diff --git a/GPU/Software/BinManager.cpp b/GPU/Software/BinManager.cpp
index 4d1233a617..faa1f648c0 100644
--- a/GPU/Software/BinManager.cpp
+++ b/GPU/Software/BinManager.cpp
@@ -20,6 +20,8 @@
 #include <mutex>
 #include "Common/Profiler/Profiler.h"
 #include "Common/Thread/ThreadManager.h"
+#include "Common/TimeUtil.h"
+#include "Core/System.h"
 #include "GPU/Software/BinManager.h"
 #include "GPU/Software/Rasterizer.h"
 #include "GPU/Software/RasterizerRectangle.h"
@@ -87,7 +89,7 @@ static inline void DrawBinItem(const BinItem &item, const RasterizerState &state
 
 class DrawBinItemsTask : public Task {
 public:
-	DrawBinItemsTask(BinWaitable *notify, BinQueue<BinItem, 1024> &items, std::atomic<bool> &status, const BinQueue<RasterizerState, 64> &states)
+	DrawBinItemsTask(BinWaitable *notify, BinManager::BinItemQueue &items, std::atomic<bool> &status, const BinManager::BinStateQueue &states)
 		: notify_(notify), items_(items), status_(status), states_(states) {
 	}
 
@@ -113,9 +115,9 @@ private:
 	}
 
 	BinWaitable *notify_;
-	BinQueue<BinItem, 1024> &items_;
+	BinManager::BinItemQueue &items_;
 	std::atomic<bool> &status_;
-	const BinQueue<RasterizerState, 64> &states_;
+	const BinManager::BinStateQueue &states_;
 };
 
 BinManager::BinManager() {
@@ -127,6 +129,13 @@ BinManager::BinManager() {
 	waitable_ = new BinWaitable();
 	for (auto &s : taskStatus_)
 		s = false;
+
+	int maxInitTasks = std::min(g_threadManager.GetNumLooperThreads(), MAX_POSSIBLE_TASKS);
+	for (int i = 0; i < maxInitTasks; ++i)
+		taskQueues_[i].Setup();
+	states_.Setup();
+	cluts_.Setup();
+	queue_.Setup();
 }
 
 BinManager::~BinManager() {
@@ -136,7 +145,7 @@ BinManager::~BinManager() {
 void BinManager::UpdateState() {
 	PROFILE_THIS_SCOPE("bin_state");
 	if (states_.Full())
-		Flush();
+		Flush("states");
 	stateIndex_ = (int)states_.Push(RasterizerState());
 	ComputeRasterizerState(&states_[stateIndex_]);
 	states_[stateIndex_].samplerID.cached.clut = cluts_[clutIndex_].readable;
@@ -165,12 +174,12 @@ void BinManager::UpdateState() {
 	// We don't want to overlap wrong, so flush any pending.
 	if (maxTasks_ != newMaxTasks) {
 		maxTasks_ = newMaxTasks;
-		Flush();
+		Flush("selfrender");
 	}
 
 	// Our bin sizes are based on offset, so if that changes we have to flush.
 	if (queueOffsetX_ != gstate.getOffsetX16() || queueOffsetY_ != gstate.getOffsetY16()) {
-		Flush();
+		Flush("offset");
 		queueOffsetX_ = gstate.getOffsetX16();
 		queueOffsetY_ = gstate.getOffsetY16();
 	}
@@ -179,7 +188,7 @@ void BinManager::UpdateState() {
 void BinManager::UpdateClut(const void *src) {
 	PROFILE_THIS_SCOPE("bin_clut");
 	if (cluts_.Full())
-		Flush();
+		Flush("cluts");
 	clutIndex_ = (int)cluts_.Push(BinClut());
 	memcpy(cluts_[clutIndex_].readable, src, sizeof(BinClut));
 }
@@ -323,7 +332,10 @@ void BinManager::Drain() {
 	}
 }
 
-void BinManager::Flush() {
+void BinManager::Flush(const char *reason) {
+	double st;
+	if (coreCollectDebugStats)
+		st = time_now_d();
 	Drain();
 	waitable_->Wait();
 	taskRanges_.clear();
@@ -341,6 +353,59 @@ void BinManager::Flush() {
 	queueRange_.y2 = 0;
 	queueOffsetX_ = -1;
 	queueOffsetY_ = -1;
+
+	if (coreCollectDebugStats) {
+		double et = time_now_d();
+		flushReasonTimes_[reason] += et - st;
+		if (et - st > slowestFlushTime_) {
+			slowestFlushTime_ = et - st;
+			slowestFlushReason_ = reason;
+		}
+	}
+}
+
+void BinManager::GetStats(char *buffer, size_t bufsize) {
+	double allTotal = 0.0;
+	double slowestTotalTime = 0.0;
+	const char *slowestTotalReason = nullptr;
+	for (auto &it : flushReasonTimes_) {
+		if (it.second > slowestTotalTime) {
+			slowestTotalTime = it.second;
+			slowestTotalReason = it.first;
+		}
+		allTotal += it.second;
+	}
+
+	// Many games are 30 FPS, so check last frame too for better stats.
+	double recentTotal = allTotal;
+	double slowestRecentTime = slowestTotalTime;
+	const char *slowestRecentReason = slowestTotalReason;
+	for (auto &it : lastFlushReasonTimes_) {
+		if (it.second > slowestRecentTime) {
+			slowestRecentTime = it.second;
+			slowestRecentReason = it.first;
+		}
+		recentTotal += it.second;
+	}
+
+	snprintf(buffer, bufsize,
+		"Slowest individual flush: %s (%0.4f)\n"
+		"Slowest frame flush: %s (%0.4f)\n"
+		"Slowest recent flush: %s (%0.4f)\n"
+		"Total flush time: %0.4f (%05.2f%%, last 2: %05.2f%%)\n",
+		slowestFlushReason_, slowestFlushTime_,
+		slowestTotalReason, slowestTotalTime,
+		slowestRecentReason, slowestRecentTime,
+		allTotal, allTotal * (6000.0 / 1.001), recentTotal * (3000.0 / 1.001));
+
+	constexpr int foo = sizeof(BinItem);
+}
+
+void BinManager::ResetStats() {
+	lastFlushReasonTimes_ = std::move(flushReasonTimes_);
+	flushReasonTimes_.clear();
+	slowestFlushReason_ = nullptr;
+	slowestFlushTime_ = 0.0;
 }
 
 inline BinCoords BinCoords::Intersect(const BinCoords &range) const {
diff --git a/GPU/Software/BinManager.h b/GPU/Software/BinManager.h
index eff0c4e6c9..883997ad10 100644
--- a/GPU/Software/BinManager.h
+++ b/GPU/Software/BinManager.h
@@ -18,10 +18,12 @@
 #pragma once
 
 #include <atomic>
+#include <unordered_map>
 #include "Common/Log.h"
 #include "GPU/Software/Rasterizer.h"
 
 struct BinWaitable;
+class DrawBinItemsTask;
 
 enum class BinItemType {
 	TRIANGLE,
@@ -56,13 +58,16 @@ struct BinItem {
 template <typename T, size_t N>
 struct BinQueue {
 	BinQueue() {
-		items_ = new T[N];
 		Reset();
 	}
 	~BinQueue() {
 		delete [] items_;
 	}
 
+	void Setup() {
+		items_ = new T[N];
+	}
+
 	void Reset() {
 		head_ = 0;
 		tail_ = 0;
@@ -166,17 +171,31 @@ public:
 	void AddPoint(const VertexData &v0);
 
 	void Drain();
-	void Flush();
+	void Flush(const char *reason);
+
+	void GetStats(char *buffer, size_t bufsize);
+	void ResetStats();
+
+protected:
+	static constexpr int MAX_POSSIBLE_TASKS = 64;
+	// This is about 1MB of state data.
+	static constexpr int QUEUED_STATES = 4096;
+	// These are 1KB each, so half an MB.
+	static constexpr int QUEUED_CLUTS = 512;
+	// About 320 KB, but we have usually 16 or less of them, so 5 MB - 20 MB.
+	static constexpr int QUEUED_PRIMS = 1024;
+
+	typedef BinQueue<Rasterizer::RasterizerState, QUEUED_STATES> BinStateQueue;
+	typedef BinQueue<BinClut, QUEUED_CLUTS> BinClutQueue;
+	typedef BinQueue<BinItem, QUEUED_PRIMS> BinItemQueue;
 
 private:
-	static constexpr int MAX_POSSIBLE_TASKS = 64;
-
-	BinQueue<Rasterizer::RasterizerState, 64> states_;
+	BinStateQueue states_;
 	int stateIndex_;
-	BinQueue<BinClut, 64> cluts_;
+	BinClutQueue cluts_;
 	int clutIndex_;
 	BinCoords scissor_;
-	BinQueue<BinItem, 1024> queue_;
+	BinItemQueue queue_;
 	BinCoords queueRange_;
 	int queueOffsetX_ = -1;
 	int queueOffsetY_ = -1;
@@ -184,13 +203,20 @@ private:
 	int maxTasks_ = 1;
 	bool tasksSplit_ = false;
 	std::vector<BinCoords> taskRanges_;
-	BinQueue<BinItem, 1024> taskQueues_[MAX_POSSIBLE_TASKS];
+	BinItemQueue taskQueues_[MAX_POSSIBLE_TASKS];
 	std::atomic<bool> taskStatus_[MAX_POSSIBLE_TASKS];
 	BinWaitable *waitable_ = nullptr;
 
+	std::unordered_map<const char *, double> flushReasonTimes_;
+	std::unordered_map<const char *, double> lastFlushReasonTimes_;
+	const char *slowestFlushReason_ = nullptr;
+	double slowestFlushTime_ = 0.0;
+
 	BinCoords Scissor(BinCoords range);
 	BinCoords Range(const VertexData &v0, const VertexData &v1, const VertexData &v2);
 	BinCoords Range(const VertexData &v0, const VertexData &v1);
 	BinCoords Range(const VertexData &v0);
 	void Expand(const BinCoords &range);
+
+	friend class DrawBinItemsTask;
 };
diff --git a/GPU/Software/SoftGpu.cpp b/GPU/Software/SoftGpu.cpp
index 7e786916ee..f1c1d7409c 100644
--- a/GPU/Software/SoftGpu.cpp
+++ b/GPU/Software/SoftGpu.cpp
@@ -900,7 +900,7 @@ void SoftGPU::FinishDeferred() {
 }
 
 void SoftGPU::GetStats(char *buffer, size_t bufsize) {
-	snprintf(buffer, bufsize, "SoftGPU: (N/A)");
+	drawEngine_->transformUnit.GetStats(buffer, bufsize);
 }
 
 void SoftGPU::InvalidateCache(u32 addr, int size, GPUInvalidationType type)
diff --git a/GPU/Software/TransformUnit.cpp b/GPU/Software/TransformUnit.cpp
index 5736f37ddd..e020f0a7f4 100644
--- a/GPU/Software/TransformUnit.cpp
+++ b/GPU/Software/TransformUnit.cpp
@@ -609,10 +609,16 @@ void TransformUnit::SubmitPrimitive(void* vertices, void* indices, GEPrimitiveTy
 }
 
 void TransformUnit::Flush(const char *reason) {
-	binner_->Flush();
+	binner_->Flush(reason);
 	GPUDebug::NotifyDraw();
 }
 
+void TransformUnit::GetStats(char *buffer, size_t bufsize) {
+	// TODO: More stats?
+	binner_->GetStats(buffer, bufsize);
+	binner_->ResetStats();
+}
+
 void TransformUnit::FlushIfOverlap(const char *reason, uint32_t addr, uint32_t sz) {
 	if (!Memory::IsVRAMAddress(addr))
 		return;
diff --git a/GPU/Software/TransformUnit.h b/GPU/Software/TransformUnit.h
index 79bd8bfbd5..0c277b21bc 100644
--- a/GPU/Software/TransformUnit.h
+++ b/GPU/Software/TransformUnit.h
@@ -123,6 +123,8 @@ public:
 	void FlushIfOverlap(const char *reason, uint32_t addr, uint32_t sz);
 	void NotifyClutUpdate(const void *src);
 
+	void GetStats(char *buffer, size_t bufsize);
+
 private:
 	VertexData ReadVertex(VertexReader &vreader, bool &outside_range_flag);