diff --git a/CMakeLists.txt b/CMakeLists.txt
index 70c11cf..d77ff84 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -111,7 +111,7 @@ add_library(aurora::shared ALIAS aurora_shared)
 target_compile_definitions(aurora_shared PUBLIC AURORA TARGET_PC)
 target_include_directories(aurora_shared PUBLIC include)
 target_link_libraries(aurora_shared PUBLIC ${AURORA_SDL3_TARGET} fmt::fmt xxhash)
-target_link_libraries(aurora_shared PRIVATE absl::btree absl::flat_hash_map)
+target_link_libraries(aurora_shared PRIVATE absl::btree absl::flat_hash_map TracyClient)
 
 if (AURORA_ENABLE_DVD)
   target_sources(aurora_shared PRIVATE lib/dolphin/dvd/dvd.cpp)
diff --git a/cmake/aurora_core.cmake b/cmake/aurora_core.cmake
index 3a9da55..9fc7bb2 100644
--- a/cmake/aurora_core.cmake
+++ b/cmake/aurora_core.cmake
@@ -9,7 +9,7 @@ add_library(aurora::core ALIAS aurora_core)
 target_compile_definitions(aurora_core PUBLIC AURORA TARGET_PC)
 target_include_directories(aurora_core PUBLIC include)
 target_link_libraries(aurora_core PUBLIC ${AURORA_SDL3_TARGET} fmt::fmt xxhash)
-target_link_libraries(aurora_core PRIVATE absl::btree absl::flat_hash_map sqlite3)
+target_link_libraries(aurora_core PRIVATE absl::btree absl::flat_hash_map sqlite3 TracyClient)
 if (AURORA_ENABLE_GX AND AURORA_CACHE_USE_ZSTD)
     target_compile_definitions(aurora_core PRIVATE AURORA_CACHE_USE_ZSTD)
     target_link_libraries(aurora_core PRIVATE libzstd_static)
diff --git a/cmake/aurora_gx.cmake b/cmake/aurora_gx.cmake
index 27bb716..12b8fa7 100644
--- a/cmake/aurora_gx.cmake
+++ b/cmake/aurora_gx.cmake
@@ -36,7 +36,7 @@ add_library(aurora_gx STATIC
 add_library(aurora::gx ALIAS aurora_gx)
 
 target_link_libraries(aurora_gx PUBLIC aurora::core xxhash)
-target_link_libraries(aurora_gx PRIVATE absl::btree absl::flat_hash_map sqlite3)
+target_link_libraries(aurora_gx PRIVATE absl::btree absl::flat_hash_map sqlite3 TracyClient)
 if (EMSCRIPTEN)
     target_link_options(aurora_gx PUBLIC -sUSE_WEBGPU=1 -sASYNCIFY -sEXIT_RUNTIME)
     target_compile_definitions(aurora_gx PRIVATE ENABLE_BACKEND_WEBGPU)
diff --git a/extern/CMakeLists.txt b/extern/CMakeLists.txt
index b713452..a59bda3 100644
--- a/extern/CMakeLists.txt
+++ b/extern/CMakeLists.txt
@@ -161,3 +161,23 @@ if (AURORA_ENABLE_GX)
     endif ()
   endif ()
 endif ()
+
+if (NOT TARGET TracyClient)
+  message(STATUS "aurora: Fetching Tracy")
+
+  # Default to off, users can enable the cache variable if they want to use it.
+  set(TRACY_ENABLE OFF CACHE BOOL "Enable Tracy profiling")
+
+  include(FetchContent)
+  FetchContent_Declare(
+          tracy
+          URL https://github.com/wolfpld/tracy/archive/refs/tags/v0.13.1.tar.gz
+          URL_HASH SHA256=d4efc50ebcb0bfcfdbba148995aeb75044c0d80f5d91223aebfaa8fa9e563d2b
+          DOWNLOAD_EXTRACT_TIMESTAMP TRUE
+  )
+  FetchContent_MakeAvailable(tracy)
+
+  if (NOT TARGET TracyClient)
+    message(FATAL_ERROR "Failed to make TracyClient available")
+  endif ()
+endif ()
diff --git a/lib/aurora.cpp b/lib/aurora.cpp
index fc7ea0a..adb7bd2 100644
--- a/lib/aurora.cpp
+++ b/lib/aurora.cpp
@@ -15,6 +15,8 @@
 #include <SDL3/SDL_filesystem.h>
 #include <magic_enum.hpp>
 
+#include "tracy/Tracy.hpp"
+
 namespace aurora {
 AuroraConfig g_config;
 uint32_t g_sdlCustomEventsStart;
@@ -167,6 +169,7 @@ void shutdown() noexcept {
 }
 
 const AuroraEvent* update() noexcept {
+  ZoneScoped;
   if (g_initialFrame) {
     g_initialFrame = false;
     input::initialize();
@@ -175,6 +178,7 @@ const AuroraEvent* update() noexcept {
 }
 
 bool begin_frame() noexcept {
+  ZoneScoped;
 #ifdef AURORA_ENABLE_GX
   if (!g_surface) {
     webgpu::refresh_surface(true);
@@ -205,6 +209,7 @@ bool begin_frame() noexcept {
     Log.error("Failed to get surface texture: {}", magic_enum::enum_name(surfaceTexture.status));
     return false;
   }
+
   imgui::new_frame(window::get_window_size());
   gfx::begin_frame();
 #endif
@@ -212,6 +217,7 @@ bool begin_frame() noexcept {
 }
 
 void end_frame() noexcept {
+  ZoneScoped;
 #ifdef AURORA_ENABLE_GX
   gx::fifo::drain();
   const auto encoderDescriptor = wgpu::CommandEncoderDescriptor{
@@ -258,6 +264,23 @@ void end_frame() noexcept {
     Log.warn("Surface present failed: {}", static_cast<int>(presentStatus));
   }
   g_currentView = {};
+
+  TracyPlotConfig("aurora: lastVertSize", tracy::PlotFormatType::Memory, false, true, 0);
+  TracyPlotConfig("aurora: lastUniformSize", tracy::PlotFormatType::Memory, false, true, 0);
+  TracyPlotConfig("aurora: lastIndexSize", tracy::PlotFormatType::Memory, false, true, 0);
+  TracyPlotConfig("aurora: lastStorageSize", tracy::PlotFormatType::Memory, false, true, 0);
+  TracyPlotConfig("aurora: lastTextureUploadSize", tracy::PlotFormatType::Memory, false, true, 0);
+
+  TracyPlot("aurora: queuedPipelines", static_cast<int64_t>(gfx::g_stats.queuedPipelines));
+  TracyPlot("aurora: createdPipelines", static_cast<int64_t>(gfx::g_stats.createdPipelines));
+  TracyPlot("aurora: drawCallCount", static_cast<int64_t>(gfx::g_stats.drawCallCount));
+  TracyPlot("aurora: mergedDrawCallCount", static_cast<int64_t>(gfx::g_stats.mergedDrawCallCount));
+  TracyPlot("aurora: lastVertSize", static_cast<int64_t>(gfx::g_stats.lastVertSize));
+  TracyPlot("aurora: lastUniformSize", static_cast<int64_t>(gfx::g_stats.lastUniformSize));
+  TracyPlot("aurora: lastIndexSize", static_cast<int64_t>(gfx::g_stats.lastIndexSize));
+  TracyPlot("aurora: lastStorageSize", static_cast<int64_t>(gfx::g_stats.lastStorageSize));
+  TracyPlot("aurora: lastTextureUploadSize", static_cast<int64_t>(gfx::g_stats.lastTextureUploadSize));
+
 #endif
 }
 } // namespace
diff --git a/lib/dolphin/gx/GXManage.cpp b/lib/dolphin/gx/GXManage.cpp
index f8114ba..475d6e2 100644
--- a/lib/dolphin/gx/GXManage.cpp
+++ b/lib/dolphin/gx/GXManage.cpp
@@ -281,10 +281,6 @@ void GXFlush() {
   if (__gx->dirtyState) {
     __GXSetDirtyState();
   }
-  for (u32 i = 32; i > 0; i--) {
-    GX_WRITE_U8(0);
-  }
-  // PPCSync();
 }
 
 void GXPixModeSync() {
diff --git a/lib/dolphin/gx/GXTexture.cpp b/lib/dolphin/gx/GXTexture.cpp
index 1c88874..c9ae65c 100644
--- a/lib/dolphin/gx/GXTexture.cpp
+++ b/lib/dolphin/gx/GXTexture.cpp
@@ -8,11 +8,14 @@
 
 #include <absl/container/flat_hash_map.h>
 
+#include "tracy/Tracy.hpp"
+
 using namespace aurora::gfx;
 
 extern "C" {
 void GXInitTexObj(GXTexObj* obj_, const void* data, u16 width, u16 height, GXTexFmt format, GXTexWrapMode wrapS,
                   GXTexWrapMode wrapT, GXBool mipmap) {
+  ZoneScopedS(3);
   if (reinterpret_cast<GXTexObj_*>(obj_)->ref)
     Log.fatal("Texture object already has a reference");
   memset(obj_, 0, sizeof(GXTexObj));
@@ -40,6 +43,7 @@ void GXInitTexObj(GXTexObj* obj_, const void* data, u16 width, u16 height, GXTex
 
 void GXInitTexObjCI(GXTexObj* obj_, const void* data, u16 width, u16 height, GXCITexFmt format, GXTexWrapMode wrapS,
                     GXTexWrapMode wrapT, GXBool mipmap, u32 tlut) {
+  ZoneScopedS(3);
   if (reinterpret_cast<GXTexObj_*>(obj_)->ref)
     Log.fatal("Texture object already has a reference");
   memset(obj_, 0, sizeof(GXTexObj));
diff --git a/lib/gfx/clear.cpp b/lib/gfx/clear.cpp
index 2799a5c..d0b4c78 100644
--- a/lib/gfx/clear.cpp
+++ b/lib/gfx/clear.cpp
@@ -1,6 +1,7 @@
 #include "clear.hpp"
 
 #include "../webgpu/gpu.hpp"
+#include "tracy/Tracy.hpp"
 
 namespace {
 wgpu::ColorWriteMask clear_write_mask(bool clearColor, bool clearAlpha) {
@@ -21,6 +22,7 @@ using webgpu::g_device;
 using webgpu::g_graphicsConfig;
 
 wgpu::RenderPipeline create_pipeline(const PipelineConfig& config) {
+  ZoneScoped;
   wgpu::ShaderSourceWGSL sourceDescriptor{};
   sourceDescriptor.code = R"""(
 struct VertexOutput {
diff --git a/lib/gfx/common.cpp b/lib/gfx/common.cpp
index faa8036..c4106b6 100644
--- a/lib/gfx/common.cpp
+++ b/lib/gfx/common.cpp
@@ -16,6 +16,8 @@
 #include <absl/container/flat_hash_map.h>
 #include <magic_enum.hpp>
 
+#include "tracy/Tracy.hpp"
+
 namespace aurora::gfx {
 static Module Log("aurora::gfx");
 
@@ -72,11 +74,11 @@ namespace aurora {
 // we create specialized methods to handle them. Note that these are highly dependent on
 // the structure definition, which could easily change with Dawn updates.
 template <>
-inline HashType xxh3_hash(const wgpu::BindGroupDescriptor& input, HashType seed) {
-  constexpr auto offset = offsetof(wgpu::BindGroupDescriptor, layout); // skip nextInChain, label
+inline HashType xxh3_hash(const WGPUBindGroupDescriptor& input, HashType seed) {
+  constexpr auto offset = offsetof(WGPUBindGroupDescriptor, layout); // skip nextInChain, label
   const auto hash = xxh3_hash_s(reinterpret_cast<const u8*>(&input) + offset,
-                                sizeof(wgpu::BindGroupDescriptor) - offset - sizeof(void*) /* skip entries */, seed);
-  return xxh3_hash_s(input.entries, sizeof(wgpu::BindGroupEntry) * input.entryCount, hash);
+                                sizeof(WGPUBindGroupDescriptor) - offset - sizeof(void*) /* skip entries */, seed);
+  return xxh3_hash_s(input.entries, sizeof(WGPUBindGroupEntry) * input.entryCount, hash);
 }
 template <>
 inline HashType xxh3_hash(const wgpu::SamplerDescriptor& input, HashType seed) {
@@ -222,7 +224,7 @@ void push_draw_command(clear::DrawData data) {
 }
 
 template <>
-PipelineRef pipeline_ref(clear::PipelineConfig config) {
+PipelineRef pipeline_ref(const clear::PipelineConfig& config) {
   return find_pipeline(ShaderType::Clear, config, [=] { return create_pipeline(config); });
 }
 
@@ -349,6 +351,7 @@ static OffscreenCacheEntry get_offscreen_textures(uint32_t width, uint32_t heigh
 }
 
 void begin_offscreen(uint32_t width, uint32_t height) {
+  ZoneScoped;
   CHECK(g_currentRenderPass != UINT32_MAX, "begin_offscreen called outside of a frame");
 
   // If the current EFB pass has no resolve target, its output is unobservable.
@@ -391,6 +394,7 @@ void begin_offscreen(uint32_t width, uint32_t height) {
 }
 
 void end_offscreen() {
+  ZoneScoped;
   CHECK(g_inOffscreen, "end_offscreen called without begin_offscreen");
 
   g_inOffscreen = false;
@@ -419,7 +423,7 @@ void push_draw_command(gx::DrawData data) {
 }
 
 template <>
-PipelineRef pipeline_ref(gx::PipelineConfig config) {
+PipelineRef pipeline_ref(const gx::PipelineConfig& config) {
   return find_pipeline(ShaderType::GX, config, [=] { return create_pipeline(config); });
 }
 
@@ -504,8 +508,12 @@ void map_staging_buffer() {
 }
 
 void begin_frame() {
-  while (!bufferMapped) {
-    g_instance.ProcessEvents();
+  ZoneScoped;
+  {
+    ZoneScopedN("Wait for buffer map");
+    while (!bufferMapped) {
+      g_instance.ProcessEvents();
+    }
   }
   size_t bufferOffset = 0;
   const auto& stagingBuf = g_stagingBuffers[currentStagingBuffer];
@@ -537,13 +545,14 @@ void begin_frame() {
 }
 
 void end_frame(const wgpu::CommandEncoder& cmd) {
+  ZoneScoped;
   ASSERT(!g_inOffscreen, "end_frame called while offscreen rendering is active");
   uint64_t bufferOffset = 0;
   const auto writeBuffer = [&](ByteBuffer& buf, wgpu::Buffer& out, uint64_t size, std::string_view label) {
     const auto writeSize = buf.size(); // Only need to copy this many bytes
     if (writeSize > 0) {
-      cmd.CopyBufferToBuffer(g_stagingBuffers[currentStagingBuffer], bufferOffset, out, 0, ALIGN(writeSize, 4));
-      buf.clear();
+      cmd.CopyBufferToBuffer(g_stagingBuffers[currentStagingBuffer], bufferOffset, out, 0, AURORA_ALIGN(writeSize, 4));
+      buf.release();
     }
     bufferOffset += size;
     return writeSize;
@@ -561,7 +570,7 @@ void end_frame(const wgpu::CommandEncoder& cmd) {
           .layout =
               wgpu::TexelCopyBufferLayout{
                   .offset = item.layout.offset + bufferOffset,
-                  .bytesPerRow = ALIGN(item.layout.bytesPerRow, 256),
+                  .bytesPerRow = AURORA_ALIGN(item.layout.bytesPerRow, 256),
                   .rowsPerImage = item.layout.rowsPerImage,
               },
           .buffer = g_stagingBuffers[currentStagingBuffer],
@@ -569,7 +578,7 @@ void end_frame(const wgpu::CommandEncoder& cmd) {
       cmd.CopyBufferToTexture(&buf, &item.tex, &item.size);
     }
     g_textureUploads.clear();
-    g_textureUpload.clear();
+    g_textureUpload.release();
   }
   currentStagingBuffer = (currentStagingBuffer + 1) % g_stagingBuffers.size();
   map_staging_buffer();
@@ -584,6 +593,7 @@ void end_frame(const wgpu::CommandEncoder& cmd) {
 uint32_t current_frame() noexcept { return g_frameIndex; }
 
 void render(wgpu::CommandEncoder& cmd) {
+  ZoneScoped;
   for (u32 i = 0; i < g_renderPasses.size(); ++i) {
     const auto& passInfo = g_renderPasses[i];
     for (const auto& conv : passInfo.paletteConvs) {
@@ -807,7 +817,7 @@ Range push_storage(const uint8_t* data, size_t length) {
 }
 Range push_texture_data(const uint8_t* data, size_t length, u32 bytesPerRow, u32 rowsPerImage) {
   // For CopyBufferToTexture, we need an alignment of 256 per row (see Dawn kTextureBytesPerRowAlignment)
-  const auto copyBytesPerRow = ALIGN(bytesPerRow, 256);
+  const auto copyBytesPerRow = AURORA_ALIGN(bytesPerRow, 256);
   const auto range = map(g_textureUpload, copyBytesPerRow * rowsPerImage, 0);
   u8* dst = g_textureUpload.data() + range.offset;
   for (u32 i = 0; i < rowsPerImage; ++i) {
@@ -835,21 +845,21 @@ std::pair<ByteBuffer, Range> map_storage(size_t length) {
 }
 
 // TODO: should we avoid caching bind groups altogether?
-BindGroupRef bind_group_ref(const wgpu::BindGroupDescriptor& descriptor) {
+BindGroupRef bind_group_ref(const WGPUBindGroupDescriptor& descriptor) {
 #ifdef EMSCRIPTEN
-  const auto bg = g_device.CreateBindGroup(&descriptor);
+  const auto bg = wgpuDeviceCreateBindGroup(g_device.Get(), &descriptor);
   BindGroupRef id = reinterpret_cast<BindGroupRef>(bg.Get());
   g_cachedBindGroups.try_emplace(id, bg);
 #else
   const auto id = xxh3_hash(descriptor);
   if (!g_cachedBindGroups.contains(id)) {
-    g_cachedBindGroups.try_emplace(id, g_device.CreateBindGroup(&descriptor));
+    g_cachedBindGroups.try_emplace(id, wgpuDeviceCreateBindGroup(g_device.Get(), &descriptor));
   }
 #endif
   return id;
 }
 
-wgpu::BindGroup find_bind_group(BindGroupRef id) {
+wgpu::BindGroup& find_bind_group(BindGroupRef id) {
 #ifdef EMSCRIPTEN
   return g_cachedBindGroups[id];
 #else
@@ -859,7 +869,7 @@ wgpu::BindGroup find_bind_group(BindGroupRef id) {
 #endif
 }
 
-wgpu::Sampler sampler_ref(const wgpu::SamplerDescriptor& descriptor) {
+wgpu::Sampler& sampler_ref(const wgpu::SamplerDescriptor& descriptor) {
   const auto id = xxh3_hash(descriptor);
   auto it = g_cachedSamplers.find(id);
   if (it == g_cachedSamplers.end()) {
@@ -868,7 +878,7 @@ wgpu::Sampler sampler_ref(const wgpu::SamplerDescriptor& descriptor) {
   return it->second;
 }
 
-uint32_t align_uniform(uint32_t value) { return ALIGN(value, g_cachedLimits.minUniformBufferOffsetAlignment); }
+uint32_t align_uniform(uint32_t value) { return AURORA_ALIGN(value, g_cachedLimits.minUniformBufferOffsetAlignment); }
 
 void insert_debug_marker(std::string label) {
 #if defined(AURORA_GFX_DEBUG_GROUPS)
diff --git a/lib/gfx/common.hpp b/lib/gfx/common.hpp
index 3516e15..a10f975 100644
--- a/lib/gfx/common.hpp
+++ b/lib/gfx/common.hpp
@@ -109,7 +109,7 @@ public:
     m_length += size;
   }
 
-  void clear() {
+  void release() {
     if (m_data != nullptr && m_owned) {
       free(m_data);
     }
@@ -119,6 +119,10 @@ public:
     m_owned = true;
   }
 
+  void clear() {
+    m_length = 0;
+  }
+
   void reserve_extra(size_t size) { resize(m_length + size, true); }
 
 private:
@@ -141,6 +145,10 @@ private:
       if (!m_owned) {
         abort();
       }
+      // Exponential expansion to avoid O(n^2) time complexity.
+      if (size < m_capacity * 2) {
+        size = m_capacity * 2;
+      }
       m_data = static_cast<uint8_t*>(realloc(m_data, size));
       if (zeroed) {
         memset(m_data + m_capacity, 0, size - m_capacity);
@@ -272,13 +280,13 @@ template <typename DrawData>
 DrawData* get_last_draw_command();
 
 template <typename PipelineConfig>
-PipelineRef pipeline_ref(PipelineConfig config);
+PipelineRef pipeline_ref(const PipelineConfig& config);
 bool bind_pipeline(PipelineRef ref, const wgpu::RenderPassEncoder& pass);
 
-BindGroupRef bind_group_ref(const wgpu::BindGroupDescriptor& descriptor);
-wgpu::BindGroup find_bind_group(BindGroupRef id);
+BindGroupRef bind_group_ref(const WGPUBindGroupDescriptor& descriptor);
+wgpu::BindGroup& find_bind_group(BindGroupRef id);
 
-wgpu::Sampler sampler_ref(const wgpu::SamplerDescriptor& descriptor);
+wgpu::Sampler& sampler_ref(const wgpu::SamplerDescriptor& descriptor);
 
 uint32_t align_uniform(uint32_t value);
 
diff --git a/lib/gfx/texture.cpp b/lib/gfx/texture.cpp
index c1c2d86..8336e39 100644
--- a/lib/gfx/texture.cpp
+++ b/lib/gfx/texture.cpp
@@ -16,6 +16,8 @@
 #include <magic_enum.hpp>
 #include <webgpu/webgpu_cpp.h>
 
+#include "tracy/Tracy.hpp"
+
 namespace aurora::gfx {
 using webgpu::g_device;
 using webgpu::g_queue;
@@ -141,6 +143,7 @@ static bool setup_swizzle(wgpu::TextureComponentSwizzleDescriptor& swizzle, u32
 
 TextureHandle new_dynamic_texture_2d(uint32_t width, uint32_t height, uint32_t mips, u32 gxFormat,
                                      const char* label) noexcept {
+  ZoneScopedS(3);
   const auto wgpuFormat = to_wgpu(gxFormat);
   const wgpu::Extent3D size{
       .width = width,
diff --git a/lib/gx/command_processor.cpp b/lib/gx/command_processor.cpp
index a6fc9e5..0a8e51d 100644
--- a/lib/gx/command_processor.cpp
+++ b/lib/gx/command_processor.cpp
@@ -14,6 +14,8 @@
 #include "dolphin/gx/GXAurora.h"
 #include <limits>
 
+#include "tracy/Tracy.hpp"
+
 namespace aurora::gx::fifo {
 static Module Log("aurora::gx::fifo");
 
@@ -104,20 +106,35 @@ static constexpr u8 CP_OPCODE_MASK = GX_OPCODE_MASK;
 static constexpr u8 CP_VAT_MASK = GX_VAT_MASK;
 
 // Read helpers for big/little endian
+#if _MSC_VER
+template<typename T>
+__forceinline // Yes, this was necessary.
+inline T unaligned_load(const T* ptr) {
+  return *static_cast<const __unaligned T*>(ptr);
+}
+#else
+template<typename T>
+inline T unaligned_load(const T* ptr) {
+  T copy;
+  memcpy(&copy, ptr, sizeof(T));
+  return copy;
+}
+#endif
+
 static inline u16 read_u16(const u8* ptr, bool bigEndian) {
+  const u16 val = unaligned_load(reinterpret_cast<const u16*>(ptr));
   if (bigEndian) {
-    return static_cast<u16>(ptr[0] << 8 | ptr[1]);
+    return bswap(val);
   }
-  return static_cast<u16>(ptr[1] << 8 | ptr[0]);
+  return val;
 }
 
 static inline u32 read_u32(const u8* ptr, bool bigEndian) {
+  const u32 val = unaligned_load(reinterpret_cast<const u32*>(ptr));
   if (bigEndian) {
-    return static_cast<u32>(ptr[0]) << 24 | static_cast<u32>(ptr[1]) << 16 | static_cast<u32>(ptr[2]) << 8 |
-           static_cast<u32>(ptr[3]);
+    return bswap(val);
   }
-  return static_cast<u32>(ptr[3]) << 24 | static_cast<u32>(ptr[2]) << 16 | static_cast<u32>(ptr[1]) << 8 |
-         static_cast<u32>(ptr[0]);
+  return val;
 }
 
 static u32 bp_get(u32 reg, u32 size, u32 shift);
@@ -317,6 +334,7 @@ static void handle_draw(u8 cmd, const u8* data, u32& pos, u32 size, bool bigEndi
 static void handle_aurora(const u8* data, u32& pos, u32 size, bool bigEndian);
 
 void process(const u8* data, u32 size, bool bigEndian) {
+  ZoneScoped;
   u32 pos = 0;
 
   while (pos < size) {
@@ -354,6 +372,7 @@ void process(const u8* data, u32 size, bool bigEndian) {
     case CP_CMD_LOAD_INDX_B:
     case CP_CMD_LOAD_INDX_C:
     case CP_CMD_LOAD_INDX_D: {
+      ZoneScopedN("LOAD_INDX");
       // Indexed XF load: 4 bytes of data
       CHECK(pos + 4 <= size, "indexed XF read overrun");
       u32 arrayType = GX_POS_MTX_ARRAY + (opcode - (CP_CMD_LOAD_INDX_A / 0x08));
@@ -426,10 +445,11 @@ void process(const u8* data, u32 size, bool bigEndian) {
 }
 
 // Helper to extract bit fields from a 32-bit register
-static u32 bp_get(u32 reg, u32 size, u32 shift) { return reg >> shift & (1u << size) - 1; }
+inline static u32 bp_get(u32 reg, u32 size, u32 shift) { return reg >> shift & (1u << size) - 1; }
 
 // BP register handler - decodes BP (RAS/pixel engine) register writes and updates g_gxState
 static void handle_bp(u32 value, bool bigEndian) {
+  ZoneScoped;
   u32 regId = (value >> 24) & 0xFF;
   // Mask off the register ID from the value for field extraction
   // (the regId is stored in bits 24-31, data is in bits 0-23)
@@ -1012,6 +1032,7 @@ static void handle_bp(u32 value, bool bigEndian) {
 
 // CP register handler - decodes CP register writes and updates g_gxState
 static void handle_cp(u8 addr, u32 value, bool bigEndian) {
+  ZoneScoped;
   switch (addr) {
   // VCD low (0x50)
   case 0x50: {
@@ -1030,6 +1051,7 @@ static void handle_cp(u8 addr, u32 value, bool bigEndian) {
     vd[GX_VA_CLR0] = static_cast<GXAttrType>(bp_get(value, 2, 13));
     vd[GX_VA_CLR1] = static_cast<GXAttrType>(bp_get(value, 2, 15));
     g_gxState.stateDirty = true;
+    g_gxState.clearVtxSizeCache();
     break;
   }
 
@@ -1045,6 +1067,7 @@ static void handle_cp(u8 addr, u32 value, bool bigEndian) {
     vd[GX_VA_TEX6] = static_cast<GXAttrType>(bp_get(value, 2, 12));
     vd[GX_VA_TEX7] = static_cast<GXAttrType>(bp_get(value, 2, 14));
     g_gxState.stateDirty = true;
+    g_gxState.clearVtxSizeCache();
     break;
   }
 
@@ -1086,6 +1109,7 @@ static void handle_cp(u8 addr, u32 value, bool bigEndian) {
       vf.attrs[GX_VA_TEX0].type = static_cast<GXCompType>(bp_get(value, 3, 22));
       vf.attrs[GX_VA_TEX0].frac = static_cast<u8>(bp_get(value, 5, 25));
       g_gxState.stateDirty = true;
+      g_gxState.clearVtxSizeCache();
     }
     // VAT B registers (0x80-0x87)
     else if (addr >= 0x80 && addr <= 0x87) {
@@ -1104,6 +1128,7 @@ static void handle_cp(u8 addr, u32 value, bool bigEndian) {
       vf.attrs[GX_VA_TEX4].type = static_cast<GXCompType>(bp_get(value, 3, 28));
       // TEX4 frac is in VAT C
       g_gxState.stateDirty = true;
+      g_gxState.clearVtxSizeCache();
     }
     // VAT C registers (0x90-0x97)
     else if (addr >= 0x90 && addr <= 0x97) {
@@ -1120,6 +1145,7 @@ static void handle_cp(u8 addr, u32 value, bool bigEndian) {
       vf.attrs[GX_VA_TEX7].type = static_cast<GXCompType>(bp_get(value, 3, 24));
       vf.attrs[GX_VA_TEX7].frac = static_cast<u8>(bp_get(value, 5, 27));
       g_gxState.stateDirty = true;
+      g_gxState.clearVtxSizeCache();
     }
     // Array base addresses (0xA0-0xAF)
     else if (addr >= 0xA0 && addr <= 0xAF) {
@@ -1143,6 +1169,7 @@ static void handle_cp(u8 addr, u32 value, bool bigEndian) {
 
 // XF register handler - decodes XF (transform unit) register writes and updates g_gxState
 static void handle_xf(const u8* data, u32& pos, u32 size, bool bigEndian) {
+  ZoneScoped;
   CHECK(pos + 4 <= size, "XF header read overrun");
   u32 header = read_u32(data + pos, bigEndian);
   pos += 4;
@@ -1362,16 +1389,24 @@ static void handle_xf(const u8* data, u32& pos, u32 size, bool bigEndian) {
   pos += dataBytes;
 }
 
+static void handle_draw_overrun [[noreturn]] (u32 totalVtxBytes, const u8* data, const u32& pos, u32 size) {
+  // Hex dump around the draw command for debugging
+  u32 cmdPos = pos - 2 - 1; // opcode byte position (before vtxCount and pos++)
+  u32 dumpStart = (cmdPos > 16) ? cmdPos - 16 : 0;
+  u32 dumpEnd = (cmdPos + 32 < size) ? cmdPos + 32 : size;
+  std::string hex;
+  for (u32 i = dumpStart; i < dumpEnd; i++) {
+    if (i == cmdPos)
+      hex += fmt::format("[{:02x}]", data[i]);
+    else
+      hex += fmt::format(" {:02x}", data[i]);
+  }
+  Log.error("  hex dump around draw cmd (pos {}-{}):{}", dumpStart, dumpEnd - 1, hex);
+  FATAL("draw vertex data overrun: need {} bytes at pos {}, have {}", totalVtxBytes, pos, size);
+}
+
 // Draw command handler - parses vertices inline and caches results
-static void handle_draw(u8 cmd, const u8* data, u32& pos, u32 size, bool bigEndian) {
-  u8 opcode = cmd & CP_OPCODE_MASK;
-  GXVtxFmt fmt = static_cast<GXVtxFmt>(cmd & CP_VAT_MASK);
-  GXPrimitive prim = static_cast<GXPrimitive>(opcode);
-
-  CHECK(pos + 2 <= size, "draw vtxCount read overrun");
-  u16 vtxCount = read_u16(data + pos, bigEndian);
-  pos += 2;
-
+static u32 calculate_last_vtx_size(GXVtxFmt fmt) {
   u32 vtxSize = 0;
   const auto& vtxFmt = g_gxState.vtxFmts[fmt];
   for (int i = GX_VA_PNMTXIDX; i <= GX_VA_TEX7; ++i) {
@@ -1392,38 +1427,53 @@ static void handle_draw(u8 cmd, const u8* data, u32& pos, u32 size, bool bigEndi
       break;
     }
   }
+
+  g_gxState.lastVtxFmt = fmt;
+  g_gxState.lastVtxSize = vtxSize;
+
+  return vtxSize;
+}
+
+static void handle_draw_unmerged(GXPrimitive prim, GXVtxFmt fmt, u16 vtxCount, gfx::Range vertRange);
+
+// Draw command handler - parses vertices inline and caches results
+static ByteBuffer handle_draw_idx_buf;
+
+static void handle_draw(u8 cmd, const u8* data, u32& pos, u32 size, bool bigEndian) {
+  ZoneScoped;
+  u8 opcode = cmd & CP_OPCODE_MASK;
+  GXVtxFmt fmt = static_cast<GXVtxFmt>(cmd & CP_VAT_MASK);
+  GXPrimitive prim = static_cast<GXPrimitive>(opcode);
+
+  CHECK(pos + 2 <= size, "draw vtxCount read overrun");
+  u16 vtxCount = read_u16(data + pos, bigEndian);
+  pos += 2;
+
+  u32 vtxSize;
+  if (g_gxState.lastVtxFmt == fmt) LIKELY {
+    vtxSize = g_gxState.lastVtxSize;
+  } else UNLIKELY {
+    vtxSize = calculate_last_vtx_size(fmt);
+  }
+
   u32 totalVtxBytes = vtxCount * vtxSize;
-  if (pos + totalVtxBytes > size) {
-    // Hex dump around the draw command for debugging
-    u32 cmdPos = pos - 2 - 1; // opcode byte position (before vtxCount and pos++)
-    u32 dumpStart = (cmdPos > 16) ? cmdPos - 16 : 0;
-    u32 dumpEnd = (cmdPos + 32 < size) ? cmdPos + 32 : size;
-    std::string hex;
-    for (u32 i = dumpStart; i < dumpEnd; i++) {
-      if (i == cmdPos)
-        hex += fmt::format("[{:02x}]", data[i]);
-      else
-        hex += fmt::format(" {:02x}", data[i]);
-    }
-    Log.error("  hex dump around draw cmd (pos {}-{}):{}", dumpStart, dumpEnd - 1, hex);
-    FATAL("draw vertex data overrun: need {} bytes at pos {}, have {}", totalVtxBytes, pos, size);
+  if (pos + totalVtxBytes > size) UNLIKELY {
+    handle_draw_overrun(totalVtxBytes, data, pos, size);
   }
 
   // Push raw vertex data to buffer
   gfx::Range vertRange = gfx::push_verts(data + pos, totalVtxBytes);
   pos += totalVtxBytes;
 
-  u32 numIndices = 0;
-  gfx::Range idxRange;
   // Try to merge with previous draw call
-  if (!g_gxState.stateDirty) {
+  if (!g_gxState.stateDirty) LIKELY {
     auto* lastDraw = gfx::get_last_draw_command<DrawData>();
     // Only if the previous draw call was a single instance draw (no lines/points handling)
     if (lastDraw != nullptr && prim != GX_LINES && prim != GX_LINESTRIP && prim != GX_POINTS &&
-        lastDraw->instanceCount == 1) {
-      ByteBuffer idxBuf;
-      numIndices = prepare_idx_buffer(idxBuf, prim, lastDraw->vtxCount, vtxCount);
-      idxRange = gfx::push_indices(idxBuf.data(), idxBuf.size());
+        lastDraw->instanceCount == 1) LIKELY {
+      u32 numIndices = prepare_idx_buffer(handle_draw_idx_buf, prim, lastDraw->vtxCount, vtxCount);
+      gfx::Range idxRange = gfx::push_indices(handle_draw_idx_buf.data(), handle_draw_idx_buf.size());
+      handle_draw_idx_buf.clear();
       CHECK(lastDraw->vertRange.offset + lastDraw->vertRange.size == vertRange.offset,
             "Non-consecutive vertex ranges ({} < {})", lastDraw->vertRange.offset + lastDraw->vertRange.size,
             vertRange.offset);
@@ -1439,10 +1489,22 @@ static void handle_draw(u8 cmd, const u8* data, u32& pos, u32 size, bool bigEndi
     }
   }
 
+  handle_draw_unmerged(prim, fmt, vtxCount, vertRange);
+}
+
+static ByteBuffer handle_draw_unmerged_idxBuf;
+
+static void handle_draw_unmerged(GXPrimitive prim, GXVtxFmt fmt, u16 vtxCount, gfx::Range vertRange) {
+  ZoneScoped;
+  u32 numIndices = 0;
+  gfx::Range idxRange;
+
   {
     ByteBuffer idxBuf;
-    numIndices = prepare_idx_buffer(idxBuf, prim, 0, vtxCount);
-    idxRange = gfx::push_indices(idxBuf.data(), idxBuf.size());
+    auto& realBuf = vtxCount < 1000 ? handle_draw_unmerged_idxBuf : idxBuf;
+    numIndices = prepare_idx_buffer(realBuf, prim, 0, vtxCount);
+    idxRange = gfx::push_indices(realBuf.data(), realBuf.size());
+    realBuf.clear();
   }
 
   // Build pipeline, bind groups, and push draw command
@@ -1501,6 +1563,7 @@ std::string read_string(const u8* data, u32& pos, u32 size, bool bigEndian) {
 }
 
 void handle_aurora(const u8* data, u32& pos, u32 size, bool bigEndian) {
+  ZoneScoped;
   CHECK(pos + 2 <= size, "Aurora cmd read overrun");
   u16 subCmd = read_u16(data + pos, bigEndian);
   pos += 2;
diff --git a/lib/gx/gx.cpp b/lib/gx/gx.cpp
index 46c94f4..bc81678 100644
--- a/lib/gx/gx.cpp
+++ b/lib/gx/gx.cpp
@@ -12,6 +12,8 @@
 #include <cfloat>
 #include <mutex>
 
+#include "tracy/Tracy.hpp"
+
 static aurora::Module Log("aurora::gx");
 
 namespace aurora::gx {
@@ -186,6 +188,7 @@ static inline wgpu::PrimitiveState to_primitive_state(GXCullMode gx_cullMode) {
 
 wgpu::RenderPipeline build_pipeline(const PipelineConfig& config, ArrayRef<wgpu::VertexBufferLayout> vtxBuffers,
                                     wgpu::ShaderModule shader, const char* label) noexcept {
+  ZoneScoped;
   const wgpu::DepthStencilState depthStencil{
       .format = g_graphicsConfig.depthFormat,
       .depthWriteEnabled = config.depthCompare && config.depthUpdate,
@@ -439,11 +442,13 @@ GXBindGroups build_bind_groups(const ShaderInfo& info, const ShaderConfig& confi
                                const BindGroupRanges& ranges) noexcept {
   const auto layouts = build_bind_group_layouts(config);
 
-  std::array<wgpu::BindGroupEntry, MaxIndexAttr + 3> uniformEntries;
+  // Using C WGPU types instead of C++ wrappers to avoid destructor overhead
+  std::array<WGPUBindGroupEntry, MaxIndexAttr + 3> uniformEntries{};
   uniformEntries[0].binding = 0;
-  uniformEntries[0].buffer = gfx::g_vertexBuffer;
+  uniformEntries[0].buffer = gfx::g_vertexBuffer.Get();
+  uniformEntries[0].size = WGPU_WHOLE_SIZE;
   uniformEntries[1].binding = 1;
-  uniformEntries[1].buffer = gfx::g_uniformBuffer;
+  uniformEntries[1].buffer = gfx::g_uniformBuffer.Get();
   uniformEntries[1].size = info.uniformSize;
   u32 uniformBindIdx = 2;
   for (u32 i = 0; i < MaxIndexAttr; ++i) {
@@ -451,46 +456,46 @@ GXBindGroups build_bind_groups(const ShaderInfo& info, const ShaderConfig& confi
     if (range.size <= 0) {
       continue;
     }
-    wgpu::BindGroupEntry& entry = uniformEntries[uniformBindIdx];
+    WGPUBindGroupEntry& entry = uniformEntries[uniformBindIdx];
     entry.binding = uniformBindIdx;
-    entry.buffer = gfx::g_storageBuffer;
+    entry.buffer = gfx::g_storageBuffer.Get();
     entry.size = range.size;
     ++uniformBindIdx;
   }
 
-  std::array<wgpu::BindGroupEntry, MaxTextures> samplerEntries;
-  std::array<wgpu::BindGroupEntry, MaxTextures> textureEntries;
+  std::array<WGPUBindGroupEntry, MaxTextures> samplerEntries{};
+  std::array<WGPUBindGroupEntry, MaxTextures> textureEntries{};
   u32 textureCount = 0;
   for (u32 i = 0; i < MaxTextures; ++i) {
     const auto& tex = g_gxState.textures[i];
-    wgpu::BindGroupEntry& samplerEntry = samplerEntries[textureCount];
-    wgpu::BindGroupEntry& textureEntry = textureEntries[textureCount];
+    WGPUBindGroupEntry& samplerEntry = samplerEntries[textureCount];
+    WGPUBindGroupEntry& textureEntry = textureEntries[textureCount];
     samplerEntry.binding = textureCount;
     textureEntry.binding = textureCount;
-    if (tex) {
-      samplerEntry.sampler = gfx::sampler_ref(tex.get_descriptor());
-      textureEntry.textureView = tex.texObj.ref->sampleTextureView;
+    if (tex && (info.sampledTextures[i] || info.sampledIndTextures[i])) {
+      samplerEntry.sampler = gfx::sampler_ref(tex.get_descriptor()).Get();
+      textureEntry.textureView = tex.texObj.ref->sampleTextureView.Get();
     } else {
-      samplerEntry.sampler = sEmptySampler;
-      textureEntry.textureView = sEmptyTextureView;
+      samplerEntry.sampler = sEmptySampler.Get();
+      textureEntry.textureView = sEmptyTextureView.Get();
     }
     ++textureCount;
   }
-  const wgpu::BindGroupDescriptor uniformBindGroupDescriptor{
-      .label = "GX Uniform Bind Group",
-      .layout = layouts.uniformLayout,
+  const WGPUBindGroupDescriptor uniformBindGroupDescriptor{
+      .label = {"GX Uniform Bind Group", WGPU_STRLEN},
+      .layout = layouts.uniformLayout.Get(),
       .entryCount = uniformBindIdx,
       .entries = uniformEntries.data(),
   };
-  const wgpu::BindGroupDescriptor samplerBindGroupDescriptor{
-      .label = "GX Sampler Bind Group",
-      .layout = layouts.samplerLayout,
+  const WGPUBindGroupDescriptor samplerBindGroupDescriptor{
+      .label = {"GX Sampler Bind Group", WGPU_STRLEN},
+      .layout = layouts.samplerLayout.Get(),
       .entryCount = textureCount,
       .entries = samplerEntries.data(),
   };
-  const wgpu::BindGroupDescriptor textureBindGroupDescriptor{
-      .label = "GX Texture Bind Group",
-      .layout = layouts.textureLayout,
+  const WGPUBindGroupDescriptor textureBindGroupDescriptor{
+      .label = {"GX Texture Bind Group", WGPU_STRLEN},
+      .layout = layouts.textureLayout.Get(),
       .entryCount = textureCount,
       .entries = textureEntries.data(),
   };
diff --git a/lib/gx/gx.hpp b/lib/gx/gx.hpp
index 78d2c80..82f800f 100644
--- a/lib/gx/gx.hpp
+++ b/lib/gx/gx.hpp
@@ -315,6 +315,8 @@ struct GXState {
   std::array<Mat3x4<float>, MaxPTTexMtx> ptTexMtxs;
   std::array<TcgConfig, MaxTexCoord> tcgs;
   std::array<TexCoordScale, MaxTexCoord> texCoordScales;
+  u16 lastVtxSize = 0;
+  GXVtxFmt lastVtxFmt = GX_MAX_VTXFMT;
   std::array<GXAttrType, MaxVtxAttr> vtxDesc;
   std::array<VtxFmt, MaxVtxFmt> vtxFmts;
   std::array<TevSwap, MaxTevSwap> tevSwapTable{
@@ -358,6 +360,10 @@ struct GXState {
   bool stateDirty = true;
   std::array<u32, 0x100> bpRegCache;
   GXState() { bpRegCache[0xFE] = 0x00FFFFFF; }
+
+  void clearVtxSizeCache() {
+    lastVtxFmt = GX_MAX_VTXFMT;
+  }
 };
 extern GXState g_gxState;
 
diff --git a/lib/gx/pipeline.cpp b/lib/gx/pipeline.cpp
index b851175..0ec6f36 100644
--- a/lib/gx/pipeline.cpp
+++ b/lib/gx/pipeline.cpp
@@ -3,11 +3,13 @@
 #include "../webgpu/gpu.hpp"
 #include "gx_fmt.hpp"
 #include "shader_info.hpp"
+#include "tracy/Tracy.hpp"
 
 namespace aurora::gx {
 static Module Log("aurora::gx");
 
 wgpu::RenderPipeline create_pipeline(const PipelineConfig& config) {
+  ZoneScoped;
   const auto shader = build_shader(config.shaderConfig);
   return build_pipeline(config, {}, shader, "GX Pipeline");
 }
diff --git a/lib/gx/shader.cpp b/lib/gx/shader.cpp
index a7bfe43..c8fedcf 100644
--- a/lib/gx/shader.cpp
+++ b/lib/gx/shader.cpp
@@ -13,6 +13,8 @@
 #include <string_view>
 #include <utility>
 
+#include "tracy/Tracy.hpp"
+
 namespace aurora::gx {
 using namespace fmt::literals;
 using namespace std::string_literals;
@@ -680,6 +682,7 @@ auto lighting_func(const ShaderConfig& config, const ColorChannelConfig& cc, u8
 }
 
 wgpu::ShaderModule build_shader(const ShaderConfig& config) noexcept {
+  ZoneScoped;
   const auto hash = xxh3_hash(config);
   {
     std::lock_guard lock{g_gxCachedShadersMutex};
diff --git a/lib/imgui.cpp b/lib/imgui.cpp
index 1e9ad52..34911b4 100644
--- a/lib/imgui.cpp
+++ b/lib/imgui.cpp
@@ -15,6 +15,7 @@
 #include "backends/imgui_impl_sdl3.h"
 #include "backends/imgui_impl_sdlrenderer3.h"
 #include "backends/imgui_impl_wgpu.h"
+#include "tracy/Tracy.hpp"
 
 namespace aurora::imgui {
 static float g_scale;
@@ -36,6 +37,7 @@ void create_context() noexcept {
 }
 
 void initialize() noexcept {
+  ZoneScoped;
   SDL_Renderer* renderer = window::get_sdl_renderer();
   ImGui_ImplSDL3_InitForSDLRenderer(window::get_sdl_window(), renderer);
   g_useSdlRenderer = renderer != nullptr;
@@ -50,6 +52,7 @@ void initialize() noexcept {
 }
 
 void shutdown() noexcept {
+  ZoneScoped;
   if (g_useSdlRenderer) {
     ImGui_ImplSDLRenderer3_Shutdown();
   } else {
@@ -69,6 +72,7 @@ void process_event(const SDL_Event& event) noexcept {
 }
 
 void new_frame(const AuroraWindowSize& size) noexcept {
+  ZoneScoped;
   const float framebufferScaleX =
       size.width > 0 ? static_cast<float>(size.native_fb_width) / static_cast<float>(size.width) : 1.0f;
   const float framebufferScaleY =
@@ -98,6 +102,7 @@ void new_frame(const AuroraWindowSize& size) noexcept {
 }
 
 void render(const wgpu::RenderPassEncoder& pass) noexcept {
+  ZoneScoped;
   ImGui::Render();
 
   auto* data = ImGui::GetDrawData();
diff --git a/lib/internal.hpp b/lib/internal.hpp
index da2a683..8c5df12 100644
--- a/lib/internal.hpp
+++ b/lib/internal.hpp
@@ -80,9 +80,7 @@ auto underlying(T value) -> std::underlying_type_t<T> {
   return static_cast<std::underlying_type_t<T>>(value);
 }
 
-#ifndef ALIGN
-#define ALIGN(x, a) (((x) + ((a) - 1)) & ~((a) - 1))
-#endif
+#define AURORA_ALIGN(x, a) (((x) + ((a) - 1)) & ~((a) - 1))
 
 #define POINTER_ADD_TYPE(type_, ptr_, offset_) ((type_)((uintptr_t)(ptr_) + (uintptr_t)(offset_)))
 #define POINTER_ADD(ptr_, offset_) POINTER_ADD_TYPE(decltype(ptr_), ptr_, offset_)
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index df1fcec..988a50b 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -56,6 +56,7 @@ if (AURORA_ENABLE_GX)
     absl::flat_hash_map
     absl::btree
     dawn::dawncpp_headers
+    TracyClient
     ${AURORA_SDL3_TARGET}
   )
 
diff --git a/tests/gx_test_stubs.cpp b/tests/gx_test_stubs.cpp
index 999edd1..bfa5bb7 100644
--- a/tests/gx_test_stubs.cpp
+++ b/tests/gx_test_stubs.cpp
@@ -97,7 +97,7 @@ void set_scissor(uint32_t x, uint32_t y, uint32_t w, uint32_t h) noexcept {}
 // --- Pipeline/draw command stubs ---
 namespace aurora::gfx {
 template <>
-PipelineRef pipeline_ref<clear::PipelineConfig>(clear::PipelineConfig config) {
+PipelineRef pipeline_ref<clear::PipelineConfig>(const clear::PipelineConfig& config) {
   return 0;
 }
 template <>
@@ -105,7 +105,7 @@ void push_draw_command<clear::DrawData>(clear::DrawData data) {
   // No-op
 }
 template <>
-PipelineRef pipeline_ref<gx::PipelineConfig>(gx::PipelineConfig config) {
+PipelineRef pipeline_ref<gx::PipelineConfig>(const gx::PipelineConfig& config) {
   return 0;
 }
 template <>