diff --git a/binaries/data/config/default.cfg b/binaries/data/config/default.cfg
index 2015cc6c8a..a89a933a21 100644
--- a/binaries/data/config/default.cfg
+++ b/binaries/data/config/default.cfg
@@ -201,6 +201,7 @@ hotkey.text.move.right = "Ctrl+RightArrow"    ; Move cursor to start of word to
 ; > PROFILER
 hotkey.profile.toggle = "F11"               ; Enable/disable real-time profiler
 hotkey.profile.save = "Shift+F11"           ; Save current profiler data to logs/profile.txt
+hotkey.profile2.enable = "F11"              ; Enable HTTP server for new profiler
 
 ; > QUICKSAVE
 hotkey.quicksave = "Shift+F5"
diff --git a/source/graphics/GameView.cpp b/source/graphics/GameView.cpp
index 7e17a4a103..39ad2f8416 100644
--- a/source/graphics/GameView.cpp
+++ b/source/graphics/GameView.cpp
@@ -483,7 +483,9 @@ void CGameView::Render()
 // Submit all objects visible in the given frustum
 void CGameView::EnumerateObjects(const CFrustum& frustum, SceneCollector* c)
 {
-	PROFILE_START("submit terrain");
+	{
+	PROFILE3("submit terrain");
+
 	CTerrain* pTerrain = m->Game->GetWorld()->GetTerrain();
 	const ssize_t patchesPerSide = pTerrain->GetPatchesPerSide();
 
@@ -548,11 +550,9 @@ void CGameView::EnumerateObjects(const CFrustum& frustum, SceneCollector* c)
 			}
 		}
 	}
-	PROFILE_END("submit terrain");
+	}
 
-	PROFILE_START("submit sim components");
 	m->Game->GetSimulation2()->RenderSubmit(*c, frustum, m->Culling);
-	PROFILE_END("submit sim components");
 }
 
 
diff --git a/source/graphics/ShaderManager.cpp b/source/graphics/ShaderManager.cpp
index ad3080c701..ba270f009c 100644
--- a/source/graphics/ShaderManager.cpp
+++ b/source/graphics/ShaderManager.cpp
@@ -23,6 +23,7 @@
 #include "lib/utf8.h"
 #include "ps/CLogger.h"
 #include "ps/Filesystem.h"
+#include "ps/Profile.h"
 #include "ps/XML/Xeromyces.h"
 #include "ps/XML/XMLWriter.h"
 #include "renderer/Renderer.h"
@@ -74,6 +75,9 @@ CShaderProgramPtr CShaderManager::LoadProgram(const char* name, const std::map<C
 
 bool CShaderManager::NewProgram(const char* name, const std::map<CStr, CStr>& baseDefines, CShaderProgramPtr& program)
 {
+	PROFILE2("loading shader");
+	PROFILE2_ATTR("name: %s", name);
+
 	if (strncmp(name, "fixed:", 6) == 0)
 	{
 		program = CShaderProgramPtr(CShaderProgram::ConstructFFP(name+6));
diff --git a/source/graphics/TextureConverter.cpp b/source/graphics/TextureConverter.cpp
index 341276dd81..151013ce48 100644
--- a/source/graphics/TextureConverter.cpp
+++ b/source/graphics/TextureConverter.cpp
@@ -26,6 +26,7 @@
 #include "maths/MD5.h"
 #include "ps/CLogger.h"
 #include "ps/CStr.h"
+#include "ps/Profiler2.h"
 #include "ps/XML/Xeromyces.h"
 
 #include "nvtt/nvtt.h"
@@ -486,12 +487,16 @@ bool CTextureConverter::IsBusy()
 void* CTextureConverter::RunThread(void* data)
 {
 	debug_SetThreadName("TextureConverter");
+	g_Profiler2.RegisterCurrentThread("texconv");
 
 	CTextureConverter* textureConverter = static_cast<CTextureConverter*>(data);
 
 	// Wait until the main thread wakes us up
 	while (SDL_SemWait(textureConverter->m_WorkerSem) == 0)
 	{
+		g_Profiler2.RecordSyncMarker();
+		PROFILE2_EVENT("wakeup");
+
 		pthread_mutex_lock(&textureConverter->m_WorkerMutex);
 		if (textureConverter->m_Shutdown)
 		{
@@ -513,9 +518,13 @@ void* CTextureConverter::RunThread(void* data)
 
 //		TIMER(L"TextureConverter compress");
 
-		// Perform the compression
-		nvtt::Compressor compressor;
-		result->ret = compressor.process(request->inputOptions, request->compressionOptions, request->outputOptions);
+		{
+			PROFILE2("compress");
+
+			// Perform the compression
+			nvtt::Compressor compressor;
+			result->ret = compressor.process(request->inputOptions, request->compressionOptions, request->outputOptions);
+		}
 
 		// Ugly hack: NVTT 2.0 doesn't set DDPF_ALPHAPIXELS for DXT1a, so we can't
 		// distinguish it from DXT1. (It's fixed in trunk by
diff --git a/source/graphics/TextureManager.cpp b/source/graphics/TextureManager.cpp
index fb30ec160e..791be0688b 100644
--- a/source/graphics/TextureManager.cpp
+++ b/source/graphics/TextureManager.cpp
@@ -29,6 +29,7 @@
 #include "ps/CacheLoader.h"
 #include "ps/CLogger.h"
 #include "ps/Filesystem.h"
+#include "ps/Profile.h"
 
 #include <iomanip>
 #include <boost/unordered_map.hpp>
@@ -170,6 +171,9 @@ public:
 		if (m_DisableGL)
 			return;
 
+		PROFILE2("load texture");
+		PROFILE2_ATTR("name: %ls", path.string().c_str());
+
 		Handle h = ogl_tex_load(m_VFS, path, RES_UNIQUE);
 		if (h <= 0)
 		{
@@ -285,6 +289,9 @@ public:
 	{
 		VfsPath sourcePath = texture->m_Properties.m_Path;
 
+		PROFILE2("convert texture");
+		PROFILE2_ATTR("name: %ls", sourcePath.string().c_str());
+
 		MD5 hash;
 		u32 version;
 		PrepareCacheKey(texture, hash, version);
diff --git a/source/gui/GUIManager.cpp b/source/gui/GUIManager.cpp
index b46cadc7a8..f7befba292 100644
--- a/source/gui/GUIManager.cpp
+++ b/source/gui/GUIManager.cpp
@@ -143,6 +143,10 @@ void CGUIManager::LoadPage(SGUIPage& page)
 		}
 
 		CStrW name (node.GetText().FromUTF8());
+
+		PROFILE2("load gui xml");
+		PROFILE2_ATTR("name: %ls", name.c_str());
+
 		TIMER(name.c_str());
 		VfsPath path = VfsPath("gui") / name;
 		page.gui->LoadXmlFile(path, page.inputs);
@@ -253,6 +257,8 @@ void CGUIManager::SendEventToAll(const CStr& eventName)
 
 void CGUIManager::TickObjects()
 {
+	PROFILE3("gui tick");
+
 	// Save an immutable copy so iterators aren't invalidated by tick handlers
 	PageStackType pageStack = m_PageStack;
 
@@ -266,6 +272,8 @@ void CGUIManager::TickObjects()
 
 void CGUIManager::Draw()
 {
+	PROFILE3("render gui");
+
 	for (PageStackType::iterator it = m_PageStack.begin(); it != m_PageStack.end(); ++it)
 		it->gui->Draw();
 }
diff --git a/source/gui/MiniMap.cpp b/source/gui/MiniMap.cpp
index ed3dd36668..05e234e277 100644
--- a/source/gui/MiniMap.cpp
+++ b/source/gui/MiniMap.cpp
@@ -262,7 +262,7 @@ void CMiniMap::DrawTexture(float coordMax, float angle, float x, float y, float
 
 void CMiniMap::Draw()
 {
-	PROFILE("minimap");
+	PROFILE3("render minimap");
 
 	// The terrain isn't actually initialized until the map is loaded, which
 	// happens when the game is started, so abort until then.
diff --git a/source/main.cpp b/source/main.cpp
index c252baa88e..ad1c9448fd 100644
--- a/source/main.cpp
+++ b/source/main.cpp
@@ -49,6 +49,7 @@ that of Atlas depending on commandline parameters.
 #include "ps/Hotkey.h"
 #include "ps/Loader.h"
 #include "ps/Profile.h"
+#include "ps/Profiler2.h"
 #include "ps/Pyrogenesis.h"
 #include "ps/Replay.h"
 #include "ps/UserReport.h"
@@ -132,6 +133,11 @@ static InReaction MainInputHandler(const SDL_Event_* ev)
 			g_VideoMode.ToggleFullscreen();
 			return IN_HANDLED;
 		}
+		else if (hotkey == "profile2.enable")
+		{
+			g_Profiler2.EnableHTTP();
+			return IN_HANDLED;
+		}
 		break;
 	}
 
@@ -142,7 +148,7 @@ static InReaction MainInputHandler(const SDL_Event_* ev)
 // dispatch all pending events to the various receivers.
 static void PumpEvents()
 {
-	PROFILE("dispatch events");
+	PROFILE3("dispatch events");
 
 	SDL_Event_ ev;
 	while(SDL_PollEvent(&ev.ev))
@@ -182,6 +188,8 @@ return false;
 
 static int ProgressiveLoad()
 {
+	PROFILE3("progressive load");
+
 	wchar_t description[100];
 	int progress_percent;
 	try
@@ -226,6 +234,8 @@ static int ProgressiveLoad()
 
 static void RendererIncrementalLoad()
 {
+	PROFILE3("renderer incremental load");
+
 	const double maxTime = 0.1f;
 
 	double startTime = timer_Time();
@@ -241,6 +251,9 @@ static bool quit = false;	// break out of main loop
 
 static void Frame()
 {
+	g_Profiler2.RecordFrameStart();
+	PROFILE2("frame");
+
 	ogl_WarnIfError();
 
 	// get elapsed time
@@ -269,7 +282,7 @@ static void Frame()
 	// debugging easier.
 	if( !g_NetClient && !g_app_has_focus )
 	{
-		PROFILE("non-focus delay");
+		PROFILE3("non-focus delay");
 		need_update = false;
 		// don't use SDL_WaitEvent: don't want the main loop to freeze until app focus is restored
 		SDL_Delay(10);
@@ -279,10 +292,7 @@ static void Frame()
 	// this is mostly relevant for "inactive" state, so that other windows
 	// get enough CPU time, but it's always nice for power+thermal management.
 
-	bool is_building_archive;	// must come before PROFILE_START's {
-	PROFILE_START("build archive");
-	is_building_archive = ProgressiveBuildArchive();
-	PROFILE_END( "build archive");
+	bool is_building_archive = ProgressiveBuildArchive();
 
 	// this scans for changed files/directories and reloads them, thus
 	// allowing hotloading (changes are immediately assimilated in-game).
@@ -290,23 +300,13 @@ static void Frame()
 	// archive file each iteration, but keeps it locked; reloading
 	// would trigger a warning because the file can't be opened.
 	if(!is_building_archive)
-	{
-		PROFILE_START("reload changed files");
 		ReloadChangedFiles();
-		PROFILE_END( "reload changed files");
-	}
 
-	PROFILE_START("progressive load");
 	ProgressiveLoad();
-	PROFILE_END("progressive load");
 
-	PROFILE_START("renderer incremental load");
 	RendererIncrementalLoad();
-	PROFILE_END("renderer incremental load");
 
-	PROFILE_START("input");
 	PumpEvents();
-	PROFILE_END("input");
 
 	// if the user quit by closing the window, the GL context will be broken and
 	// may crash when we call Render() on some drivers, so leave this loop
@@ -321,16 +321,12 @@ static void Frame()
 		g_ResizedW = g_ResizedH = 0;
 	}
 
-	PROFILE_START("network poll");
 	if (g_NetClient)
 		g_NetClient->Poll();
-	PROFILE_END("network poll");
 
 	ogl_WarnIfError();
 
-	PROFILE_START("gui tick");
 	g_GUI->TickObjects();
-	PROFILE_END("gui tick");
 
 	ogl_WarnIfError();
 
@@ -338,11 +334,8 @@ static void Frame()
 	{
 		g_Game->Update(TimeSinceLastFrame);
 
-		PROFILE_START( "camera update" );
 		g_Game->GetView()->Update(float(TimeSinceLastFrame));
-		PROFILE_END( "camera update" );
 
-		PROFILE_START( "sound update" );
 		CCamera* camera = g_Game->GetView()->GetCamera();
 		CMatrix3D& orientation = camera->m_Orientation;
 		float* pos = &orientation._data[12];
@@ -352,37 +345,37 @@ static void Frame()
 		// is going wrong, because the listener and camera are supposed to
 		// coincide in position and orientation.
 		float down[3] = { -up[0], -up[1], -up[2] };
-		if(snd_update(pos, dir, down) < 0)
-			debug_printf(L"snd_update failed\n");
-		PROFILE_END( "sound update" );
+
+		{
+			PROFILE3("sound update");
+			if (snd_update(pos, dir, down) < 0)
+				debug_printf(L"snd_update failed\n");
+		}
 	}
 	else
 	{
-		if(snd_update(0, 0, 0) < 0)
+		PROFILE3("sound update (0)");
+		if (snd_update(0, 0, 0) < 0)
 			debug_printf(L"snd_update (pos=0 version) failed\n");
 	}
 
 	// Immediately flush any messages produced by simulation code
-	PROFILE_START("network flush");
 	if (g_NetClient)
 		g_NetClient->Flush();
-	PROFILE_END("network flush");
 
 	g_UserReporter.Update();
 
 	g_Console->Update(TimeSinceLastFrame);
 
-	PROFILE_START("render");
 	ogl_WarnIfError();
 	if(need_render)
 	{
 		Render();
-		PROFILE_START( "swap buffers" );
+
+		PROFILE3("swap buffers");
 		SDL_GL_SwapBuffers();
-		PROFILE_END( "swap buffers" );
 	}
 	ogl_WarnIfError();
-	PROFILE_END("render");
 
 	g_Profiler.Frame();
 
diff --git a/source/network/NetSession.cpp b/source/network/NetSession.cpp
index 28ca8445fc..5d92ef0bce 100644
--- a/source/network/NetSession.cpp
+++ b/source/network/NetSession.cpp
@@ -94,6 +94,8 @@ void CNetClientSession::Disconnect(u32 reason)
 
 void CNetClientSession::Poll()
 {
+	PROFILE3("net client poll");
+
 	ENSURE(m_Host && m_Server);
 
 	m_FileTransferer.Poll();
@@ -152,6 +154,8 @@ void CNetClientSession::Poll()
 
 void CNetClientSession::Flush()
 {
+	PROFILE3("net client flush");
+
 	ENSURE(m_Host && m_Server);
 
 	enet_host_flush(m_Host);
diff --git a/source/network/NetTurnManager.cpp b/source/network/NetTurnManager.cpp
index dddedde8dd..e80ae743a7 100644
--- a/source/network/NetTurnManager.cpp
+++ b/source/network/NetTurnManager.cpp
@@ -149,7 +149,7 @@ bool CNetTurnManager::Update(float frameLength, size_t maxTurns)
 		// Save the current state for rewinding, if enabled
 		if (m_TimeWarpNumTurns && (m_CurrentTurn % m_TimeWarpNumTurns) == 0)
 		{
-			PROFILE("time warp serialization");
+			PROFILE3("time warp serialization");
 			std::stringstream stream;
 			m_Simulation2.SerializeState(stream);
 			m_TimeWarpStates.push_back(stream.str());
@@ -390,7 +390,7 @@ void CNetClientTurnManager::NotifyFinishedUpdate(u32 turn)
 	bool quick = !TurnNeedsFullHash(turn);
 	std::string hash;
 	{
-		PROFILE("state hash check");
+		PROFILE3("state hash check");
 		bool ok = m_Simulation2.ComputeStateHash(hash, quick);
 		ENSURE(ok);
 	}
@@ -435,7 +435,7 @@ void CNetLocalTurnManager::NotifyFinishedUpdate(u32 UNUSED(turn))
 #if 0 // this hurts performance and is only useful for verifying log replays
 	std::string hash;
 	{
-		PROFILE("state hash check");
+		PROFILE3("state hash check");
 		bool ok = m_Simulation2.ComputeStateHash(hash);
 		ENSURE(ok);
 	}
diff --git a/source/ps/CConsole.cpp b/source/ps/CConsole.cpp
index 3e6ef200da..b4a338de34 100644
--- a/source/ps/CConsole.cpp
+++ b/source/ps/CConsole.cpp
@@ -175,6 +175,8 @@ void CConsole::Render()
 {
 	if (! (m_bVisible || m_bToggle) ) return;
 
+	PROFILE3("render console");
+
 	CFont font(CONSOLE_FONT);
 	font.Bind();
 
diff --git a/source/ps/CLogger.cpp b/source/ps/CLogger.cpp
index b5ed7ee917..f9747425e0 100644
--- a/source/ps/CLogger.cpp
+++ b/source/ps/CLogger.cpp
@@ -26,6 +26,7 @@
 #include "lib/res/graphics/unifont.h"
 #include "lib/sysdep/sysdep.h"
 #include "ps/Font.h"
+#include "ps/Profile.h"
 
 #include <ctime>
 #include <iostream>
@@ -276,6 +277,8 @@ void CLogger::LogError(const wchar_t* fmt, ...)
 
 void CLogger::Render()
 {
+	PROFILE3("render logger");
+
 	CleanupRenderQueue();
 
 	CFont font(L"mono-stroke-10");
diff --git a/source/ps/Filesystem.cpp b/source/ps/Filesystem.cpp
index c2968984c5..60a9ca01f7 100644
--- a/source/ps/Filesystem.cpp
+++ b/source/ps/Filesystem.cpp
@@ -20,6 +20,7 @@
 
 #include "gui/GUIManager.h"
 #include "ps/CLogger.h"
+#include "ps/Profile.h"
 
 #include "lib/res/h_mgr.h"	// h_reload
 #include "lib/sysdep/dir_watch.h"
@@ -67,6 +68,8 @@ static bool CanIgnore(const DirWatchNotification& notification)
 
 Status ReloadChangedFiles()
 {
+	PROFILE3("hotload");
+
 	std::vector<DirWatchNotification> notifications;
 	RETURN_STATUS_IF_ERR(dir_watch_Poll(notifications));
 	for(size_t i = 0; i < notifications.size(); i++)
diff --git a/source/ps/Game.cpp b/source/ps/Game.cpp
index 6307ace09d..d9ad8aa388 100644
--- a/source/ps/Game.cpp
+++ b/source/ps/Game.cpp
@@ -273,7 +273,6 @@ bool CGame::Update(double deltaTime, bool doInterpolate)
 		// so just use the sim rate itself as the number of turns per frame.
 		size_t maxTurns = (size_t)m_SimRate;
 
-		PROFILE("simulation update");
 		if (m_TurnManager->Update(deltaTime, maxTurns))
 		{
 			g_GUI->SendEventToAll("SimulationUpdate");
@@ -283,7 +282,6 @@ bool CGame::Update(double deltaTime, bool doInterpolate)
 
 	if (doInterpolate)
 	{
-		PROFILE("interpolate");
 		m_TurnManager->Interpolate(deltaTime);
 	}
 	
diff --git a/source/ps/GameSetup/GameSetup.cpp b/source/ps/GameSetup/GameSetup.cpp
index 791b1bd392..a9175709fd 100644
--- a/source/ps/GameSetup/GameSetup.cpp
+++ b/source/ps/GameSetup/GameSetup.cpp
@@ -47,6 +47,7 @@
 #include "ps/Overlay.h"
 #include "ps/Profile.h"
 #include "ps/ProfileViewer.h"
+#include "ps/Profiler2.h"
 #include "ps/UserReport.h"
 #include "ps/Util.h"
 #include "ps/VideoMode.h"
@@ -193,6 +194,8 @@ void GUI_DisplayLoadProgress(int percent, const wchar_t* pending_task)
 
 void Render()
 {
+	PROFILE3("render");
+
 	ogl_WarnIfError();
 
 	CStr skystring = "255 0 255";
@@ -235,10 +238,8 @@ void Render()
 
 	g_Renderer.RenderTextOverlays();
 
-	// Temp GUI message GeeTODO
-	PROFILE_START("render gui");
-	if(g_DoRenderGui) g_GUI->Draw();
-	PROFILE_END("render gui");
+	if (g_DoRenderGui)
+		g_GUI->Draw();
 
 	ogl_WarnIfError();
 
@@ -253,30 +254,25 @@ void Render()
 
 	glLoadIdentity();
 
-	PROFILE_START("render console");
 	g_Console->Render();
-	PROFILE_END("render console");
 
 	ogl_WarnIfError();
 
-	PROFILE_START("render logger");
-	if(g_DoRenderLogger) g_Logger->Render();
-	PROFILE_END("render logger");
+	if (g_DoRenderLogger)
+		g_Logger->Render();
 
 	ogl_WarnIfError();
 
 	// Profile information
 
-	PROFILE_START("render profiling");
 	g_ProfileViewer.RenderProfile();
-	PROFILE_END("render profiling");
 
 	ogl_WarnIfError();
 
 	// Draw the cursor (or set the Windows cursor, on Windows)
 	if (g_DoRenderCursor)
 	{
-		PROFILE("render cursor");
+		PROFILE3("render cursor");
 		CStrW cursorName = g_CursorName;
 		if (cursorName.empty())
 		{
@@ -298,6 +294,14 @@ void Render()
 
 	g_Renderer.EndFrame();
 
+	PROFILE2_ATTR("draw calls: %d", (int)g_Renderer.GetStats().m_DrawCalls);
+	PROFILE2_ATTR("terrain tris: %d", (int)g_Renderer.GetStats().m_TerrainTris);
+	PROFILE2_ATTR("water tris: %d", (int)g_Renderer.GetStats().m_WaterTris);
+	PROFILE2_ATTR("model tris: %d", (int)g_Renderer.GetStats().m_ModelTris);
+	PROFILE2_ATTR("overlay tris: %d", (int)g_Renderer.GetStats().m_OverlayTris);
+	PROFILE2_ATTR("blend splats: %d", (int)g_Renderer.GetStats().m_BlendSplats);
+	PROFILE2_ATTR("particles: %d", (int)g_Renderer.GetStats().m_Particles);
+
 	ogl_WarnIfError();
 }
 
@@ -694,6 +698,8 @@ void Shutdown(int UNUSED(flags))
 	wmi_Shutdown();
 	TIMER_END(L"shutdown wmi");
 #endif
+
+	g_Profiler2.Shutdown();
 }
 
 #if OS_UNIX
@@ -773,6 +779,10 @@ void EarlyInit()
 
 	timer_LatchStartTime();
 
+	// initialise profiler early so it can profile startup
+	g_Profiler2.Initialise();
+	g_Profiler2.RegisterCurrentThread("main");
+
 	FixLocales();
 
 	// Because we do GL calls from a secondary thread, Xlib needs to
@@ -844,6 +854,8 @@ void Init(const CmdLineArgs& args, int UNUSED(flags))
 
 	if (!g_Quickstart)
 		g_UserReporter.Initialize(); // after config
+
+	PROFILE2_EVENT("Init finished");
 }
 
 void InitGraphics(const CmdLineArgs& args, int flags)
diff --git a/source/ps/Profile.h b/source/ps/Profile.h
index 8ee102806e..9c32d4f719 100644
--- a/source/ps/Profile.h
+++ b/source/ps/Profile.h
@@ -25,6 +25,7 @@
 #include <vector>
 
 #include "lib/adts/ring_buf.h"
+#include "ps/Profiler2.h"
 #include "ps/Singleton.h"
 #include "ps/ThreadUtil.h"
 
@@ -146,14 +147,14 @@ public:
 class CProfileSample
 {
 public:
-	CProfileSample( const char* name )
+	CProfileSample(const char* name)
 	{
 		if (CProfileManager::IsInitialised())
 		{
 			// The profiler is only safe to use on the main thread
 			ENSURE(ThreadUtil::IsMainThread());
 
-			g_Profiler.Start( name );
+			g_Profiler.Start(name);
 		}
 	}
 	~CProfileSample()
@@ -187,11 +188,14 @@ public:
 	}
 };
 
-// Put a PROFILE( xyz ) block at the start of all code to be profiled.
+// Put a PROFILE("xyz") block at the start of all code to be profiled.
 // Profile blocks last until the end of the containing scope.
-#define PROFILE( name ) CProfileSample __profile( name )
+#define PROFILE(name) CProfileSample __profile(name)
 // Cheat a bit to make things slightly easier on the user
-#define PROFILE_START( name ) { CProfileSample __profile( name )
-#define PROFILE_END( name ) }
+#define PROFILE_START(name) { CProfileSample __profile(name)
+#define PROFILE_END(name) }
+
+// Do both old and new profilers simultaneously (1+2=3), for convenience.
+#define PROFILE3(name) PROFILE(name); PROFILE2(name)
 
 #endif // INCLUDED_PROFILE
diff --git a/source/ps/ProfileViewer.cpp b/source/ps/ProfileViewer.cpp
index fb22a8e46a..407e30f0ab 100644
--- a/source/ps/ProfileViewer.cpp
+++ b/source/ps/ProfileViewer.cpp
@@ -159,6 +159,8 @@ void CProfileViewer::RenderProfile()
 		return;
 	}
 
+	PROFILE3("render profiler");
+
 	glEnable(GL_BLEND);
 	glBlendFunc(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA);
 
diff --git a/source/ps/ThreadUtil.h b/source/ps/ThreadUtil.h
index aa00ae27be..d4e853ad0b 100644
--- a/source/ps/ThreadUtil.h
+++ b/source/ps/ThreadUtil.h
@@ -19,7 +19,6 @@
 #define INCLUDED_THREADUTIL
 
 #include "lib/posix/posix_pthread.h"
-#include "lib/posix/posix_filesystem.h"	// O_CREAT
 
 #ifdef DEBUG_LOCKS
 
diff --git a/source/renderer/OverlayRenderer.cpp b/source/renderer/OverlayRenderer.cpp
index a8cd900a00..4afbc7970b 100644
--- a/source/renderer/OverlayRenderer.cpp
+++ b/source/renderer/OverlayRenderer.cpp
@@ -116,7 +116,7 @@ void OverlayRenderer::EndFrame()
 
 void OverlayRenderer::PrepareForRendering()
 {
-	PROFILE("prepare overlays");
+	PROFILE3("prepare overlays");
 
 	// This is where we should do something like sort the overlays by
 	// colour/sprite/etc for more efficient rendering
@@ -138,7 +138,7 @@ void OverlayRenderer::PrepareForRendering()
 
 void OverlayRenderer::RenderOverlaysBeforeWater()
 {
-	PROFILE("render overlays (before water)");
+	PROFILE3("render overlays (before)");
 
 	glDisable(GL_TEXTURE_2D);
 	glEnable(GL_BLEND);
@@ -166,7 +166,7 @@ void OverlayRenderer::RenderOverlaysBeforeWater()
 
 void OverlayRenderer::RenderOverlaysAfterWater()
 {
-	PROFILE("render overlays (after water)");
+	PROFILE3("render overlays (after)");
 
 	if (!m->texlines.empty())
 	{
@@ -262,7 +262,7 @@ void OverlayRenderer::RenderOverlaysAfterWater()
 
 void OverlayRenderer::RenderForegroundOverlays(const CCamera& viewCamera)
 {
-	PROFILE("render overlays (fg)");
+	PROFILE3("render overlays (fg)");
 
 	glEnable(GL_TEXTURE_2D);
 	glEnable(GL_BLEND);
diff --git a/source/renderer/ParticleRenderer.cpp b/source/renderer/ParticleRenderer.cpp
index aa817b8351..84e46337cd 100644
--- a/source/renderer/ParticleRenderer.cpp
+++ b/source/renderer/ParticleRenderer.cpp
@@ -76,6 +76,8 @@ struct SortEmitterDistance
 
 void ParticleRenderer::PrepareForRendering()
 {
+	PROFILE3("prepare particles");
+
 	// Can't load the shader in the constructor because it's called before the
 	// renderer initialisation is complete, so load it the first time through here
 	if (!m->shader)
diff --git a/source/renderer/PatchRData.cpp b/source/renderer/PatchRData.cpp
index 5f9b5dc711..2354816b57 100644
--- a/source/renderer/PatchRData.cpp
+++ b/source/renderer/PatchRData.cpp
@@ -146,6 +146,8 @@ struct SBlendLayer
 
 void CPatchRData::BuildBlends()
 {
+	PROFILE3("build blends");
+
 	m_BlendSplats.clear();
 
 	std::vector<SBlendVertex> blendVertices;
@@ -428,6 +430,8 @@ void CPatchRData::AddBlend(std::vector<SBlendVertex>& blendVertices, std::vector
 
 void CPatchRData::BuildIndices()
 {
+	PROFILE3("build indices");
+
 	CTerrain* terrain = m_Patch->m_Parent;
 
 	ssize_t px = m_Patch->m_X * PATCH_SIZE;
@@ -520,6 +524,8 @@ void CPatchRData::BuildIndices()
 
 void CPatchRData::BuildVertices()
 {
+	PROFILE3("build vertices");
+
 	// create both vertices and lighting colors
 
 	// number of vertices in each direction in each patch
@@ -613,6 +619,8 @@ void CPatchRData::BuildSide(std::vector<SSideVertex>& vertices, CPatchSideFlags
 
 void CPatchRData::BuildSides()
 {
+	PROFILE3("build sides");
+
 	std::vector<SSideVertex> sideVertices;
 
 	int sideFlags = m_Patch->GetSideFlags();
@@ -1142,6 +1150,8 @@ void CPatchRData::RenderPriorities()
 // Build vertex buffer for water vertices over our patch
 void CPatchRData::BuildWater()
 {
+	PROFILE3("build water");
+
 	// number of vertices in each direction in each patch
 	ENSURE((PATCH_SIZE % water_cell_size) == 0);
 
diff --git a/source/renderer/Renderer.cpp b/source/renderer/Renderer.cpp
index b8308bc18f..739bf4a73a 100644
--- a/source/renderer/Renderer.cpp
+++ b/source/renderer/Renderer.cpp
@@ -933,7 +933,7 @@ void CRenderer::SetClearColor(SColor4ub color)
 
 void CRenderer::RenderShadowMap()
 {
-	PROFILE("render shadow map");
+	PROFILE3("render shadow map");
 
 	m->shadow->BeginRender();
 
@@ -978,7 +978,7 @@ void CRenderer::RenderShadowMap()
 
 void CRenderer::RenderPatches(const CFrustum* frustum)
 {
-	PROFILE("render patches");
+	PROFILE3("render patches");
 
 	bool filtered = false;
 	if (frustum)
@@ -1050,7 +1050,7 @@ private:
 
 void CRenderer::RenderModels(const CFrustum* frustum)
 {
-	PROFILE("render models");
+	PROFILE3("render models");
 
 	int flags = 0;
 	if (frustum)
@@ -1087,7 +1087,7 @@ void CRenderer::RenderModels(const CFrustum* frustum)
 
 void CRenderer::RenderTransparentModels(ETransparentMode transparentMode, const CFrustum* frustum)
 {
-	PROFILE("render transparent models");
+	PROFILE3("render transparent models");
 
 	int flags = 0;
 	if (frustum)
@@ -1360,7 +1360,7 @@ SScreenRect CRenderer::RenderRefractions(const CBound &scissor)
 
 void CRenderer::RenderSilhouettes()
 {
-	PROFILE("render silhouettes");
+	PROFILE3("render silhouettes");
 
 	// Render silhouettes of units hidden behind terrain or occluders.
 	// To avoid breaking the standard rendering of alpha-blended objects, this
@@ -1473,7 +1473,7 @@ void CRenderer::RenderParticles()
 	if (GetRenderPath() != RP_SHADER)
 		return;
 
-	PROFILE("render particles");
+	PROFILE3("render particles");
 
 	m->particleRenderer.RenderParticles();
 
@@ -1499,7 +1499,7 @@ void CRenderer::RenderParticles()
 // RenderSubmissions: force rendering of any batched objects
 void CRenderer::RenderSubmissions()
 {
-	PROFILE("render submissions");
+	PROFILE3("render submissions");
 
 	ogl_WarnIfError();
 
@@ -1507,7 +1507,8 @@ void CRenderer::RenderSubmissions()
 	m->SetOpenGLCamera(m_ViewCamera);
 
 	// Prepare model renderers
-	PROFILE_START("prepare models");
+	{
+	PROFILE3("prepare models");
 	m->Model.Normal->PrepareModels();
 	m->Model.Player->PrepareModels();
 	if (m->Model.Normal != m->Model.NormalInstancing)
@@ -1515,19 +1516,13 @@ void CRenderer::RenderSubmissions()
 	if (m->Model.Player != m->Model.PlayerInstancing)
 		m->Model.PlayerInstancing->PrepareModels();
 	m->Model.Transp->PrepareModels();
-	PROFILE_END("prepare models");
+	}
 
-	PROFILE_START("prepare terrain");
 	m->terrainRenderer->PrepareForRendering();
-	PROFILE_END("prepare terrain");
 
-	PROFILE_START("prepare overlays");
 	m->overlayRenderer.PrepareForRendering();
-	PROFILE_END("prepare overlays");
 
-	PROFILE_START("prepare particles");
 	m->particleRenderer.PrepareForRendering();
-	PROFILE_END("prepare particles");
 
 	if (m_Caps.m_Shadows && m_Options.m_Shadows && GetRenderPath() == RP_SHADER)
 	{
@@ -1639,7 +1634,7 @@ void CRenderer::RenderSubmissions()
 // EndFrame: signal frame end
 void CRenderer::EndFrame()
 {
-	PROFILE("end frame");
+	PROFILE3("end frame");
 
 	// empty lists
 	m->terrainRenderer->EndFrame();
@@ -1657,9 +1652,12 @@ void CRenderer::EndFrame()
 
 	ogl_tex_bind(0, 0);
 
-	if (glGetError())
 	{
-		ONCE(LOGERROR(L"CRenderer::EndFrame: GL errors occurred"));
+		PROFILE3("error check");
+		if (glGetError())
+		{
+			ONCE(LOGERROR(L"CRenderer::EndFrame: GL errors occurred"));
+		}
 	}
 }
 
diff --git a/source/renderer/TerrainOverlay.cpp b/source/renderer/TerrainOverlay.cpp
index f598d42c4a..84e1a40bd0 100644
--- a/source/renderer/TerrainOverlay.cpp
+++ b/source/renderer/TerrainOverlay.cpp
@@ -93,7 +93,7 @@ void TerrainOverlay::RenderOverlays()
 	if (g_TerrainOverlayList.size() == 0)
 		return;
 
-	PROFILE("render terrain overlays");
+	PROFILE3("render terrain overlays");
 
 	glEnable(GL_BLEND);
 	glBlendFunc(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA);
diff --git a/source/renderer/TerrainRenderer.cpp b/source/renderer/TerrainRenderer.cpp
index 13b71bccf2..475635668d 100644
--- a/source/renderer/TerrainRenderer.cpp
+++ b/source/renderer/TerrainRenderer.cpp
@@ -610,7 +610,7 @@ CBound TerrainRenderer::ScissorWater(const CMatrix3D &viewproj)
 // Render fancy water
 bool TerrainRenderer::RenderFancyWater()
 {
-	PROFILE("render fancy water");
+	PROFILE3("render fancy water");
 
 	// If we're using fancy water, make sure its shader is loaded
 	if (!m->fancyWaterShader)
@@ -739,7 +739,7 @@ bool TerrainRenderer::RenderFancyWater()
 
 void TerrainRenderer::RenderSimpleWater()
 {
-	PROFILE("render simple water");
+	PROFILE3("render simple water");
 
 	WaterManager* WaterMgr = g_Renderer.GetWaterManager();
 	CLOSTexture& losTexture = g_Game->GetView()->GetLOSTexture();
diff --git a/source/simulation2/Simulation2.cpp b/source/simulation2/Simulation2.cpp
index c56f30f069..4aca8fa4f2 100644
--- a/source/simulation2/Simulation2.cpp
+++ b/source/simulation2/Simulation2.cpp
@@ -324,6 +324,9 @@ void CSimulation2Impl::ReportSerializationFailure(
 
 void CSimulation2Impl::Update(int turnLength, const std::vector<SimulationCommand>& commands)
 {
+	PROFILE3("sim update");
+	PROFILE2_ATTR("turn %d", (int)m_TurnNumber);
+
 	fixed turnLengthFixed = fixed::FromInt(turnLength) / 1000;
 
 	/*
@@ -510,6 +513,8 @@ void CSimulation2Impl::UpdateComponents(CSimContext& simContext, fixed turnLengt
 
 void CSimulation2Impl::Interpolate(float frameLength, float frameOffset)
 {
+	PROFILE3("sim interpolate");
+
 	m_LastFrameOffset = frameOffset;
 
 	CMessageInterpolate msg(frameLength, frameOffset);
@@ -654,6 +659,8 @@ void CSimulation2::Interpolate(float frameLength, float frameOffset)
 
 void CSimulation2::RenderSubmit(SceneCollector& collector, const CFrustum& frustum, bool culling)
 {
+	PROFILE3("sim submit");
+
 	CMessageRenderSubmit msg(collector, frustum, culling);
 	m->m_ComponentManager.BroadcastMessage(msg);
 }
diff --git a/source/simulation2/components/CCmpAIManager.cpp b/source/simulation2/components/CCmpAIManager.cpp
index c222bac993..f16e006365 100644
--- a/source/simulation2/components/CCmpAIManager.cpp
+++ b/source/simulation2/components/CCmpAIManager.cpp
@@ -440,7 +440,7 @@ private:
 		// Deserialize the game state, to pass to the AI's HandleMessage
 		CScriptVal state;
 		{
-			PROFILE("AI compute read state");
+			PROFILE3("AI compute read state");
 			state = m_ScriptInterface.ReadStructuredClone(m_GameState);
 			m_ScriptInterface.SetProperty(state.get(), "passabilityMap", m_PassabilityMapVal, true);
 			m_ScriptInterface.SetProperty(state.get(), "territoryMap", m_TerritoryMapVal, true);
@@ -452,10 +452,12 @@ private:
 		// affecting other AI scripts they share it with. But the performance
 		// cost is far too high, so we won't do that.
 
+		for (size_t i = 0; i < m_Players.size(); ++i)
 		{
-			PROFILE("AI compute scripts");
-			for (size_t i = 0; i < m_Players.size(); ++i)
-				m_Players[i]->Run(state);
+			PROFILE3("AI script");
+			PROFILE2_ATTR("player: %d", m_Players[i]->m_Player);
+			PROFILE2_ATTR("script: %ls", m_Players[i]->m_AIName.c_str());
+			m_Players[i]->Run(state);
 		}
 
 		// Run the GC every so often.
@@ -463,7 +465,7 @@ private:
 		// since it avoids random GC delays while running other scripts)
 		if (m_TurnNum++ % 25 == 0)
 		{
-			PROFILE("AI compute GC");
+			PROFILE3("AI compute GC");
 			m_ScriptInterface.MaybeGC();
 		}
 	}
diff --git a/source/simulation2/components/CCmpRangeManager.cpp b/source/simulation2/components/CCmpRangeManager.cpp
index b01dd110bb..68dff4613a 100644
--- a/source/simulation2/components/CCmpRangeManager.cpp
+++ b/source/simulation2/components/CCmpRangeManager.cpp
@@ -649,7 +649,7 @@ public:
 	 */
 	void ExecuteActiveQueries()
 	{
-		PROFILE("ExecuteActiveQueries");
+		PROFILE3("ExecuteActiveQueries");
 
 		// Store a queue of all messages before sending any, so we can assume
 		// no entities will move until we've finished checking all the ranges
diff --git a/source/tools/atlas/GameInterface/MessagePasserImpl.cpp b/source/tools/atlas/GameInterface/MessagePasserImpl.cpp
index 9ea2df6629..9f9e04b071 100644
--- a/source/tools/atlas/GameInterface/MessagePasserImpl.cpp
+++ b/source/tools/atlas/GameInterface/MessagePasserImpl.cpp
@@ -24,6 +24,7 @@
 
 #include "lib/timer.h"
 #include "lib/rand.h"
+#include "lib/posix/posix_filesystem.h"
 
 using namespace AtlasMessage;