diff --git a/pcsx2/GS/GS.cpp b/pcsx2/GS/GS.cpp
index e1c6803ef2..ce2809c4cf 100644
--- a/pcsx2/GS/GS.cpp
+++ b/pcsx2/GS/GS.cpp
@@ -35,6 +35,7 @@
 #include "pcsx2/Config.h"
 #include "pcsx2/Host.h"
 #include "pcsx2/HostDisplay.h"
+#include "pcsx2/GS.h"
 #ifdef PCSX2_CORE
 #include "pcsx2/HostSettings.h"
 #endif
@@ -629,26 +630,27 @@ void GSgetInternalResolution(int* width, int* height)
 
 void GSgetStats(std::string& info)
 {
-	GSPerfMon& pm = s_gs->m_perfmon;
+	GSPerfMon& pm = g_perfmon;
 
 	const char* api_name = HostDisplay::RenderAPIToString(s_render_api);
 
 	if (GSConfig.Renderer == GSRendererType::SW)
 	{
-		int sum = 0;
-		for (int i = 0; i < 16; i++)
-			sum += pm.CPU(GSPerfMon::WorkerDraw0 + i);
+		float sum = 0.0f;
+		for (int i = GSPerfMon::WorkerDraw0; i < GSPerfMon::TimerLast; i++)
+			sum += pm.GetTimer(static_cast<GSPerfMon::timer_t>(i));
 
-		const double fps = 1000.0f / pm.Get(GSPerfMon::Frame);
+		const double fps = GetVerticalFrequency();
 		const double fillrate = pm.Get(GSPerfMon::Fillrate);
-		info = format("%d S | %d P | %d D | %.2f U | %.2f D | %.2f mpps | %d%% WCPU",
+		info = format("%s SW | %d S | %d P | %d D | %.2f U | %.2f D | %.2f mpps | %d%% WCPU",
+			api_name,
 			(int)pm.Get(GSPerfMon::SyncPoint),
 			(int)pm.Get(GSPerfMon::Prim),
 			(int)pm.Get(GSPerfMon::Draw),
 			pm.Get(GSPerfMon::Swizzle) / 1024,
 			pm.Get(GSPerfMon::Unswizzle) / 1024,
 			fps * fillrate / (1024 * 1024),
-			sum);
+			static_cast<int>(std::lround(sum)));
 	}
 	else if (GSConfig.Renderer == GSRendererType::Null)
 	{
@@ -656,12 +658,14 @@ void GSgetStats(std::string& info)
 	}
 	else
 	{
-		info = format("%d S | %d P | %d D | %.2f U | %.2f D",
-			(int)pm.Get(GSPerfMon::SyncPoint),
+		info = format("%s HW | %d P | %d D | %d DC | %d RB | %d TC | %d TU",
+			api_name,
 			(int)pm.Get(GSPerfMon::Prim),
 			(int)pm.Get(GSPerfMon::Draw),
-			pm.Get(GSPerfMon::Swizzle) / 1024,
-			pm.Get(GSPerfMon::Unswizzle) / 1024);
+			(int)std::ceil(pm.Get(GSPerfMon::DrawCalls)),
+			(int)std::ceil(pm.Get(GSPerfMon::Readbacks)),
+			(int)std::ceil(pm.Get(GSPerfMon::TextureCopies)),
+			(int)std::ceil(pm.Get(GSPerfMon::TextureUploads)));
 	}
 }
 
diff --git a/pcsx2/GS/GSPerfMon.cpp b/pcsx2/GS/GSPerfMon.cpp
index be64adfad5..921a404ce1 100644
--- a/pcsx2/GS/GSPerfMon.cpp
+++ b/pcsx2/GS/GSPerfMon.cpp
@@ -15,6 +15,9 @@
 
 #include "PrecompiledHeader.h"
 #include "GSPerfMon.h"
+#include "GS.h"
+
+GSPerfMon g_perfmon;
 
 GSPerfMon::GSPerfMon()
 	: m_frame(0)
@@ -23,41 +26,15 @@ GSPerfMon::GSPerfMon()
 {
 	memset(m_counters, 0, sizeof(m_counters));
 	memset(m_stats, 0, sizeof(m_stats));
+	memset(m_timer_stats, 0, sizeof(m_timer_stats));
 	memset(m_total, 0, sizeof(m_total));
 	memset(m_begin, 0, sizeof(m_begin));
 }
 
-void GSPerfMon::Put(counter_t c, double val)
+void GSPerfMon::EndFrame()
 {
-#ifndef DISABLE_PERF_MON
-	if (c == Frame)
-	{
-#if defined(__unix__) || defined(__APPLE__)
-		struct timespec ts;
-# ifdef CLOCK_MONOTONIC_RAW
-		clock_gettime(CLOCK_MONOTONIC_RAW, &ts);
-# else
-		clock_gettime(CLOCK_MONOTONIC, &ts);
-# endif
-		u64 now = (u64)ts.tv_sec * (u64)1e6 + (u64)ts.tv_nsec / (u64)1e3;
-#else
-		clock_t now = clock();
-#endif
-
-		if (m_lastframe != 0)
-		{
-			m_counters[c] += (now - m_lastframe) * 1000 / CLOCKS_PER_SEC;
-		}
-
-		m_lastframe = now;
-		m_frame++;
-		m_count++;
-	}
-	else
-	{
-		m_counters[c] += val;
-	}
-#endif
+	m_frame++;
+	m_count++;
 }
 
 void GSPerfMon::Update()
@@ -71,6 +48,30 @@ void GSPerfMon::Update()
 		}
 
 		m_count = 0;
+
+		// Update CPU usage for SW renderer.
+		if (GSConfig.Renderer == GSRendererType::SW)
+		{
+			const u64 current = __rdtsc();
+
+			for (size_t i = WorkerDraw0; i < TimerLast; i++)
+			{
+				if (m_begin[i] == 0)
+				{
+					m_timer_stats[i] = 0.0f;
+					continue;
+				}
+
+				m_timer_stats[i] =
+					static_cast<float>(static_cast<double>(m_total[i]) / static_cast<double>(current - m_begin[i])
+						* 100.0);
+
+				m_begin[i] = 0;
+				m_start[i] = 0;
+				m_total[i] = 0;
+			}
+		}
+
 	}
 
 	memset(m_counters, 0, sizeof(m_counters));
@@ -99,17 +100,3 @@ void GSPerfMon::Stop(int timer)
 	}
 #endif
 }
-
-int GSPerfMon::CPU(int timer, bool reset)
-{
-	int percent = (int)(100 * m_total[timer] / (__rdtsc() - m_begin[timer]));
-
-	if (reset)
-	{
-		m_begin[timer] = 0;
-		m_start[timer] = 0;
-		m_total[timer] = 0;
-	}
-
-	return percent;
-}
diff --git a/pcsx2/GS/GSPerfMon.h b/pcsx2/GS/GSPerfMon.h
index f7cc7b0009..02e0bb291d 100644
--- a/pcsx2/GS/GSPerfMon.h
+++ b/pcsx2/GS/GSPerfMon.h
@@ -29,20 +29,26 @@ public:
 
 	enum counter_t
 	{
-		Frame,
 		Prim,
 		Draw,
+		DrawCalls,
+		Readbacks,
 		Swizzle,
 		Unswizzle,
 		Fillrate,
 		Quad,
 		SyncPoint,
 		CounterLast,
+
+		// Reused counters for HW.
+		TextureCopies = Fillrate,
+		TextureUploads = SyncPoint,
 	};
 
 protected:
 	double m_counters[CounterLast];
 	double m_stats[CounterLast];
+	float m_timer_stats[TimerLast];
 	u64 m_begin[TimerLast], m_total[TimerLast], m_start[TimerLast];
 	u64 m_frame;
 	clock_t m_lastframe;
@@ -55,14 +61,15 @@ public:
 
 	void SetFrame(u64 frame) { m_frame = frame; }
 	u64 GetFrame() { return m_frame; }
+	void EndFrame();
 
-	void Put(counter_t c, double val = 0);
+	void Put(counter_t c, double val = 0) { m_counters[c] += val; }
 	double Get(counter_t c) { return m_stats[c]; }
+	float GetTimer(timer_t t) { return m_timer_stats[t]; }
 	void Update();
 
 	void Start(int timer = Main);
 	void Stop(int timer = Main);
-	int CPU(int timer = Main, bool reset = true);
 };
 
 class GSPerfMonAutoTimer
@@ -78,3 +85,5 @@ public:
 	}
 	~GSPerfMonAutoTimer() { m_pm->Stop(m_timer); }
 };
+
+extern GSPerfMon g_perfmon;
\ No newline at end of file
diff --git a/pcsx2/GS/GSState.cpp b/pcsx2/GS/GSState.cpp
index b443036edd..634b571eb5 100644
--- a/pcsx2/GS/GSState.cpp
+++ b/pcsx2/GS/GSState.cpp
@@ -1397,7 +1397,7 @@ void GSState::FlushWrite()
 
 	m_tr.start += len;
 
-	m_perfmon.Put(GSPerfMon::Swizzle, len);
+	g_perfmon.Put(GSPerfMon::Swizzle, len);
 }
 
 void GSState::FlushPrim()
@@ -1477,8 +1477,8 @@ void GSState::FlushPrim()
 
 		m_context->RestoreReg();
 
-		m_perfmon.Put(GSPerfMon::Draw, 1);
-		m_perfmon.Put(GSPerfMon::Prim, m_index.tail / GSUtil::GetVertexCount(PRIM->PRIM));
+		g_perfmon.Put(GSPerfMon::Draw, 1);
+		g_perfmon.Put(GSPerfMon::Prim, m_index.tail / GSUtil::GetVertexCount(PRIM->PRIM));
 
 		m_index.tail = 0;
 
@@ -1557,7 +1557,7 @@ void GSState::Write(const u8* mem, int len)
 
 		m_tr.start = m_tr.end = m_tr.total;
 
-		m_perfmon.Put(GSPerfMon::Swizzle, len);
+		g_perfmon.Put(GSPerfMon::Swizzle, len);
 	}
 	else
 	{
@@ -1791,7 +1791,7 @@ void GSState::SoftReset(u32 mask)
 
 void GSState::ReadFIFO(u8* mem, int size)
 {
-	GSPerfMonAutoTimer pmat(&m_perfmon);
+	GSPerfMonAutoTimer pmat(&g_perfmon);
 
 	Flush();
 
@@ -1811,7 +1811,7 @@ template void GSState::Transfer<3>(const u8* mem, u32 size);
 template <int index>
 void GSState::Transfer(const u8* mem, u32 size)
 {
-	GSPerfMonAutoTimer pmat(&m_perfmon);
+	GSPerfMonAutoTimer pmat(&g_perfmon);
 
 	const u8* start = mem;
 
@@ -2208,7 +2208,7 @@ int GSState::Defrost(const freezeData* fd)
 
 	UpdateScissor();
 
-	m_perfmon.SetFrame(5000);
+	g_perfmon.SetFrame(5000);
 
 	return 0;
 }
diff --git a/pcsx2/GS/GSState.h b/pcsx2/GS/GSState.h
index ceb482187c..3fe46f60f8 100644
--- a/pcsx2/GS/GSState.h
+++ b/pcsx2/GS/GSState.h
@@ -212,7 +212,6 @@ public:
 	GSLocalMemory m_mem;
 	GSDrawingEnvironment m_env;
 	GSDrawingContext* m_context;
-	GSPerfMon m_perfmon;
 	u32 m_crc;
 	CRC::Game m_game;
 	std::unique_ptr<GSDumpBase> m_dump;
diff --git a/pcsx2/GS/Renderers/Common/GSRenderer.cpp b/pcsx2/GS/Renderers/Common/GSRenderer.cpp
index c600a55d9f..c651814e48 100644
--- a/pcsx2/GS/Renderers/Common/GSRenderer.cpp
+++ b/pcsx2/GS/Renderers/Common/GSRenderer.cpp
@@ -384,15 +384,13 @@ static GSVector4 CalculateDrawRect(s32 window_width, s32 window_height, s32 text
 
 void GSRenderer::VSync(u32 field)
 {
-	GSPerfMonAutoTimer pmat(&m_perfmon);
-
-	m_perfmon.Put(GSPerfMon::Frame);
+	GSPerfMonAutoTimer pmat(&g_perfmon);
 
 	Flush();
 
 	if (s_dump && s_n >= s_saven)
 	{
-		m_regs->Dump(root_sw + format("%05d_f%lld_gs_reg.txt", s_n, m_perfmon.GetFrame()));
+		m_regs->Dump(root_sw + format("%05d_f%lld_gs_reg.txt", s_n, g_perfmon.GetFrame()));
 	}
 
 	g_gs_device->AgePool();
@@ -409,8 +407,9 @@ void GSRenderer::VSync(u32 field)
 		return;
 	}
 
-	if ((m_perfmon.GetFrame() & 0x1f) == 0)
-		m_perfmon.Update();
+	g_perfmon.EndFrame();
+	if ((g_perfmon.GetFrame() & 0x1f) == 0)
+		g_perfmon.Update();
 
 	g_gs_device->ResetAPIState();
 	if (Host::BeginPresentFrame(false))
diff --git a/pcsx2/GS/Renderers/DX11/GSDevice11.cpp b/pcsx2/GS/Renderers/DX11/GSDevice11.cpp
index e91f3b17d7..cdaac46bba 100644
--- a/pcsx2/GS/Renderers/DX11/GSDevice11.cpp
+++ b/pcsx2/GS/Renderers/DX11/GSDevice11.cpp
@@ -18,6 +18,7 @@
 #include "GSDevice11.h"
 #include "GS/Renderers/DX11/D3D.h"
 #include "GS/GSExtra.h"
+#include "GS/GSPerfMon.h"
 #include "GS/GSUtil.h"
 #include "Host.h"
 #include "HostDisplay.h"
@@ -370,6 +371,8 @@ void GSDevice11::RestoreAPIState()
 
 void GSDevice11::BeforeDraw()
 {
+	g_perfmon.Put(GSPerfMon::DrawCalls, 1);
+
 	// DX can't read from the FB
 	// So let's copy it and send that to the shader instead
 
@@ -546,6 +549,8 @@ bool GSDevice11::DownloadTexture(GSTexture* src, const GSVector4i& rect, GSTextu
 {
 	ASSERT(src);
 	ASSERT(!m_download_tex);
+	g_perfmon.Put(GSPerfMon::Readbacks, 1);
+
 	m_download_tex.reset(static_cast<GSTexture11*>(CreateOffscreen(rect.width(), rect.height(), src->GetFormat())));
 	if (!m_download_tex)
 		return false;
@@ -570,6 +575,8 @@ void GSDevice11::CopyRect(GSTexture* sTex, GSTexture* dTex, const GSVector4i& r)
 		return;
 	}
 
+	g_perfmon.Put(GSPerfMon::TextureCopies, 1);
+
 	D3D11_BOX box = {(UINT)r.left, (UINT)r.top, 0U, (UINT)r.right, (UINT)r.bottom, 1U};
 
 	// DX api isn't happy if we pass a box for depth copy
diff --git a/pcsx2/GS/Renderers/DX11/GSTexture11.cpp b/pcsx2/GS/Renderers/DX11/GSTexture11.cpp
index c057c5db27..eb03badbd1 100644
--- a/pcsx2/GS/Renderers/DX11/GSTexture11.cpp
+++ b/pcsx2/GS/Renderers/DX11/GSTexture11.cpp
@@ -16,6 +16,7 @@
 #include "PrecompiledHeader.h"
 #include "GSTexture11.h"
 #include "GS/GSPng.h"
+#include "GS/GSPerfMon.h"
 
 GSTexture11::GSTexture11(wil::com_ptr_nothrow<ID3D11Texture2D> texture, GSTexture::Format format)
 	: m_texture(std::move(texture)), m_layer(0)
@@ -56,6 +57,8 @@ bool GSTexture11::Update(const GSVector4i& r, const void* data, int pitch, int l
 
 	if (m_dev && m_texture)
 	{
+		g_perfmon.Put(GSPerfMon::TextureUploads, 1);
+
 		D3D11_BOX box = {(UINT)r.left, (UINT)r.top, 0U, (UINT)r.right, (UINT)r.bottom, 1U};
 		UINT subresource = layer; // MipSlice + (ArraySlice * MipLevels).
 
diff --git a/pcsx2/GS/Renderers/HW/GSRendererHW.cpp b/pcsx2/GS/Renderers/HW/GSRendererHW.cpp
index 2787001904..ae96a034c8 100644
--- a/pcsx2/GS/Renderers/HW/GSRendererHW.cpp
+++ b/pcsx2/GS/Renderers/HW/GSRendererHW.cpp
@@ -331,7 +331,7 @@ GSTexture* GSRendererHW::GetOutput(int i, int& y_offset)
 		{
 			if (s_savef && s_n >= s_saven)
 			{
-				t->Save(m_dump_root + format("%05d_f%lld_fr%d_%05x_%s.bmp", s_n, m_perfmon.GetFrame(), i, (int)TEX0.TBP0, psm_str(TEX0.PSM)));
+				t->Save(m_dump_root + format("%05d_f%lld_fr%d_%05x_%s.bmp", s_n, g_perfmon.GetFrame(), i, (int)TEX0.TBP0, psm_str(TEX0.PSM)));
 			}
 		}
 #endif
@@ -354,7 +354,7 @@ GSTexture* GSRendererHW::GetFeedbackOutput()
 
 #ifdef ENABLE_OGL_DEBUG
 	if (s_dump && s_savef && s_n >= s_saven)
-		t->Save(m_dump_root + format("%05d_f%lld_fr%d_%05x_%s.bmp", s_n, m_perfmon.GetFrame(), 3, (int)TEX0.TBP0, psm_str(TEX0.PSM)));
+		t->Save(m_dump_root + format("%05d_f%lld_fr%d_%05x_%s.bmp", s_n, g_perfmon.GetFrame(), 3, (int)TEX0.TBP0, psm_str(TEX0.PSM)));
 #endif
 
 	return t;
@@ -1513,7 +1513,7 @@ void GSRendererHW::Draw()
 
 	if (s_dump)
 	{
-		const u64 frame = m_perfmon.GetFrame();
+		const u64 frame = g_perfmon.GetFrame();
 
 		std::string s;
 
@@ -1680,7 +1680,7 @@ void GSRendererHW::Draw()
 
 	if (s_dump)
 	{
-		const u64 frame = m_perfmon.GetFrame();
+		const u64 frame = g_perfmon.GetFrame();
 
 		std::string s;
 
diff --git a/pcsx2/GS/Renderers/HW/GSTextureCache.cpp b/pcsx2/GS/Renderers/HW/GSTextureCache.cpp
index ced88e4a16..670579a2cb 100644
--- a/pcsx2/GS/Renderers/HW/GSTextureCache.cpp
+++ b/pcsx2/GS/Renderers/HW/GSTextureCache.cpp
@@ -1956,7 +1956,7 @@ void GSTextureCache::Source::Update(const GSVector4i& rect, int layer)
 
 	if (blocks > 0)
 	{
-		m_renderer->m_perfmon.Put(GSPerfMon::Unswizzle, bs.x * bs.y * blocks << (m_palette ? 2 : 0));
+		g_perfmon.Put(GSPerfMon::Unswizzle, bs.x * bs.y * blocks << (m_palette ? 2 : 0));
 
 		Flush(m_write.count, layer);
 	}
diff --git a/pcsx2/GS/Renderers/OpenGL/GSDeviceOGL.cpp b/pcsx2/GS/Renderers/OpenGL/GSDeviceOGL.cpp
index 9491ee1ec1..9d1280e068 100644
--- a/pcsx2/GS/Renderers/OpenGL/GSDeviceOGL.cpp
+++ b/pcsx2/GS/Renderers/OpenGL/GSDeviceOGL.cpp
@@ -680,6 +680,7 @@ void GSDeviceOGL::RestoreAPIState()
 
 void GSDeviceOGL::DrawPrimitive()
 {
+	g_perfmon.Put(GSPerfMon::DrawCalls, 1);
 	glDrawArrays(m_draw_topology, m_vertex.start, m_vertex.count);
 }
 
@@ -687,6 +688,7 @@ void GSDeviceOGL::DrawIndexedPrimitive()
 {
 	if (!m_disable_hw_gl_draw)
 	{
+		g_perfmon.Put(GSPerfMon::DrawCalls, 1);
 		glDrawElementsBaseVertex(m_draw_topology, static_cast<u32>(m_index.count), GL_UNSIGNED_INT,
 			reinterpret_cast<void*>(static_cast<u32>(m_index.start) * sizeof(u32)), static_cast<GLint>(m_vertex.start));
 	}
@@ -698,6 +700,7 @@ void GSDeviceOGL::DrawIndexedPrimitive(int offset, int count)
 
 	if (!m_disable_hw_gl_draw)
 	{
+		g_perfmon.Put(GSPerfMon::DrawCalls, 1);
 		glDrawElementsBaseVertex(m_draw_topology, count, GL_UNSIGNED_INT,
 			reinterpret_cast<void*>((static_cast<u32>(m_index.start) + static_cast<u32>(offset)) * sizeof(u32)),
 			static_cast<GLint>(m_vertex.start));
@@ -1105,6 +1108,7 @@ std::string GSDeviceOGL::GetPSSource(PSSelector sel)
 bool GSDeviceOGL::DownloadTexture(GSTexture* src, const GSVector4i& rect, GSTexture::GSMap& out_map)
 {
 	ASSERT(src);
+	g_perfmon.Put(GSPerfMon::Readbacks, 1);
 
 	GSTextureOGL* srcgl = static_cast<GSTextureOGL*>(src);
 
@@ -1116,6 +1120,7 @@ bool GSDeviceOGL::DownloadTexture(GSTexture* src, const GSVector4i& rect, GSText
 void GSDeviceOGL::BlitRect(GSTexture* sTex, const GSVector4i& r, const GSVector2i& dsize, bool at_origin, bool linear)
 {
 	GL_PUSH(format("CopyRectConv from %d", static_cast<GSTextureOGL*>(sTex)->GetID()).c_str());
+	g_perfmon.Put(GSPerfMon::TextureCopies, 1);
 
 	// NOTE: This previously used glCopyTextureSubImage2D(), but this appears to leak memory in
 	// the loading screens of Evolution Snowboarding in Intel/NVIDIA drivers.
@@ -1153,6 +1158,7 @@ void GSDeviceOGL::CopyRect(GSTexture* sTex, GSTexture* dTex, const GSVector4i& r
 #endif
 
 	dTex->CommitRegion(GSVector2i(r.z, r.w));
+	g_perfmon.Put(GSPerfMon::TextureCopies, 1);
 
 	ASSERT(GLExtension::Has("GL_ARB_copy_image") && glCopyImageSubData);
 	glCopyImageSubData(sid, GL_TEXTURE_2D,
diff --git a/pcsx2/GS/Renderers/OpenGL/GSTextureOGL.cpp b/pcsx2/GS/Renderers/OpenGL/GSTextureOGL.cpp
index 3df8de2377..c9c0f187d5 100644
--- a/pcsx2/GS/Renderers/OpenGL/GSTextureOGL.cpp
+++ b/pcsx2/GS/Renderers/OpenGL/GSTextureOGL.cpp
@@ -17,6 +17,7 @@
 #include <limits.h>
 #include "GSTextureOGL.h"
 #include "GLState.h"
+#include "GS/GSPerfMon.h"
 #include "GS/GSPng.h"
 
 #ifdef ENABLE_OGL_DEBUG_MEM_BW
@@ -393,6 +394,7 @@ bool GSTextureOGL::Update(const GSVector4i& r, const void* data, int pitch, int
 #endif
 
 	GL_PUSH("Upload Texture %d", m_texture_id);
+	g_perfmon.Put(GSPerfMon::TextureUploads, 1);
 
 	// The easy solution without PBO
 #if 0
@@ -449,6 +451,7 @@ bool GSTextureOGL::Map(GSMap& m, const GSVector4i* _r, int layer)
 	if (m_type == Type::Texture || m_type == Type::RenderTarget)
 	{
 		GL_PUSH_("Upload Texture %d", m_texture_id); // POP is in Unmap
+		g_perfmon.Put(GSPerfMon::TextureUploads, 1);
 
 		m_clean = false;
 
diff --git a/pcsx2/GS/Renderers/SW/GSRendererSW.cpp b/pcsx2/GS/Renderers/SW/GSRendererSW.cpp
index 65d4eda374..7ad2839dfa 100644
--- a/pcsx2/GS/Renderers/SW/GSRendererSW.cpp
+++ b/pcsx2/GS/Renderers/SW/GSRendererSW.cpp
@@ -35,7 +35,7 @@ GSRendererSW::GSRendererSW(int threads)
 
 	memset(m_texture, 0, sizeof(m_texture));
 
-	m_rl = GSRasterizerList::Create<GSDrawScanline>(threads, &m_perfmon);
+	m_rl = GSRasterizerList::Create<GSDrawScanline>(threads, &g_perfmon);
 
 	m_output = (u8*)_aligned_malloc(1024 * 1024 * sizeof(u32), 32);
 
@@ -97,7 +97,7 @@ void GSRendererSW::VSync(u32 field)
 
 	if (0) if (LOG)
 	{
-		fprintf(s_fp, "%llu\n", m_perfmon.GetFrame());
+		fprintf(s_fp, "%llu\n", g_perfmon.GetFrame());
 
 		GSVector4i dr = GetDisplayRect();
 		GSVector4i fr = GetFrameRect();
@@ -162,7 +162,7 @@ GSTexture* GSRendererSW::GetOutput(int i, int& y_offset)
 		{
 			if (s_savef && s_n >= s_saven)
 			{
-				m_texture[i]->Save(m_dump_root + format("%05d_f%lld_fr%d_%05x_%s.bmp", s_n, m_perfmon.GetFrame(), i, (int)DISPFB.Block(), psm_str(DISPFB.PSM)));
+				m_texture[i]->Save(m_dump_root + format("%05d_f%lld_fr%d_%05x_%s.bmp", s_n, g_perfmon.GetFrame(), i, (int)DISPFB.Block(), psm_str(DISPFB.PSM)));
 			}
 		}
 	}
@@ -362,7 +362,7 @@ void GSRendererSW::Draw()
 
 	sd->scissor = scissor;
 	sd->bbox = bbox;
-	sd->frame = m_perfmon.GetFrame();
+	sd->frame = g_perfmon.GetFrame();
 
 	if (!GetScanlineGlobalData(sd))
 	{
@@ -434,7 +434,7 @@ void GSRendererSW::Draw()
 	{
 		Sync(2);
 
-		u64 frame = m_perfmon.GetFrame();
+		u64 frame = g_perfmon.GetFrame();
 		// Dump the texture in 32 bits format. It helps to debug texture shuffle effect
 		// It will breaks the few games that really uses 16 bits RT
 		bool texture_shuffle = ((context->FRAME.PSM & 0x2) && ((context->TEX0.PSM & 3) == 2) && (m_vt.m_primclass == GS_SPRITE_CLASS));
@@ -581,7 +581,7 @@ void GSRendererSW::Sync(int reason)
 {
 	//printf("sync %d\n", reason);
 
-	GSPerfMonAutoTimer pmat(&m_perfmon, GSPerfMon::Sync);
+	GSPerfMonAutoTimer pmat(&g_perfmon, GSPerfMon::Sync);
 
 	u64 t = __rdtsc();
 
@@ -593,14 +593,14 @@ void GSRendererSW::Sync(int reason)
 
 		if (s_save)
 		{
-			s = format("%05d_f%lld_rt1_%05x_%s.bmp", s_n, m_perfmon.GetFrame(), m_context->FRAME.Block(), psm_str(m_context->FRAME.PSM));
+			s = format("%05d_f%lld_rt1_%05x_%s.bmp", s_n, g_perfmon.GetFrame(), m_context->FRAME.Block(), psm_str(m_context->FRAME.PSM));
 
 			m_mem.SaveBMP(m_dump_root + s, m_context->FRAME.Block(), m_context->FRAME.FBW, m_context->FRAME.PSM, GetFrameRect().width(), 512);
 		}
 
 		if (s_savez)
 		{
-			s = format("%05d_f%lld_zb1_%05x_%s.bmp", s_n, m_perfmon.GetFrame(), m_context->ZBUF.Block(), psm_str(m_context->ZBUF.PSM));
+			s = format("%05d_f%lld_zb1_%05x_%s.bmp", s_n, g_perfmon.GetFrame(), m_context->ZBUF.Block(), psm_str(m_context->ZBUF.PSM));
 
 			m_mem.SaveBMP(m_dump_root + s, m_context->ZBUF.Block(), m_context->FRAME.FBW, m_context->ZBUF.PSM, GetFrameRect().width(), 512);
 		}
@@ -616,7 +616,7 @@ void GSRendererSW::Sync(int reason)
 		fflush(s_fp);
 	}
 
-	m_perfmon.Put(GSPerfMon::Fillrate, pixels);
+	g_perfmon.Put(GSPerfMon::Fillrate, pixels);
 }
 
 void GSRendererSW::InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& r)
@@ -1541,7 +1541,7 @@ void GSRendererSW::SharedData::UpdateSource()
 
 	if (m_parent->s_dump)
 	{
-		u64 frame = m_parent->m_perfmon.GetFrame();
+		u64 frame = g_perfmon.GetFrame();
 
 		std::string s;
 
diff --git a/pcsx2/GS/Renderers/SW/GSTextureCacheSW.cpp b/pcsx2/GS/Renderers/SW/GSTextureCacheSW.cpp
index 6cd5954271..c326fa06d5 100644
--- a/pcsx2/GS/Renderers/SW/GSTextureCacheSW.cpp
+++ b/pcsx2/GS/Renderers/SW/GSTextureCacheSW.cpp
@@ -281,7 +281,7 @@ bool GSTextureCacheSW::Texture::Update(const GSVector4i& rect)
 
 	if (blocks > 0)
 	{
-		m_state->m_perfmon.Put(GSPerfMon::Unswizzle, bs.x * bs.y * blocks << shift);
+		g_perfmon.Put(GSPerfMon::Unswizzle, bs.x * bs.y * blocks << shift);
 	}
 
 	return true;
diff --git a/pcsx2/GS/config.h b/pcsx2/GS/config.h
index bdb84c5b10..26c309b9af 100644
--- a/pcsx2/GS/config.h
+++ b/pcsx2/GS/config.h
@@ -39,6 +39,4 @@
 //#define ENABLE_EXTRA_LOG // print extra log
 #endif
 
-#if (defined(__unix__) || defined(__APPLE__)) && !(defined(_DEBUG) || defined(_DEVEL))
-#define DISABLE_PERF_MON // Burn cycle for nothing in release mode
-#endif
+//#define DISABLE_PERF_MON // Burn cycle for nothing in release mode