diff --git a/Common/GPU/OpenGL/GLFrameData.cpp b/Common/GPU/OpenGL/GLFrameData.cpp
index fa5a051d30..91300da9d5 100644
--- a/Common/GPU/OpenGL/GLFrameData.cpp
+++ b/Common/GPU/OpenGL/GLFrameData.cpp
@@ -3,6 +3,21 @@
 #include "Common/GPU/OpenGL/GLRenderManager.h"
 #include "Common/Log.h"
 
+void GLCachedReadback::Destroy(bool skipGLCalls) {
+	if (buffer && !skipGLCalls) {
+		glDeleteBuffers(1, &buffer);
+	}
+	buffer = 0;
+}
+
+void GLFrameData::Destroy(bool skipGLCalls) {
+	readbacks_.IterateMut([=](const GLReadbackKey &key, GLCachedReadback *value) {
+		value->Destroy(skipGLCalls);
+	delete value;
+		});
+	readbacks_.Clear();
+}
+
 void GLDeleter::Take(GLDeleter &other) {
 	_assert_msg_(IsEmpty(), "Deleter already has stuff");
 	shaders = std::move(other.shaders);
diff --git a/Common/GPU/OpenGL/GLFrameData.h b/Common/GPU/OpenGL/GLFrameData.h
index a50fbe7d55..3dc5487162 100644
--- a/Common/GPU/OpenGL/GLFrameData.h
+++ b/Common/GPU/OpenGL/GLFrameData.h
@@ -6,6 +6,7 @@
 #include <set>
 
 #include "Common/GPU/OpenGL/GLCommon.h"
+#include "Common/Data/Collections/Hashmaps.h"
 
 class GLRShader;
 class GLRBuffer;
@@ -35,8 +36,24 @@ public:
 	std::vector<GLPushBuffer *> pushBuffers;
 };
 
+// TODO: To be safe, should probably add some more stuff here, like format and even readback count, maybe.
+struct GLReadbackKey {
+	const GLRFramebuffer *framebuf;
+	int width;
+	int height;
+};
+
+struct GLCachedReadback {
+	GLuint buffer;  // PBO
+	size_t bufferSize;
+
+	void Destroy(bool skipGLCalls);
+};
+
 // Per-frame data, round-robin so we can overlap submission with execution of the previous frame.
 struct GLFrameData {
+	GLFrameData() : readbacks_(8) {}
+
 	bool skipSwap = false;
 
 	std::mutex fenceMutex;
@@ -49,4 +66,8 @@ struct GLFrameData {
 	GLDeleter deleter;
 	GLDeleter deleter_prev;
 	std::set<GLPushBuffer *> activePushBuffers;
+
+	DenseHashMap<GLReadbackKey, GLCachedReadback *, nullptr> readbacks_;
+
+	void Destroy(bool skipGLCalls);
 };
diff --git a/Common/GPU/OpenGL/GLRenderManager.cpp b/Common/GPU/OpenGL/GLRenderManager.cpp
index b10caa2541..e6766868da 100644
--- a/Common/GPU/OpenGL/GLRenderManager.cpp
+++ b/Common/GPU/OpenGL/GLRenderManager.cpp
@@ -47,7 +47,9 @@ GLRenderManager::~GLRenderManager() {
 	for (int i = 0; i < MAX_INFLIGHT_FRAMES; i++) {
 		_assert_(frameData_[i].deleter.IsEmpty());
 		_assert_(frameData_[i].deleter_prev.IsEmpty());
+		frameData_[i].Destroy(skipGLCalls_);
 	}
+
 	// Was anything deleted during shutdown?
 	deleter_.Perform(this, skipGLCalls_);
 	_assert_(deleter_.IsEmpty());
@@ -108,6 +110,7 @@ void GLRenderManager::ThreadEnd() {
 		// Since we're in shutdown, we should skip the GL calls on Android.
 		frameData_[i].deleter.Perform(this, skipGLCalls_);
 		frameData_[i].deleter_prev.Perform(this, skipGLCalls_);
+		frameData_[i].Destroy(skipGLCalls_);
 	}
 	deleter_.Perform(this, skipGLCalls_);
 	for (int i = 0; i < (int)steps_.size(); i++) {
@@ -292,6 +295,8 @@ void GLRenderManager::BlitFramebuffer(GLRFramebuffer *src, GLRect2D srcRect, GLR
 bool GLRenderManager::CopyFramebufferToMemory(GLRFramebuffer *src, int aspectBits, int x, int y, int w, int h, Draw::DataFormat destFormat, uint8_t *pixels, int pixelStride, Draw::ReadbackMode mode, const char *tag) {
 	_assert_(pixels);
 
+	curRenderStep_ = nullptr;
+
 	GLRStep *step = new GLRStep{ GLRStepType::READBACK };
 	step->readback.src = src;
 	step->readback.srcRect = { x, y, w, h };
@@ -301,8 +306,9 @@ bool GLRenderManager::CopyFramebufferToMemory(GLRFramebuffer *src, int aspectBit
 	step->tag = tag;
 	steps_.push_back(step);
 
-	curRenderStep_ = nullptr;
-	FlushSync();
+	if (mode == Draw::ReadbackMode::BLOCK) {
+		FlushSync();
+	}
 
 	Draw::DataFormat srcFormat;
 	if (aspectBits & GL_COLOR_BUFFER_BIT) {
@@ -316,13 +322,17 @@ bool GLRenderManager::CopyFramebufferToMemory(GLRFramebuffer *src, int aspectBit
 	} else {
 		return false;
 	}
-	queueRunner_.CopyFromReadbackBuffer(src, w, h, srcFormat, destFormat, pixelStride, pixels);
+
+	queueRunner_.CopyFromReadbackBuffer(mode == Draw::ReadbackMode::OLD_DATA_OK ? src : nullptr, w, h, srcFormat, destFormat, pixelStride, pixels);
 	return true;
 }
 
 void GLRenderManager::CopyImageToMemorySync(GLRTexture *texture, int mipLevel, int x, int y, int w, int h, Draw::DataFormat destFormat, uint8_t *pixels, int pixelStride, const char *tag) {
 	_assert_(texture);
 	_assert_(pixels);
+
+	curRenderStep_ = nullptr;
+
 	GLRStep *step = new GLRStep{ GLRStepType::READBACK_IMAGE };
 	step->readback_image.texture = texture;
 	step->readback_image.mipLevel = mipLevel;
@@ -330,7 +340,6 @@ void GLRenderManager::CopyImageToMemorySync(GLRTexture *texture, int mipLevel, i
 	step->tag = tag;
 	steps_.push_back(step);
 
-	curRenderStep_ = nullptr;
 	FlushSync();
 
 	queueRunner_.CopyFromReadbackBuffer(nullptr, w, h, Draw::DataFormat::R8G8B8A8_UNORM, destFormat, pixelStride, pixels);
diff --git a/Common/GPU/Vulkan/VulkanFrameData.h b/Common/GPU/Vulkan/VulkanFrameData.h
index 0e1344f24e..2d27f1c2d1 100644
--- a/Common/GPU/Vulkan/VulkanFrameData.h
+++ b/Common/GPU/Vulkan/VulkanFrameData.h
@@ -28,6 +28,7 @@ struct QueueProfileContext {
 
 class VKRFramebuffer;
 
+// TODO: To be safe, should probably add some more stuff here, like format and even readback count, maybe.
 struct ReadbackKey {
 	const VKRFramebuffer *framebuf;
 	int width;
diff --git a/Common/GPU/thin3d.h b/Common/GPU/thin3d.h
index 3db2d7b8db..e355ebbc91 100644
--- a/Common/GPU/thin3d.h
+++ b/Common/GPU/thin3d.h
@@ -581,9 +581,10 @@ struct DeviceCaps {
 	bool multiViewSupported;
 	bool isTilingGPU;  // This means that it benefits from correct store-ops, msaa without backing memory, etc.
 	bool sampleRateShadingSupported;
-
 	bool verySlowShaderCompiler;
 
+	bool pboSupported;  // Only makes sense in OpenGL.
+
 	// From the other backends, we can detect if D3D9 support is known bad (like on Xe) and disable it.
 	bool supportsD3D9;