810 PCI 基址与 halt 标志、模拟网络板状态都进了存档,VF3 与 Scud Race 不会再卡在 DMA 链上。Daytona 2 的 9.9 秒前奏和 12.6 秒比赛曲前奏恢复;修掉了「所有赛道都 放 Track 8」的隐藏彩蛋误触发。
  • 画面与界面:逐像素半透明排序、双 Subpass 片上合成、MSAA、精确四边形插值、DDS/.dds.zst 高清纹理包、毛玻璃三主题界面、4K/Retina、高刷插值与 VRR 倍频。
  • 平台:AMD Windows 驱动 128 字节 push constant 限制、Linux 固定 Ubuntu 24.04 构建(CI 守 GLIBC_2.39,实际产物最高只要求 GLIBC_2.38)、GNOME Wayland 标题栏、macOS Retina、 非 Windows 游戏内面板输入全部处理。
  • 下载:Windows x64、Linux x86_64、macOS Intel、macOS Apple Silicon 四个包,另附 SHA256SUMS.txt;Linux 需要 glibc 2.38+,macOS 需要 13.3+(13.x 上 .zst 存档回退为 .st)。

默认不打开:录音、高清纹理加载/导出、原始显存/贴图导出、统计浮层、准星、Sinden 白边、 Vulkan 校验层、独占全屏/独占刷新率、逐像素 OIT、实例插值、MSAA/CRT/辉光/FXAA、点对点缩放/ 宽屏拉伸、力反馈。默认打开:混合 JIT、4.0 四声道、自动刷新节奏 + 运动插值、强力追尾镜头 稳定与手柄转向平滑、暖色毛玻璃界面、机柜背景图、网络自连、异步声板线程。

兼容性:SUPERMODEL_VERSION 未改,旧存档与 NVRAM 继续可用;旧 Supermodel.ini 会自动迁移为 Model3-Vulkan.ini;旧 .st 存档仍可读取,新存档优先 .zst。


Model 3-Vulkan is licensed under GPLv3. Upstream Supermodel attribution and third-party notices are preserved in Data/NOTICE.md, Data/LICENSE-ADDITIONAL-TERMS.md and Data/TRADEMARK.md.

Réagir sur le forum

Une question, un retour d’expérience ? Ouvrez la discussion sur le forum.

Réagir sur le forum