diff --git a/android/app/build.gradle.kts b/android/app/build.gradle.kts
index 4ff4e132ee..3f5cd42162 100644
--- a/android/app/build.gradle.kts
+++ b/android/app/build.gradle.kts
@@ -43,7 +43,11 @@ android {
           "-DXEMU_ANDROID_BUILD_ID=3",
           "-DXEMU_ENABLE_XISO_CONVERTER=ON",
           "-DCMAKE_C_FLAGS_DEBUG=-O2 -g0",
-          "-DCMAKE_CXX_FLAGS_DEBUG=-O2 -g0"
+          "-DCMAKE_CXX_FLAGS_DEBUG=-O2 -g0",
+          "-DCMAKE_C_FLAGS_RELEASE=-O3 -g0 -march=armv8.2-a -ffunction-sections -fdata-sections",
+          "-DCMAKE_CXX_FLAGS_RELEASE=-O3 -g0 -march=armv8.2-a -ffunction-sections -fdata-sections",
+          "-DCMAKE_EXE_LINKER_FLAGS_RELEASE=-Wl,--gc-sections",
+          "-DCMAKE_SHARED_LINKER_FLAGS_RELEASE=-Wl,--gc-sections"
         )
         cppFlags += listOf("-std=c++17", "-fexceptions", "-frtti")
       }
diff --git a/android/app/src/main/cpp/CMakeLists.txt b/android/app/src/main/cpp/CMakeLists.txt
index 9999514f54..8c70eb21a3 100644
--- a/android/app/src/main/cpp/CMakeLists.txt
+++ b/android/app/src/main/cpp/CMakeLists.txt
@@ -844,6 +844,7 @@ target_compile_definitions(xemu_core PRIVATE
 
 target_compile_options(xemu_core PRIVATE -g0
   $<$<OR:$<CONFIG:Release>,$<CONFIG:RelWithDebInfo>,$<CONFIG:MinSizeRel>>:-UNDEBUG>
+  $<$<CONFIG:Release>:-O3 -march=armv8.2-a -ffunction-sections -fdata-sections>
 )
 
 target_include_directories(xemu_core PRIVATE
@@ -913,6 +914,7 @@ endif()
 
 target_compile_options(xemu PRIVATE -fexceptions -frtti -g0
   $<$<OR:$<CONFIG:Release>,$<CONFIG:RelWithDebInfo>,$<CONFIG:MinSizeRel>>:-UNDEBUG>
+  $<$<CONFIG:Release>:-O3 -march=armv8.2-a -ffunction-sections -fdata-sections>
 )
 
 target_include_directories(xemu PRIVATE
diff --git a/hw/xbox/nv2a/pgraph/vk/draw.c b/hw/xbox/nv2a/pgraph/vk/draw.c
index 6db8de4c83..cb42be706b 100644
--- a/hw/xbox/nv2a/pgraph/vk/draw.c
+++ b/hw/xbox/nv2a/pgraph/vk/draw.c
@@ -22,6 +22,10 @@
 #include "renderer.h"
 #include "hw/xbox/nv2a/pgraph/prim_rewrite.h"
 #include <math.h>
+#ifdef __ANDROID__
+#include <SDL.h>
+#include <android/log.h>
+#endif
 
 void pgraph_vk_draw_begin(NV2AState *d)
 {
@@ -99,6 +103,22 @@ static bool pipeline_cache_entry_compare(Lru *lru, LruNode *node,
     return memcmp(&snode->key, key, sizeof(PipelineKey));
 }
 
+#ifdef __ANDROID__
+static char *get_pipeline_cache_path(PGRAPHVkState *r)
+{
+    char *pref_path = SDL_GetPrefPath("xemu", "xemu");
+    if (!pref_path) {
+        return NULL;
+    }
+    char *path = g_strdup_printf("%svk_pipeline_cache_%08x_%08x.bin",
+                                 pref_path,
+                                 r->device_props.deviceID,
+                                 r->device_props.driverVersion);
+    SDL_free(pref_path);
+    return path;
+}
+#endif
+
 static void init_pipeline_cache(PGRAPHState *pg)
 {
     PGRAPHVkState *r = pg->vk_renderer_state;
@@ -110,9 +130,33 @@ static void init_pipeline_cache(PGRAPHState *pg)
         .pInitialData = NULL,
         .pNext = NULL,
     };
+
+#ifdef __ANDROID__
+    g_autofree char *cache_path = get_pipeline_cache_path(r);
+    gchar *cache_data = NULL;
+    gsize cache_data_size = 0;
+    if (cache_path) {
+        GError *err = NULL;
+        if (g_file_get_contents(cache_path, &cache_data, &cache_data_size, &err)) {
+            cache_info.initialDataSize = (size_t)cache_data_size;
+            cache_info.pInitialData = (const void *)cache_data;
+            __android_log_print(ANDROID_LOG_INFO, "xemu-vk",
+                                "Loaded pipeline cache: %zu bytes", cache_data_size);
+        } else {
+            if (err) {
+                g_error_free(err);
+            }
+        }
+    }
+#endif
+
     VK_CHECK(vkCreatePipelineCache(r->device, &cache_info, NULL,
                                    &r->vk_pipeline_cache));
 
+#ifdef __ANDROID__
+    g_free(cache_data);
+#endif
+
     const size_t pipeline_cache_size = 2048;
     lru_init(&r->pipeline_cache);
     r->pipeline_cache_entries =
@@ -135,6 +179,34 @@ static void finalize_pipeline_cache(PGRAPHState *pg)
     g_free(r->pipeline_cache_entries);
     r->pipeline_cache_entries = NULL;
 
+#ifdef __ANDROID__
+    g_autofree char *cache_path = get_pipeline_cache_path(r);
+    if (cache_path) {
+        size_t data_size = 0;
+        VkResult res = vkGetPipelineCacheData(r->device, r->vk_pipeline_cache,
+                                              &data_size, NULL);
+        if (res == VK_SUCCESS && data_size > 0) {
+            g_autofree void *data = g_malloc(data_size);
+            res = vkGetPipelineCacheData(r->device, r->vk_pipeline_cache,
+                                         &data_size, data);
+            if (res == VK_SUCCESS) {
+                GError *err = NULL;
+                if (g_file_set_contents(cache_path, (const gchar *)data,
+                                        (gssize)data_size, &err)) {
+                    __android_log_print(ANDROID_LOG_INFO, "xemu-vk",
+                                        "Saved pipeline cache: %zu bytes", data_size);
+                } else {
+                    __android_log_print(ANDROID_LOG_WARN, "xemu-vk",
+                                        "Failed to save pipeline cache");
+                    if (err) {
+                        g_error_free(err);
+                    }
+                }
+            }
+        }
+    }
+#endif
+
     vkDestroyPipelineCache(r->device, r->vk_pipeline_cache, NULL);
 }
 
diff --git a/hw/xbox/nv2a/pgraph/vk/surface.c b/hw/xbox/nv2a/pgraph/vk/surface.c
index f10e78c55f..915e7d0cbd 100644
--- a/hw/xbox/nv2a/pgraph/vk/surface.c
+++ b/hw/xbox/nv2a/pgraph/vk/surface.c
@@ -498,15 +498,14 @@ void pgraph_vk_wait_for_surface_download(SurfaceBinding *surface)
     NV2AState *d = g_nv2a;
     bool require_download = qatomic_read(&surface->draw_dirty);
 
-#ifdef __ANDROID__
     /*
-     * Android Vulkan currently presents through the CPU/VGA fallback path.
-     * Force framebuffer surface download so fallback upload sees fresh pixels.
+     * Android presents through the CPU/VGA fallback path (no external memory
+     * interop). We rely on draw_dirty to gate downloads: it is set atomically
+     * whenever the Vulkan renderer draws to this surface and cleared after each
+     * successful download. Skipping the download when draw_dirty is false
+     * avoids a GPU stall on frames where the Xbox GPU has not redrawn the
+     * framebuffer (static screens, menus, load screens, etc.).
      */
-    if (!d->pgraph.vk_renderer_state->display.use_external_memory) {
-        require_download = true;
-    }
-#endif
 
     if (require_download) {
         qemu_mutex_lock(&d->pfifo.lock);
diff --git a/ui/xemu.c b/ui/xemu.c
index d94ef1ce77..2ad936f2b3 100644
--- a/ui/xemu.c
+++ b/ui/xemu.c
@@ -1381,7 +1381,7 @@ void xemu_android_display_loop(void)
             sdl2_poll_events(&sdl2_console[0]);
             bql_unlock();
             qemu_mutex_unlock_main_loop();
-            SDL_Delay(16);
+            SDL_Delay(100);
             continue;
         }
         sdl2_gl_refresh(&sdl2_console[0].dcl);
@@ -1570,7 +1570,7 @@ void sdl2_gl_refresh(DisplayChangeListener *dcl)
         sdl2_poll_events(scon);
         bql_unlock();
         qemu_mutex_unlock_main_loop();
-        SDL_Delay(16);
+        SDL_Delay(100);
         return;
     }
 #endif
@@ -1776,7 +1776,11 @@ void sdl2_gl_refresh(DisplayChangeListener *dcl)
     bql_unlock();
     qemu_mutex_unlock_main_loop();
 
+#ifdef __ANDROID__
+    glFlush();
+#else
     glFinish();
+#endif
     nv2a_release_framebuffer_surface();
 #ifdef __ANDROID__
     android_log_gl_error("refresh-finish");
@@ -1807,7 +1811,9 @@ void sdl2_gl_refresh(DisplayChangeListener *dcl)
     int64_t spin_acc = 0;
 #endif
 
-#ifndef _WIN32
+#ifdef __ANDROID__
+    const int64_t sleep_threshold = 500000;   // 0.5ms — Android CFS scheduler jitter is ~0.2ms
+#elif !defined(_WIN32)
     const int64_t sleep_threshold = 2000000;
 #else
     const int64_t sleep_threshold = 250000;