diff --git a/src/86box.c b/src/86box.c
index 6b64cd2ec..aa3afb085 100644
--- a/src/86box.c
+++ b/src/86box.c
@@ -213,6 +213,7 @@ int      do_auto_pause                          = 0;              /* (C) Auto-pa
 int      hook_enabled                           = 1;              /* (C) Keyboard hook is enabled */
 int      test_mode                              = 0;              /* (C) Test mode */
 char     uuid[MAX_UUID_LEN]                     = { '\0' };       /* (C) UUID or machine identifier */
+int      sound_muted                            = 0;              /* (C) Is sound muted? */
 
 int      other_ide_present = 0;                                   /* IDE controllers from non-IDE cards are
                                                                      present */
diff --git a/src/chipset/ali1531.c b/src/chipset/ali1531.c
index e765e6d45..53324f8e6 100644
--- a/src/chipset/ali1531.c
+++ b/src/chipset/ali1531.c
@@ -22,6 +22,7 @@
 #include <wchar.h>
 #define HAVE_STDARG_H
 #include <86box/86box.h>
+#include "cpu.h"
 #include <86box/timer.h>
 
 #include <86box/device.h>
diff --git a/src/chipset/ali1541.c b/src/chipset/ali1541.c
index eed6ddbc5..ebbcd7e63 100644
--- a/src/chipset/ali1541.c
+++ b/src/chipset/ali1541.c
@@ -22,6 +22,7 @@
 #include <wchar.h>
 #define HAVE_STDARG_H
 #include <86box/86box.h>
+#include "cpu.h"
 #include <86box/timer.h>
 
 #include <86box/device.h>
diff --git a/src/chipset/ali1543.c b/src/chipset/ali1543.c
index 69a8990e9..2f9273736 100644
--- a/src/chipset/ali1543.c
+++ b/src/chipset/ali1543.c
@@ -22,6 +22,7 @@
 #include <wchar.h>
 #define HAVE_STDARG_H
 #include <86box/86box.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/device.h>
 #include <86box/io.h>
diff --git a/src/chipset/ali6117.c b/src/chipset/ali6117.c
index 707b528b2..88bb4b572 100644
--- a/src/chipset/ali6117.c
+++ b/src/chipset/ali6117.c
@@ -26,6 +26,7 @@
 #include <86box/io.h>
 #include <86box/pci.h>
 #include <86box/pic.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/pit.h>
 #include <86box/device.h>
diff --git a/src/chipset/gc100.c b/src/chipset/gc100.c
index 23455b952..e9eb05ecf 100644
--- a/src/chipset/gc100.c
+++ b/src/chipset/gc100.c
@@ -27,6 +27,7 @@
 #define HAVE_STDARG_H
 #include <86box/86box.h>
 #include <86box/nmi.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/pit.h>
 #include <86box/mem.h>
diff --git a/src/chipset/sis_5511_h2p.c b/src/chipset/sis_5511_h2p.c
index 543fcacc5..b94e69f26 100644
--- a/src/chipset/sis_5511_h2p.c
+++ b/src/chipset/sis_5511_h2p.c
@@ -22,6 +22,7 @@
 #include <86box/86box.h>
 #include <86box/device.h>
 #include <86box/io.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/dma.h>
 #include <86box/mem.h>
diff --git a/src/chipset/sis_5513_p2i.c b/src/chipset/sis_5513_p2i.c
index 95a890fcf..b2346bbc4 100644
--- a/src/chipset/sis_5513_p2i.c
+++ b/src/chipset/sis_5513_p2i.c
@@ -22,6 +22,7 @@
 #include <86box/86box.h>
 #include <86box/device.h>
 #include <86box/io.h>
+#include "cpu.h"
 #include <86box/timer.h>
 
 #include <86box/dma.h>
diff --git a/src/chipset/sis_5571_h2p.c b/src/chipset/sis_5571_h2p.c
index d04964581..a28553487 100644
--- a/src/chipset/sis_5571_h2p.c
+++ b/src/chipset/sis_5571_h2p.c
@@ -22,6 +22,7 @@
 #include <86box/86box.h>
 #include <86box/device.h>
 #include <86box/io.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/dma.h>
 #include <86box/mem.h>
diff --git a/src/chipset/sis_5581_h2p.c b/src/chipset/sis_5581_h2p.c
index d01e9dd28..83983376d 100644
--- a/src/chipset/sis_5581_h2p.c
+++ b/src/chipset/sis_5581_h2p.c
@@ -22,6 +22,7 @@
 #include <86box/86box.h>
 #include <86box/device.h>
 #include <86box/io.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/dma.h>
 #include <86box/mem.h>
diff --git a/src/chipset/sis_5591_h2p.c b/src/chipset/sis_5591_h2p.c
index 048e7deea..74dd6dfbd 100644
--- a/src/chipset/sis_5591_h2p.c
+++ b/src/chipset/sis_5591_h2p.c
@@ -22,6 +22,7 @@
 #include <86box/86box.h>
 #include <86box/device.h>
 #include <86box/io.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/dma.h>
 #include <86box/mem.h>
diff --git a/src/chipset/sis_5600_h2p.c b/src/chipset/sis_5600_h2p.c
index a15c6fff5..c23309de4 100644
--- a/src/chipset/sis_5600_h2p.c
+++ b/src/chipset/sis_5600_h2p.c
@@ -22,6 +22,7 @@
 #include <86box/86box.h>
 #include <86box/device.h>
 #include <86box/io.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/dma.h>
 #include <86box/mem.h>
diff --git a/src/chipset/sis_85c50x.c b/src/chipset/sis_85c50x.c
index 192ae3767..e932ff6ca 100644
--- a/src/chipset/sis_85c50x.c
+++ b/src/chipset/sis_85c50x.c
@@ -24,6 +24,7 @@
 #include <86box/86box.h>
 #include <86box/device.h>
 #include <86box/io.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/apm.h>
 #include <86box/machine.h>
diff --git a/src/config.c b/src/config.c
index 6c34c4790..07e770822 100644
--- a/src/config.c
+++ b/src/config.c
@@ -168,6 +168,7 @@ load_general(void)
     kbd_req_capture = ini_section_get_int(cat, "kbd_req_capture", 0);
     hide_status_bar = ini_section_get_int(cat, "hide_status_bar", 0);
     hide_tool_bar   = ini_section_get_int(cat, "hide_tool_bar", 0);
+    sound_muted     = ini_section_get_int(cat, "sound_muted", 0);
 
     confirm_reset = ini_section_get_int(cat, "confirm_reset", 1);
     confirm_exit  = ini_section_get_int(cat, "confirm_exit", 1);
@@ -1846,6 +1847,10 @@ save_general(void)
 
     const char *va_name;
 
+    ini_section_set_int(cat, "sound_muted", sound_muted);
+    if (sound_muted == 0)
+        ini_section_delete_var(cat, "sound_muted");
+
     ini_section_set_int(cat, "vid_resize", vid_resize);
     if (vid_resize == 0)
         ini_section_delete_var(cat, "vid_resize");
diff --git a/src/device/cartridge.c b/src/device/cartridge.c
index d2e91ecb3..00464026a 100644
--- a/src/device/cartridge.c
+++ b/src/device/cartridge.c
@@ -22,6 +22,7 @@
 #include <wchar.h>
 #define HAVE_STDARG_H
 #include <86box/86box.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/plat.h>
 #include <86box/ui.h>
diff --git a/src/device/keyboard_xt.c b/src/device/keyboard_xt.c
index 644168004..ddbcae61b 100644
--- a/src/device/keyboard_xt.c
+++ b/src/device/keyboard_xt.c
@@ -29,6 +29,7 @@
 #include <wchar.h>
 #include <86box/86box.h>
 #include <86box/device.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/fdd.h>
 #include <86box/machine.h>
diff --git a/src/device/serial.c b/src/device/serial.c
index c25da0070..deb97225a 100644
--- a/src/device/serial.c
+++ b/src/device/serial.c
@@ -27,6 +27,7 @@
 #define HAVE_STDARG_H
 #include <86box/86box.h>
 #include <86box/device.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/machine.h>
 #include <86box/io.h>
diff --git a/src/include/86box/86box.h b/src/include/86box/86box.h
index e4e5ecae1..79bb84009 100644
--- a/src/include/86box/86box.h
+++ b/src/include/86box/86box.h
@@ -158,6 +158,7 @@ extern int      other_scsi_present;         /* SCSI controllers from non-SCSI ca
 extern int    hard_reset_pending;
 extern int    fixed_size_x;
 extern int    fixed_size_y;
+extern int    sound_muted;                  /* (C) Is sound muted? */
 extern int    do_auto_pause;                /* (C) Auto-pause the emulator on focus loss */
 extern int    auto_paused;
 extern double mouse_sensitivity;            /* (C) Mouse sensitivity scale */
diff --git a/src/include/86box/timer.h b/src/include/86box/timer.h
index 25aff6b2f..37a03d9ca 100644
--- a/src/include/86box/timer.h
+++ b/src/include/86box/timer.h
@@ -1,7 +1,7 @@
 #ifndef _TIMER_H_
 #define _TIMER_H_
 
-#include "cpu.h"
+extern uint64_t tsc;
 
 /* Maximum period, currently 1 second. */
 #define MAX_USEC64    1000000ULL
diff --git a/src/include/86box/vid_8514a.h b/src/include/86box/vid_8514a.h
index 57e98cc44..bfde22d5e 100644
--- a/src/include/86box/vid_8514a.h
+++ b/src/include/86box/vid_8514a.h
@@ -213,7 +213,7 @@ typedef struct ibm8514_t {
     int     vdisp2;
     int     disp_cntl;
     int     interlace;
-    uint8_t subsys_cntl;
+    uint16_t subsys_cntl;
     uint8_t subsys_stat;
 
     atomic_int fifo_idx;
diff --git a/src/include/86box/video.h b/src/include/86box/video.h
index 7ad29838c..2985559e1 100644
--- a/src/include/86box/video.h
+++ b/src/include/86box/video.h
@@ -69,6 +69,8 @@ enum {
 #define VIDEO_FLAG_TYPE_NONE    5
 #define VIDEO_FLAG_TYPE_MASK    7
 
+#define VIDEO_FLAG_TYPE_SECONDARY VIDEO_FLAG_TYPE_SPECIAL
+
 typedef struct video_timings_t {
     int type;
     int write_b;
diff --git a/src/machine/m_at_socket5.c b/src/machine/m_at_socket5.c
index b6e82301b..02922b425 100644
--- a/src/machine/m_at_socket5.c
+++ b/src/machine/m_at_socket5.c
@@ -29,6 +29,7 @@
 #include <86box/fdc_ext.h>
 #include <86box/hdc.h>
 #include <86box/hdc_ide.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/fdd.h>
 #include <86box/fdc.h>
diff --git a/src/machine/m_xt_olivetti.c b/src/machine/m_xt_olivetti.c
index b6e4fb94a..34ca441ec 100644
--- a/src/machine/m_xt_olivetti.c
+++ b/src/machine/m_xt_olivetti.c
@@ -30,6 +30,7 @@
 #include <wchar.h>
 #define HAVE_STDARG_H
 #include <86box/86box.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/io.h>
 #include <86box/pic.h>
diff --git a/src/machine/m_xt_philips.c b/src/machine/m_xt_philips.c
index 604ccebaf..eed54e07f 100644
--- a/src/machine/m_xt_philips.c
+++ b/src/machine/m_xt_philips.c
@@ -23,6 +23,7 @@
 #define HAVE_STDARG_H
 #include <86box/86box.h>
 #include <86box/nmi.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/pit.h>
 #include <86box/mem.h>
diff --git a/src/mem/sst_flash.c b/src/mem/sst_flash.c
index b41e0f2e2..34bfae83a 100644
--- a/src/mem/sst_flash.c
+++ b/src/mem/sst_flash.c
@@ -25,6 +25,7 @@
 #include <86box/device.h>
 #include <86box/mem.h>
 #include <86box/machine.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/nvr.h>
 #include <86box/plat.h>
diff --git a/src/qt/icons/sound_mute.ico b/src/qt/icons/sound_mute.ico
new file mode 100644
index 000000000..89a840735
Binary files /dev/null and b/src/qt/icons/sound_mute.ico differ
diff --git a/src/qt/qt_machinestatus.cpp b/src/qt/qt_machinestatus.cpp
index d1335873c..3c3f4019f 100644
--- a/src/qt/qt_machinestatus.cpp
+++ b/src/qt/qt_machinestatus.cpp
@@ -19,9 +19,6 @@
 #include "qt_machinestatus.hpp"
 
 extern "C" {
-#define EMU_CPU_H // superhack - don't want timer.h to include cpu.h here, and some combo is preventing a compile
-extern uint64_t tsc;
-
 #include <86box/hdd.h>
 #include <86box/timer.h>
 #include <86box/86box.h>
@@ -42,6 +39,7 @@ extern uint64_t tsc;
 #include <86box/network.h>
 #include <86box/ui.h>
 #include <86box/machine_status.h>
+#include <86box/config.h>
 };
 
 #include <QIcon>
@@ -92,7 +90,7 @@ struct Pixmaps {
     PixmapSetEmptyActive mo;
     PixmapSetActive      hd;
     PixmapSetEmptyActive net;
-    QPixmap              sound;
+    QPixmap              sound, soundMuted;
 };
 
 struct StateActive {
@@ -218,6 +216,7 @@ struct MachineStatus::States {
         pixmaps.hd.load("/hard_disk%1.ico");
         pixmaps.net.load("/network%1.ico");
         pixmaps.sound = ProgSettings::loadIcon("/sound.ico").pixmap(pixmap_size);
+        pixmaps.soundMuted = ProgSettings::loadIcon("/sound_mute.ico").pixmap(pixmap_size);
 
         cartridge[0].pixmaps = &pixmaps.cartridge;
         cartridge[1].pixmaps = &pixmaps.cartridge;
@@ -259,12 +258,19 @@ MachineStatus::MachineStatus(QObject *parent)
     , refreshTimer(new QTimer(this))
 {
     d = std::make_unique<MachineStatus::States>(this);
+    muteUnmuteAction = nullptr;
     connect(refreshTimer, &QTimer::timeout, this, &MachineStatus::refreshIcons);
     refreshTimer->start(75);
 }
 
 MachineStatus::~MachineStatus() = default;
 
+void
+MachineStatus::setSoundGainAction(QAction* action)
+{
+    soundGainAction = action;
+}
+
 bool
 MachineStatus::hasCassette()
 {
@@ -497,6 +503,28 @@ MachineStatus::refresh(QStatusBar *sbar)
     }
     sbar->removeWidget(d->sound.get());
 
+    if (!muteUnmuteAction) {
+        muteUnmuteAction = new QAction;
+        connect(muteUnmuteAction, &QAction::triggered, this, [this]() {
+            sound_muted ^= 1;
+            config_save();
+            if (d->sound)
+                d->sound->setPixmap(sound_muted ? d->pixmaps.soundMuted : d->pixmaps.sound);
+            
+            muteUnmuteAction->setText(sound_muted ? tr("&Unmute") : tr("&Mute"));
+        });
+    }
+
+    if (!soundMenu) {
+        soundMenu = new QMenu((QWidget*)parent());
+
+        soundMenu->addAction(muteUnmuteAction);
+        soundMenu->addSeparator();
+        soundMenu->addAction(soundGainAction);
+
+        muteUnmuteAction->setParent(soundMenu);
+    }
+
     if (cassette_enable) {
         d->cassette.label = std::make_unique<ClickableLabel>();
         d->cassette.setEmpty(QString(cassette_fname).isEmpty());
@@ -665,12 +693,14 @@ MachineStatus::refresh(QStatusBar *sbar)
     }
 
     d->sound = std::make_unique<ClickableLabel>();
-    d->sound->setPixmap(d->pixmaps.sound);
-
-    connect(d->sound.get(), &ClickableLabel::doubleClicked, d->sound.get(), [](QPoint pos) {
-        SoundGain gain(main_window);
-        gain.exec();
+    d->sound->setPixmap(sound_muted ? d->pixmaps.soundMuted : d->pixmaps.sound);
+    if (muteUnmuteAction)
+        muteUnmuteAction->setText(sound_muted ? tr("&Unmute") : tr("&Mute"));
+    
+    connect(d->sound.get(), &ClickableLabel::clicked, this, [this](QPoint pos) {
+        this->soundMenu->popup(pos - QPoint(0, this->soundMenu->sizeHint().height()));
     });
+
     d->sound->setToolTip(tr("Sound"));
     sbar->addWidget(d->sound.get());
     d->text = std::make_unique<QLabel>();
diff --git a/src/qt/qt_machinestatus.hpp b/src/qt/qt_machinestatus.hpp
index fc0e33e91..90b420763 100644
--- a/src/qt/qt_machinestatus.hpp
+++ b/src/qt/qt_machinestatus.hpp
@@ -1,6 +1,8 @@
 #ifndef QT_MACHINESTATUS_HPP
 #define QT_MACHINESTATUS_HPP
 
+#include <QAction>
+#include <QMenu>
 #include <QWidget>
 #include <QLabel>
 #include <QMouseEvent>
@@ -71,6 +73,7 @@ public:
 
     QString getMessage();
     void    clearActivity();
+    void    setSoundGainAction(QAction* action);
 public slots:
     void refresh(QStatusBar *sbar);
     void message(const QString &msg);
@@ -82,6 +85,9 @@ private:
     struct States;
     std::unique_ptr<States> d;
     QTimer                 *refreshTimer;
+    QAction                *soundGainAction;
+    QAction                *muteUnmuteAction;
+    QMenu                  *soundMenu;
 };
 
 #endif // QT_MACHINESTATUS_HPP
diff --git a/src/qt/qt_main.cpp b/src/qt/qt_main.cpp
index 7723261b4..719208c18 100644
--- a/src/qt/qt_main.cpp
+++ b/src/qt/qt_main.cpp
@@ -86,6 +86,7 @@ extern MainWindow   *main_window;
 
 extern "C" {
 #include <86box/keyboard.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/nvr.h>
 extern int qt_nvr_save(void);
diff --git a/src/qt/qt_mainwindow.cpp b/src/qt/qt_mainwindow.cpp
index 89d17a07c..61f5f6fe1 100644
--- a/src/qt/qt_mainwindow.cpp
+++ b/src/qt/qt_mainwindow.cpp
@@ -48,11 +48,10 @@ extern "C" {
 #include <86box/machine.h>
 #include <86box/vid_ega.h>
 #include <86box/version.h>
-#if 0
-#include <86box/acpi.h> /* Requires timer.h include, which conflicts with Qt headers */
-#endif
-extern atomic_int acpi_pwrbut_pressed;
-extern int acpi_enabled;
+#include <86box/timer.h>
+#include <86box/apm.h>
+#include <86box/nvr.h>
+#include <86box/acpi.h>
 
 #ifdef USE_VNC
 #    include <86box/vnc.h>
@@ -179,6 +178,7 @@ MainWindow::MainWindow(QWidget *parent)
     extern MainWindow *main_window;
     main_window = this;
     ui->setupUi(this);
+    status->setSoundGainAction(ui->actionSound_gain);
     ui->stackedWidget->setMouseTracking(true);
     statusBar()->setVisible(!hide_status_bar);
     statusBar()->setStyleSheet("QStatusBar::item {border: None; } QStatusBar QLabel { margin-right: 2px; margin-bottom: 1px; }");
diff --git a/src/qt/qt_settingsdisplay.cpp b/src/qt/qt_settingsdisplay.cpp
index 75f8d376b..b7a930711 100644
--- a/src/qt/qt_settingsdisplay.cpp
+++ b/src/qt/qt_settingsdisplay.cpp
@@ -220,9 +220,9 @@ SettingsDisplay::on_comboBoxVideo_currentIndexChanged(int index)
         int secondaryFlags = video_card_get_flags(c);
         if (video_card_available(c)
             && device_is_valid(video_dev, machineId)
-            && !((secondaryFlags == primaryFlags) && (secondaryFlags != VIDEO_FLAG_TYPE_SPECIAL))
-            && !(((primaryFlags == VIDEO_FLAG_TYPE_8514) || (primaryFlags == VIDEO_FLAG_TYPE_XGA)) && (secondaryFlags != VIDEO_FLAG_TYPE_MDA) && (secondaryFlags != VIDEO_FLAG_TYPE_SPECIAL))
-            && !((primaryFlags != VIDEO_FLAG_TYPE_MDA) && (primaryFlags != VIDEO_FLAG_TYPE_SPECIAL) && ((secondaryFlags == VIDEO_FLAG_TYPE_8514) || (secondaryFlags == VIDEO_FLAG_TYPE_XGA)))) {
+            && !((secondaryFlags == primaryFlags) && (secondaryFlags != VIDEO_FLAG_TYPE_SECONDARY))
+            && !(((primaryFlags == VIDEO_FLAG_TYPE_8514) || (primaryFlags == VIDEO_FLAG_TYPE_XGA)) && (secondaryFlags != VIDEO_FLAG_TYPE_MDA) && (secondaryFlags != VIDEO_FLAG_TYPE_SECONDARY))
+            && !((primaryFlags != VIDEO_FLAG_TYPE_MDA) && (primaryFlags != VIDEO_FLAG_TYPE_SECONDARY) && ((secondaryFlags == VIDEO_FLAG_TYPE_8514) || (secondaryFlags == VIDEO_FLAG_TYPE_XGA)))) {
             ui->comboBoxVideoSecondary->addItem(name, c);
             if (c == curVideoCard_2)
                 ui->comboBoxVideoSecondary->setCurrentIndex(ui->comboBoxVideoSecondary->count() - 1);
diff --git a/src/qt/qt_settingsmachine.cpp b/src/qt/qt_settingsmachine.cpp
index 771f1cbec..9b1f9849e 100644
--- a/src/qt/qt_settingsmachine.cpp
+++ b/src/qt/qt_settingsmachine.cpp
@@ -32,13 +32,9 @@ extern "C" {
 #include <86box/config.h>
 #include <86box/device.h>
 #include <86box/machine.h>
+#include <86box/nvr.h>
 }
 
-// from nvr.h, which we can't import into CPP code
-#define TIME_SYNC_DISABLED 0
-#define TIME_SYNC_ENABLED  1
-#define TIME_SYNC_UTC      2
-
 #include "qt_deviceconfig.hpp"
 #include "qt_models_common.hpp"
 
diff --git a/src/qt_resources.qrc b/src/qt_resources.qrc
index 1f6341786..6ca323b89 100644
--- a/src/qt_resources.qrc
+++ b/src/qt_resources.qrc
@@ -43,6 +43,7 @@
         <file>qt/icons/other_removable_devices.ico</file>
         <file>qt/icons/ports.ico</file>
         <file>qt/icons/sound.ico</file>
+        <file>qt/icons/sound_mute.ico</file>
         <file>qt/icons/storage_controllers.ico</file>
         <file>qt/icons/zip.ico</file>
         <file>qt/icons/zip_active.ico</file>
diff --git a/src/scsi/scsi_t128.c b/src/scsi/scsi_t128.c
index a40d803f1..5ac3b5d67 100644
--- a/src/scsi/scsi_t128.c
+++ b/src/scsi/scsi_t128.c
@@ -29,6 +29,7 @@
 #define HAVE_STDARG_H
 #include <86box/86box.h>
 #include <86box/io.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/dma.h>
 #include <86box/pic.h>
diff --git a/src/sound/openal.c b/src/sound/openal.c
index a41199b23..90f626362 100644
--- a/src/sound/openal.c
+++ b/src/sound/openal.c
@@ -278,7 +278,7 @@ givealbuffer_common(const void *buf, const uint8_t src, const int size, const in
 
     alGetSourcei(source[src], AL_BUFFERS_PROCESSED, &processed);
     if (processed >= 1) {
-        const double gain = pow(10.0, (double) sound_gain / 20.0);
+        const double gain = sound_muted ? 0.0 : pow(10.0, (double) sound_gain / 20.0);
         alListenerf(AL_GAIN, (float) gain);
 
         alSourceUnqueueBuffers(source[src], 1, &buffer);
diff --git a/src/sound/snd_audiopci.c b/src/sound/snd_audiopci.c
index e3e2b9f1a..8d1de442b 100644
--- a/src/sound/snd_audiopci.c
+++ b/src/sound/snd_audiopci.c
@@ -39,6 +39,7 @@
 #include <86box/pci.h>
 #include <86box/snd_ac97.h>
 #include <86box/sound.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/plat_unused.h>
 #include <86box/snd_akm4531.h>
diff --git a/src/sound/snd_gus.c b/src/sound/snd_gus.c
index 43638473c..1885581a5 100644
--- a/src/sound/snd_gus.c
+++ b/src/sound/snd_gus.c
@@ -15,6 +15,7 @@
 #include <86box/nmi.h>
 #include <86box/pic.h>
 #include <86box/sound.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #ifdef USE_GUSMAX
 #    include <86box/snd_ad1848.h>
diff --git a/src/sound/snd_opl_esfm.c b/src/sound/snd_opl_esfm.c
index bcd2a56e9..b80d264d5 100644
--- a/src/sound/snd_opl_esfm.c
+++ b/src/sound/snd_opl_esfm.c
@@ -32,6 +32,7 @@
 #include <86box/86box.h>
 #include <86box/sound.h>
 #include <86box/device.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/snd_opl.h>
 #include <86box/plat_unused.h>
diff --git a/src/sound/snd_opl_nuked.c b/src/sound/snd_opl_nuked.c
index 03851a589..60f5ed2a6 100644
--- a/src/sound/snd_opl_nuked.c
+++ b/src/sound/snd_opl_nuked.c
@@ -46,6 +46,7 @@
 #define HAVE_STDARG_H
 #include <86box/86box.h>
 #include <86box/sound.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/device.h>
 #include <86box/snd_opl.h>
diff --git a/src/sound/snd_sb.c b/src/sound/snd_sb.c
index 4f8b8a0bd..689d0b91e 100644
--- a/src/sound/snd_sb.c
+++ b/src/sound/snd_sb.c
@@ -39,6 +39,7 @@
 #include <86box/pic.h>
 #include <86box/rom.h>
 #include <86box/sound.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/snd_sb.h>
 #include <86box/plat_unused.h>
diff --git a/src/sound/xaudio2.c b/src/sound/xaudio2.c
index 9b341f574..2aee97efc 100644
--- a/src/sound/xaudio2.c
+++ b/src/sound/xaudio2.c
@@ -245,7 +245,7 @@ givealbuffer_common(const void *buf, IXAudio2SourceVoice *sourcevoice, const siz
     if (!initialized)
         return;
 
-    (void) IXAudio2MasteringVoice_SetVolume(mastervoice, pow(10.0, (double) sound_gain / 20.0),
+    (void) IXAudio2MasteringVoice_SetVolume(mastervoice, sound_muted ? 0.0 : pow(10.0, (double) sound_gain / 20.0),
                                             XAUDIO2_COMMIT_NOW);
     XAUDIO2_BUFFER buffer = { 0 };
     buffer.Flags          = 0;
diff --git a/src/timer.c b/src/timer.c
index 9fed37511..03908890f 100644
--- a/src/timer.c
+++ b/src/timer.c
@@ -3,6 +3,7 @@
 #include <string.h>
 #include <wchar.h>
 #include <86box/86box.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/nv/vid_nv_rivatimer.h>
 
diff --git a/src/unix/unix.c b/src/unix/unix.c
index 7f653b9b6..33b78ddb1 100644
--- a/src/unix/unix.c
+++ b/src/unix/unix.c
@@ -38,6 +38,7 @@
 #include <86box/device.h>
 #include <86box/gameport.h>
 #include <86box/unix_sdl.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/nvr.h>
 #include <86box/version.h>
diff --git a/src/video/vid_8514a.c b/src/video/vid_8514a.c
index 379772820..765258237 100644
--- a/src/video/vid_8514a.c
+++ b/src/video/vid_8514a.c
@@ -715,27 +715,13 @@ ibm8514_accel_out(uint16_t port, uint32_t val, svga_t *svga, int len)
                 break;
 
             case 0x42e8:
-                if (val & 0x01)
-                    dev->subsys_stat &= ~0x01;
-                if (val & 0x02)
-                    dev->subsys_stat &= ~0x02;
-                if (val & 0x04)
-                    dev->subsys_stat &= ~0x04;
-                if (val & 0x08)
-                    dev->subsys_stat &= ~0x08;
+                ibm8514_log("VBLANK stat=%02x, val=%02x.\n", dev->subsys_stat, val);
+                dev->subsys_cntl = (dev->subsys_cntl & 0xff00) | val;
+                dev->subsys_stat &= ~val;
                 break;
             case 0x42e9:
-                dev->subsys_cntl = val;
-                if (val & 0x01)
-                    dev->subsys_stat |= 0x01;
-                if (val & 0x02)
-                    dev->subsys_stat |= 0x02;
-                if (val & 0x04)
-                    dev->subsys_stat |= 0x04;
-                if (val & 0x08)
-                    dev->subsys_stat |= 0x08;
-
-                if ((val & 0xc0) == 0xc0) {
+                dev->subsys_cntl = (dev->subsys_cntl & 0xff) | (val << 8);
+                if ((val & 0xc0) == 0x80) {
                     dev->fifo_idx = 0;
                     dev->force_busy = 0;
                     dev->force_busy2 = 0;
@@ -882,10 +868,10 @@ ibm8514_accel_in(uint16_t port, svga_t *svga)
 
     switch (port) {
         case 0x2e8:
-            if (dev->vc == dev->v_syncstart)
+            if (dev->vc == dev->dispend)
                 temp |= 0x02;
 
-            ibm8514_log("0x2E8 read: Display Status=%02x.\n", temp);
+            ibm8514_log("Read: Display Status1=%02x.\n", temp);
             break;
 
         case 0x6e8:
@@ -910,21 +896,25 @@ ibm8514_accel_in(uint16_t port, svga_t *svga)
 
         case 0x42e8:
         case 0x42e9:
-            if (dev->vc == dev->v_syncstart)
-                dev->subsys_stat |= 0x01;
+            if ((dev->subsys_cntl & 0x01) && !(dev->subsys_stat & 0x01) && (dev->vc == dev->dispend))
+                temp |= 0x01;
 
             if (cmd == 6) {
-                if ((dev->accel.dx >= clip_l) &&
+                if ((dev->subsys_cntl & 0x02) &&
+                    !(dev->subsys_stat & 0x02) &&
+                    (dev->accel.dx >= clip_l) &&
                     (dev->accel.dx <= clip_r_ibm) &&
                     (dev->accel.dy >= clip_t) &&
                     (dev->accel.dy <= clip_b_ibm))
-                    dev->subsys_stat |= 0x02;
+                    temp |= 0x02;
             } else {
-                if ((dev->accel.cx >= clip_l) &&
+                if ((dev->subsys_cntl & 0x02) &&
+                    !(dev->subsys_stat & 0x02) &&
+                    (dev->accel.cx >= clip_l) &&
                     (dev->accel.cx <= clip_r_ibm) &&
                     (dev->accel.cy >= clip_t) &&
                     (dev->accel.cy <= clip_b_ibm))
-                    dev->subsys_stat |= 0x02;
+                    temp |= 0x02;
             }
 
             if (!dev->fifo_idx) {
@@ -932,9 +922,10 @@ ibm8514_accel_in(uint16_t port, svga_t *svga)
                     temp |= 0x08;
             }
 
-            if (port & 1)
+            if (port & 1) {
                 temp = dev->vram_512k_8514 ? 0x00 : 0x80;
-            else {
+                temp |= (dev->subsys_cntl >> 8);
+            } else {
                 temp |= (dev->subsys_stat | (dev->vram_512k_8514 ? 0x00 : 0x80));
                 temp |= 0x20;
             }
@@ -1155,6 +1146,7 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                         (dev->accel.cx <= clip_r) &&
                         (dev->accel.cy >= clip_t) &&
                         (dev->accel.cy <= clip_b)) {
+                        dev->subsys_stat |= 0x02;
                         switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                             case 0:
                                 src_dat = bkgd_color;
@@ -1250,6 +1242,7 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                         (dev->accel.cx <= clip_r) &&
                         (dev->accel.cy >= clip_t) &&
                         (dev->accel.cy <= clip_b)) {
+                        dev->subsys_stat |= 0x02;
                         switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                             case 0:
                                 src_dat = bkgd_color;
@@ -1385,6 +1378,7 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                         (dev->accel.cx <= clip_r) &&
                         (dev->accel.cy >= clip_t) &&
                         (dev->accel.cy <= clip_b)) {
+                        dev->subsys_stat |= 0x02;
                         if (ibm8514_cpu_dest(svga) && (pixcntl == 0)) {
                             mix_dat = mix_mask; /* Mix data = forced to foreground register. */
                         } else if (ibm8514_cpu_dest(svga) && (pixcntl == 3)) {
@@ -1543,6 +1537,7 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                             (dev->accel.cx <= clip_r) &&
                             (dev->accel.cy >= clip_t) &&
                             (dev->accel.cy <= clip_b)) {
+                            dev->subsys_stat |= 0x02;
                             if (ibm8514_cpu_dest(svga)) {
                                 READ((dev->accel.cy * dev->pitch) + dev->accel.cx, src_dat);
                             } else
@@ -1634,6 +1629,7 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                             (dev->accel.cx <= clip_r) &&
                             (dev->accel.cy >= clip_t) &&
                             (dev->accel.cy <= clip_b)) {
+                            dev->subsys_stat |= 0x02;
                             if (ibm8514_cpu_dest(svga) && (pixcntl == 0)) {
                                 mix_dat = mix_mask; /* Mix data = forced to foreground register. */
                             } else if (ibm8514_cpu_dest(svga) && (pixcntl == 3)) {
@@ -1832,6 +1828,7 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                                 (dev->accel.cx <= clip_r) &&
                                 (dev->accel.cy >= clip_t) &&
                                 (dev->accel.cy <= clip_b)) {
+                                dev->subsys_stat |= 0x02;
                                 switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                                     case 0:
                                         src_dat = bkgd_color;
@@ -1989,6 +1986,7 @@ skip_vector_rect_write:
                             (dev->accel.cx <= clip_r) &&
                             (dev->accel.cy >= clip_t) &&
                             (dev->accel.cy <= clip_b)) {
+                            dev->subsys_stat |= 0x02;
                             if (ibm8514_cpu_dest(svga) && (pixcntl == 0)) {
                                 mix_dat = mix_mask; /* Mix data = forced to foreground register. */
                             } else if (ibm8514_cpu_dest(svga) && (pixcntl == 3)) {
@@ -2150,6 +2148,7 @@ skip_nibble_rect_write:
                                     (dev->accel.cx <= clip_r) &&
                                     (dev->accel.cy >= clip_t) &&
                                     (dev->accel.cy <= clip_b)) {
+                                    dev->subsys_stat |= 0x02;
                                     switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                                         case 0:
                                             src_dat = bkgd_color;
@@ -2232,6 +2231,7 @@ skip_nibble_rect_write:
                                     (dev->accel.cx <= clip_r) &&
                                     (dev->accel.cy >= clip_t) &&
                                     (dev->accel.cy <= clip_b)) {
+                                    dev->subsys_stat |= 0x02;
                                     switch ((mix_dat & 0x01) ? frgd_mix : bkgd_mix) {
                                         case 0:
                                             src_dat = bkgd_color;
@@ -2313,6 +2313,7 @@ skip_nibble_rect_write:
                                 (dev->accel.cx <= clip_r) &&
                                 (dev->accel.cy >= clip_t) &&
                                 (dev->accel.cy <= clip_b)) {
+                                dev->subsys_stat |= 0x02;
                                 switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                                     case 0:
                                         src_dat = bkgd_color;
@@ -2424,6 +2425,7 @@ skip_nibble_rect_write:
                                 (dev->accel.cx <= clip_r) &&
                                 (dev->accel.cy >= clip_t) &&
                                 (dev->accel.cy <= clip_b)) {
+                                dev->subsys_stat |= 0x02;
                                 switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                                     case 0:
                                         src_dat = bkgd_color;
@@ -2546,6 +2548,7 @@ skip_nibble_rect_write:
                         (dev->accel.cx <= clip_r) &&
                         (dev->accel.cy >= clip_t) &&
                         (dev->accel.cy <= clip_b)) {
+                        dev->subsys_stat |= 0x02;
                         switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                             case 0:
                                 src_dat = bkgd_color;
@@ -2650,9 +2653,10 @@ skip_nibble_rect_write:
                     dev->accel.cx = CLAMP(dev->accel.cx, clip_l, clip_r);
 
                     if ((dev->accel.cx >= clip_l) &&
-                        (dev->accel.cx < clip_r) &&
+                        (dev->accel.cx <= clip_r) &&
                         (dev->accel.cy >= clip_t) &&
                         (dev->accel.cy <= clip_b)) {
+                        dev->subsys_stat |= 0x02;
                         switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                             case 0:
                                 src_dat = bkgd_color;
@@ -2811,6 +2815,7 @@ skip_nibble_rect_write:
                         (dev->accel.dx <= clip_r) &&
                         (dev->accel.dy >= clip_t) &&
                         (dev->accel.dy <= clip_b)) {
+                        dev->subsys_stat |= 0x02;
                         if (pixcntl == 3) {
                             if (!(dev->accel.cmd & 0x10) && ((frgd_mix != 3) || (bkgd_mix != 3))) {
                                 READ(dev->accel.src + dev->accel.cx, mix_dat);
@@ -2978,6 +2983,7 @@ skip_nibble_bitblt_write:
                                 (dev->accel.dx <= clip_r) &&
                                 (dev->accel.dy >= clip_t) &&
                                 (dev->accel.dy <= clip_b)) {
+                                dev->subsys_stat |= 0x02;
                                 switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                                     case 0:
                                         src_dat = bkgd_color;
@@ -3075,6 +3081,7 @@ skip_nibble_bitblt_write:
                                 (dev->accel.dx <= clip_r) &&
                                 (dev->accel.dy >= clip_t) &&
                                 (dev->accel.dy <= clip_b)) {
+                                dev->subsys_stat |= 0x02;
                                 switch ((mix_dat & 0x01) ? frgd_mix : bkgd_mix) {
                                     case 0:
                                         src_dat = bkgd_color;
@@ -3174,6 +3181,7 @@ skip_nibble_bitblt_write:
                                 (dx <= (((uint64_t)clip_r) * 3)) &&
                                 (dev->accel.dy >= (clip_t << 1)) &&
                                 (dev->accel.dy <= (clip_b << 1))) {
+                                dev->subsys_stat |= 0x02;
 
                                 READ(dev->accel.src + cx, src_dat);
                                 READ(dev->accel.dest + dx, dest_dat);
@@ -3196,6 +3204,7 @@ skip_nibble_bitblt_write:
                                 (dev->accel.dx <= clip_r) &&
                                 (dev->accel.dy >= clip_t) &&
                                 (dev->accel.dy <= clip_b)) {
+                                dev->subsys_stat |= 0x02;
                                 if (pixcntl == 3) {
                                     if (!(dev->accel.cmd & 0x10) && ((frgd_mix != 3) || (bkgd_mix != 3))) {
                                         READ(dev->accel.src + dev->accel.cx, mix_dat);
@@ -3704,6 +3713,8 @@ ibm8514_poll(void *priv)
             dev->vc &= 0xfff;
 
             if (dev->vc == dev->dispend) {
+                dev->subsys_stat |= 0x01;
+                ibm8514_log("VBLANK irq.\n");
                 dev->dispon = 0;
 
                 for (x = 0; x < ((dev->vram_mask + 1) >> 12); x++) {
diff --git a/src/video/vid_ati68860_ramdac.c b/src/video/vid_ati68860_ramdac.c
index bb5ccbf31..17b2a6d22 100644
--- a/src/video/vid_ati68860_ramdac.c
+++ b/src/video/vid_ati68860_ramdac.c
@@ -70,20 +70,19 @@ void
 ati68860_ramdac_out(uint16_t addr, uint8_t val, void *priv, svga_t *svga)
 {
     ati68860_ramdac_t *ramdac = (ati68860_ramdac_t *) priv;
-    const ibm8514_t   *dev    = (ibm8514_t *) svga->dev8514;
 
     switch (addr) {
         case 0:
-            svga_out((dev && dev->on) ? 0x2ec : 0x3c8, val, svga);
+            svga_out(0x3c8, val, svga);
             break;
         case 1:
-            svga_out((dev && dev->on) ? 0x2ed : 0x3c9, val, svga);
+            svga_out(0x3c9, val, svga);
             break;
         case 2:
-            svga_out((dev && dev->on) ? 0x2ea : 0x3c6, val, svga);
+            svga_out(0x3c6, val, svga);
             break;
         case 3:
-            svga_out((dev && dev->on) ? 0x2eb : 0x3c7, val, svga);
+            svga_out(0x3c7, val, svga);
             break;
         default:
             ramdac->regs[addr & 0xf] = val;
@@ -176,21 +175,20 @@ uint8_t
 ati68860_ramdac_in(uint16_t addr, void *priv, svga_t *svga)
 {
     const ati68860_ramdac_t *ramdac = (ati68860_ramdac_t *) priv;
-    const ibm8514_t         *dev    = (ibm8514_t *) svga->dev8514;
     uint8_t                  temp   = 0;
 
     switch (addr) {
         case 0:
-            temp = svga_in((dev && dev->on) ? 0x2ec : 0x3c8, svga);
+            temp = svga_in(0x3c8, svga);
             break;
         case 1:
-            temp = svga_in((dev && dev->on) ? 0x2ed : 0x3c9, svga);
+            temp = svga_in(0x3c9, svga);
             break;
         case 2:
-            temp = svga_in((dev && dev->on) ? 0x2ea : 0x3c6, svga);
+            temp = svga_in(0x3c6, svga);
             break;
         case 3:
-            temp = svga_in((dev && dev->on) ? 0x2eb : 0x3c7, svga);
+            temp = svga_in(0x3c7, svga);
             break;
         case 4:
         case 8:
diff --git a/src/video/vid_ati_mach64.c b/src/video/vid_ati_mach64.c
index 2df7782ff..24935e3a2 100644
--- a/src/video/vid_ati_mach64.c
+++ b/src/video/vid_ati_mach64.c
@@ -28,6 +28,7 @@
 #include <86box/device.h>
 #include <86box/io.h>
 #include <86box/mem.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/pci.h>
 #include <86box/rom.h>
diff --git a/src/video/vid_ati_mach8.c b/src/video/vid_ati_mach8.c
index 71dd973da..6e2798295 100644
--- a/src/video/vid_ati_mach8.c
+++ b/src/video/vid_ati_mach8.c
@@ -27,6 +27,7 @@
 #include <86box/device.h>
 #include <86box/io.h>
 #include <86box/mem.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/mca.h>
 #include <86box/pci.h>
@@ -437,6 +438,7 @@ mach_accel_start(int cmd_type, int cpu_input, int count, uint32_t mix_dat, uint3
                         (dev->accel.dx <= clip_r) &&
                         (dev->accel.dy >= clip_t) &&
                         (dev->accel.dy <= clip_b)) {
+                        dev->subsys_stat |= 0x02;
                         switch (mix ? frgd_sel : bkgd_sel) {
                             case 0:
                                 src_dat = dev->accel.bkgd_color;
@@ -660,6 +662,7 @@ mach_accel_start(int cmd_type, int cpu_input, int count, uint32_t mix_dat, uint3
                         (dev->accel.dx <= clip_r) &&
                         (dev->accel.dy >= clip_t) &&
                         (dev->accel.dy <= clip_b)) {
+                        dev->subsys_stat |= 0x02;
                         switch (mix ? frgd_sel : bkgd_sel) {
                             case 0:
                                 src_dat = dev->accel.bkgd_color;
@@ -1066,6 +1069,7 @@ mach_accel_start(int cmd_type, int cpu_input, int count, uint32_t mix_dat, uint3
                     (dev->accel.dx <= clip_r) &&
                     (dev->accel.dy >= clip_t) &&
                     (dev->accel.dy <= clip_b)) {
+                    dev->subsys_stat |= 0x02;
                     if (mach->accel.dp_config & 0x02) {
                         READ(dev->accel.src + dev->accel.cx, poly_src);
                         poly_src = ((poly_src & rd_mask) == rd_mask);
@@ -1307,6 +1311,7 @@ mach_accel_start(int cmd_type, int cpu_input, int count, uint32_t mix_dat, uint3
                             (dev->accel.cx <= clip_r) &&
                             (dev->accel.cy >= clip_t) &&
                             (dev->accel.cy <= clip_b)) {
+                            dev->subsys_stat |= 0x02;
                             mach->accel.clip_overrun = 0;
                             switch (mix ? frgd_sel : bkgd_sel) {
                                 case 0:
@@ -1440,6 +1445,7 @@ mach_accel_start(int cmd_type, int cpu_input, int count, uint32_t mix_dat, uint3
                             (dev->accel.cx <= clip_r) &&
                             (dev->accel.cy >= clip_t) &&
                             (dev->accel.cy <= clip_b)) {
+                            dev->subsys_stat |= 0x02;
                             mach->accel.clip_overrun = 0;
                             if (mach->accel.linedraw_opt & 0x02) {
                                 if (dev->bpp) {
@@ -1598,6 +1604,7 @@ mach_accel_start(int cmd_type, int cpu_input, int count, uint32_t mix_dat, uint3
                             (dev->accel.cx <= clip_r) &&
                             (dev->accel.cy >= clip_t) &&
                             (dev->accel.cy <= clip_b)) {
+                            dev->subsys_stat |= 0x02;
                             mach->accel.clip_overrun = 0;
                             switch (mix ? frgd_sel : bkgd_sel) {
                                 case 0:
@@ -1732,6 +1739,7 @@ mach_accel_start(int cmd_type, int cpu_input, int count, uint32_t mix_dat, uint3
                             (dev->accel.cx <= clip_r) &&
                             (dev->accel.cy >= clip_t) &&
                             (dev->accel.cy <= clip_b)) {
+                            dev->subsys_stat |= 0x02;
                             mach->accel.clip_overrun = 0;
                             switch (mix ? frgd_sel : bkgd_sel) {
                                 case 0:
@@ -2027,6 +2035,7 @@ mach_accel_start(int cmd_type, int cpu_input, int count, uint32_t mix_dat, uint3
                     (dev->accel.dx <= clip_r) &&
                     (dev->accel.dy >= clip_t) &&
                     (dev->accel.dy <= clip_b)) {
+                    dev->subsys_stat |= 0x02;
                     switch (mix ? frgd_sel : bkgd_sel) {
                         case 0:
                             src_dat = dev->accel.bkgd_color;
@@ -2332,6 +2341,7 @@ mach_out(uint16_t addr, uint8_t val, void *priv)
         case 0x2ed:
             rs2 = !!(mach->accel.ext_ge_config & 0x1000);
             rs3 = !!(mach->accel.ext_ge_config & 0x2000);
+            mach_log("8514/A RS2=%d, RS3=%d, addr=%03x.\n", rs2, rs3, addr);
             if ((dev->local & 0xff) >= 0x02) {
                 if (mach->regs[0xb0] & 0x20) { /*ATI extended 8514/A mode.*/
                     mach_log("Extended 8514/A mode.\n");
@@ -2340,10 +2350,14 @@ mach_out(uint16_t addr, uint8_t val, void *priv)
                     svga_recalctimings(svga);
                     mach32_updatemapping(mach, svga);
                 }
-                if (mach->pci_bus && !mach->ramdac_type)
-                    ati68860_ramdac_out((addr & 0x03) | (rs2 << 2) | (rs3 << 3), val, svga->ramdac, svga);
-                else
-                    ati68875_ramdac_out(addr, rs2, rs3, val, svga->ramdac, svga);
+                if (dev->on)
+                    svga_out(addr, val, svga);
+                else {
+                    if (mach->pci_bus && !mach->ramdac_type)
+                        ati68860_ramdac_out((addr & 0x03) | (rs2 << 2) | (rs3 << 3), val, svga->ramdac, svga);
+                    else
+                        ati68875_ramdac_out(addr, rs2, rs3, val, svga->ramdac, svga);
+                }
             } else
                 svga_out(addr, val, svga);
             return;
@@ -2354,6 +2368,7 @@ mach_out(uint16_t addr, uint8_t val, void *priv)
         case 0x3C9:
             rs2 = !!(mach->regs[0xa0] & 0x20);
             rs3 = !!(mach->regs[0xa0] & 0x40);
+            mach_log("VGA RS2=%d, RS3=%d, addr=%03x.\n", rs2, rs3, addr);
             if ((dev->local & 0xff) >= 0x02) {
                 if (svga->attrregs[0x10] & 0x40) {
                     mach_log("VGA mode.\n");
@@ -2362,10 +2377,14 @@ mach_out(uint16_t addr, uint8_t val, void *priv)
                     svga_recalctimings(svga);
                     mach32_updatemapping(mach, svga);
                 }
-                if (mach->pci_bus && !mach->ramdac_type)
-                    ati68860_ramdac_out((addr & 0x03) | (rs2 << 2) | (rs3 << 3), val, svga->ramdac, svga);
-                else
-                    ati68875_ramdac_out(addr, rs2, rs3, val, svga->ramdac, svga);
+                if (dev->on)
+                    svga_out(addr, val, svga);
+                else {
+                    if (mach->pci_bus && !mach->ramdac_type)
+                        ati68860_ramdac_out((addr & 0x03) | (rs2 << 2) | (rs3 << 3), val, svga->ramdac, svga);
+                    else
+                        ati68875_ramdac_out(addr, rs2, rs3, val, svga->ramdac, svga);
+                }
             } else
                 svga_out(addr, val, svga);
             return;
@@ -2480,10 +2499,14 @@ mach_in(uint16_t addr, void *priv)
             rs2 = !!(mach->accel.ext_ge_config & 0x1000);
             rs3 = !!(mach->accel.ext_ge_config & 0x2000);
             if ((dev->local & 0xff) >= 0x02) {
-                if (mach->pci_bus && !mach->ramdac_type)
-                    temp = ati68860_ramdac_in((addr & 3) | (rs2 << 2) | (rs3 << 3), svga->ramdac, svga);
-                else
-                    temp = ati68875_ramdac_in(addr, rs2, rs3, svga->ramdac, svga);
+                if (dev->on)
+                    temp = svga_in(addr, svga);
+                else {
+                    if (mach->pci_bus && !mach->ramdac_type)
+                        temp = ati68860_ramdac_in((addr & 3) | (rs2 << 2) | (rs3 << 3), svga->ramdac, svga);
+                    else
+                        temp = ati68875_ramdac_in(addr, rs2, rs3, svga->ramdac, svga);
+                }
             } else
                 temp = svga_in(addr, svga);
             break;
@@ -3406,7 +3429,7 @@ mach_accel_out_fifo(mach_t *mach, svga_t *svga, ibm8514_t *dev, uint16_t port, u
 static void
 mach_accel_out_call(uint16_t port, uint8_t val, mach_t *mach, svga_t *svga, ibm8514_t *dev)
 {
-    if (port != 0x42e8 && port != 0x42e9)
+    if (port == 0x42e8 || port == 0x42e9)
         mach_log("[%04X:%08X]: Port CALL OUT=%04x, val=%02x.\n", CS, cpu_state.pc, port, val);
 
     switch (port) {
@@ -3421,7 +3444,7 @@ mach_accel_out_call(uint16_t port, uint8_t val, mach_t *mach, svga_t *svga, ibm8
             break;
         case 0x42e9:
             ibm8514_accel_out(port, val, svga, 2);
-            if ((val & 0xc0) == 0xc0) {
+            if ((val & 0xc0) == 0x80) {
                 dev->ext_fifo_idx = 0;
                 mach->force_busy = 0;
             }
@@ -3512,6 +3535,11 @@ mach_accel_out_call(uint16_t port, uint8_t val, mach_t *mach, svga_t *svga, ibm8
             svga_recalctimings(svga);
             break;
 
+        case 0x46e8:
+        case 0x46e9:
+            mach_log("0x%04x write: VGA subsystem enable add-on=%02x.\n", port, val);
+            break;
+
         case 0x4ae8:
             dev->accel.advfunc_cntl = val;
             dev->on = dev->accel.advfunc_cntl & 0x01;
@@ -4200,6 +4228,7 @@ mach_accel_in_call(uint16_t port, mach_t *mach, svga_t *svga, ibm8514_t *dev)
 
     switch (port) {
         case 0x2e8:
+        case 0x2e9:
         case 0x6e8:
         case 0x22e8:
         case 0x26e8:
@@ -4211,18 +4240,22 @@ mach_accel_in_call(uint16_t port, mach_t *mach, svga_t *svga, ibm8514_t *dev)
 
         case 0x42e8:
         case 0x42e9:
-            if (dev->vc == dev->v_syncstart)
+            if ((dev->subsys_cntl & 0x01) && !(dev->subsys_stat & 0x01) && (dev->vc == dev->dispend))
                 temp |= 0x01;
 
             if (mach->accel.cmd_type == -1) {
                 if (cmd == 6) {
-                    if ((dev->accel.dx >= clip_l) &&
+                    if ((dev->subsys_cntl & 0x02) &&
+                        !(dev->subsys_stat & 0x02) &&
+                        (dev->accel.dx >= clip_l) &&
                         (dev->accel.dx <= clip_r_ibm) &&
                         (dev->accel.dy >= clip_t) &&
                         (dev->accel.dy <= clip_b_ibm))
                         temp |= 0x02;
                 } else {
-                    if ((dev->accel.cx >= clip_l) &&
+                    if ((dev->subsys_cntl & 0x02) &&
+                        !(dev->subsys_stat & 0x02) &&
+                        (dev->accel.cx >= clip_l) &&
                         (dev->accel.cx <= clip_r_ibm) &&
                         (dev->accel.cy >= clip_t) &&
                         (dev->accel.cy <= clip_b_ibm))
@@ -4233,7 +4266,9 @@ mach_accel_in_call(uint16_t port, mach_t *mach, svga_t *svga, ibm8514_t *dev)
                     case 1:
                     case 2:
                     case 5:
-                        if ((dev->accel.dx >= clip_l) &&
+                        if ((dev->subsys_cntl & 0x02) &&
+                            !(dev->subsys_stat & 0x02) &&
+                            (dev->accel.dx >= clip_l) &&
                             (dev->accel.dx <= clip_r) &&
                             (dev->accel.dy >= clip_t) &&
                             (dev->accel.dy <= clip_b))
@@ -4241,7 +4276,9 @@ mach_accel_in_call(uint16_t port, mach_t *mach, svga_t *svga, ibm8514_t *dev)
                         break;
                     case 3:
                     case 4:
-                        if ((dev->accel.cx >= clip_l) &&
+                        if ((dev->subsys_cntl & 0x02) &&
+                            !(dev->subsys_stat & 0x02) &&
+                            (dev->accel.cx >= clip_l) &&
                             (dev->accel.cx <= clip_r) &&
                             (dev->accel.cy >= clip_t) &&
                             (dev->accel.cy <= clip_b))
@@ -4256,16 +4293,17 @@ mach_accel_in_call(uint16_t port, mach_t *mach, svga_t *svga, ibm8514_t *dev)
                 if ((!dev->force_busy && !dev->force_busy2) || !mach->force_busy)
                     temp |= 0x08;
             }
-            if (port & 1)
+            if (port & 1) {
                 temp = dev->vram_512k_8514 ? 0x00 : 0x80;
-            else {
+                temp |= (dev->subsys_cntl >> 8);
+            } else {
                 temp |= (dev->subsys_stat | (dev->vram_512k_8514 ? 0x00 : 0x80));
                 if (mach->accel.ext_ge_config & 0x08)
                     temp |= ((mach->accel.ext_ge_config & 0x07) << 4);
                 else
                     temp |= 0x20;
             }
-            mach_log("0x%04x read: Subsystem Status=%02x.\n", port, temp);
+            mach_log("0x%04x read: Subsystem Status=%02x, monitoralias=%02x.\n", port, temp, mach->accel.ext_ge_config & 0x07);
             break;
 
             /*ATI Mach8/32 specific registers*/
@@ -4402,7 +4440,8 @@ mach_accel_in_call(uint16_t port, mach_t *mach, svga_t *svga, ibm8514_t *dev)
         default:
             break;
     }
-    mach_log("[%04X:%08X]: Port NORMAL IN=%04x, temp=%04x.\n", CS, cpu_state.pc, port, temp);
+    if (port == 0x2ee8 || port == 0x2ee9 || port == 0x42e8 || port == 0x42e9)
+        mach_log("[%04X:%08X]: Port NORMAL IN=%04x, temp=%04x.\n", CS, cpu_state.pc, port, temp);
 
     return temp;
 }
@@ -5418,6 +5457,10 @@ mach32_updatemapping(mach_t *mach, svga_t *svga)
             mem_mapping_set_handler(&svga->mapping, mach32_read, mach32_readw, mach32_readl, mach32_write, mach32_writew, mach32_writel);
             mem_mapping_set_p(&svga->mapping, mach);
         } else {
+            if (!dev->on) {
+                memset(dev->vram, 0, dev->vram_size);
+                memset(dev->changedvram, 0, (dev->vram_size >> 12) + 1);
+            }
             mach_log("IBM compatible banked mapping.\n");
             mem_mapping_set_handler(&svga->mapping, svga_read, svga_readw, svga_readl, svga_write, svga_writew, svga_writel);
             mem_mapping_set_p(&svga->mapping, svga);
diff --git a/src/video/vid_chips_69000.c b/src/video/vid_chips_69000.c
index 012a16348..a387e99fa 100644
--- a/src/video/vid_chips_69000.c
+++ b/src/video/vid_chips_69000.c
@@ -27,6 +27,7 @@
 #include <86box/mem.h>
 #include <86box/rom.h>
 #include <86box/device.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/video.h>
 #include <86box/vid_svga.h>
diff --git a/src/video/vid_mga.c b/src/video/vid_mga.c
index f11983a20..00570bd9f 100644
--- a/src/video/vid_mga.c
+++ b/src/video/vid_mga.c
@@ -23,6 +23,7 @@
 #include <stdatomic.h>
 #include <86box/86box.h>
 #include <86box/io.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/mem.h>
 #include <86box/pci.h>
diff --git a/src/video/vid_nga.c b/src/video/vid_nga.c
index 3640e5106..ef1c6cd40 100644
--- a/src/video/vid_nga.c
+++ b/src/video/vid_nga.c
@@ -30,6 +30,7 @@
 #include <86box/io.h>
 #include <86box/video.h>
 #include <86box/86box.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/mem.h>
 #include <86box/pit.h>
diff --git a/src/video/vid_ogc.c b/src/video/vid_ogc.c
index 7fe5fbbeb..a49cd8a22 100644
--- a/src/video/vid_ogc.c
+++ b/src/video/vid_ogc.c
@@ -31,6 +31,7 @@
 #include <86box/io.h>
 #include <86box/video.h>
 #include <86box/86box.h>
+#include "cpu.h"
 #include <86box/timer.h>
 #include <86box/mem.h>
 #include <86box/pit.h>
diff --git a/src/video/vid_table.c b/src/video/vid_table.c
index ff3073c73..d1df2ba20 100644
--- a/src/video/vid_table.c
+++ b/src/video/vid_table.c
@@ -141,10 +141,10 @@ video_cards[] = {
     { .device = &chips_69000_device,                            .flags = VIDEO_FLAG_TYPE_NONE },
     { .device = &gd5430_pci_device,                             .flags = VIDEO_FLAG_TYPE_NONE },
     { .device = &gd5434_pci_device,                             .flags = VIDEO_FLAG_TYPE_NONE },
-    { .device = &gd5436_pci_device,                             .flags = VIDEO_FLAG_TYPE_SPECIAL },
+    { .device = &gd5436_pci_device,                             .flags = VIDEO_FLAG_TYPE_SECONDARY },
     { .device = &gd5440_pci_device,                             .flags = VIDEO_FLAG_TYPE_NONE },
-    { .device = &gd5446_pci_device,                             .flags = VIDEO_FLAG_TYPE_SPECIAL },
-    { .device = &gd5446_stb_pci_device,                         .flags = VIDEO_FLAG_TYPE_SPECIAL },
+    { .device = &gd5446_pci_device,                             .flags = VIDEO_FLAG_TYPE_SECONDARY },
+    { .device = &gd5446_stb_pci_device,                         .flags = VIDEO_FLAG_TYPE_SECONDARY },
     { .device = &gd5480_pci_device,                             .flags = VIDEO_FLAG_TYPE_NONE },
     { .device = &et4000w32p_videomagic_revb_pci_device,         .flags = VIDEO_FLAG_TYPE_NONE },
     { .device = &et4000w32p_revc_pci_device,                    .flags = VIDEO_FLAG_TYPE_NONE },