diff --git a/rpcs3/Emu/RSX/VK/VKGSRender.cpp b/rpcs3/Emu/RSX/VK/VKGSRender.cpp
index ed39b205e1..5353c404df 100644
--- a/rpcs3/Emu/RSX/VK/VKGSRender.cpp
+++ b/rpcs3/Emu/RSX/VK/VKGSRender.cpp
@@ -509,17 +509,17 @@ VKGSRender::VKGSRender(utils::serial* ar) noexcept : GSRender(ar)
 	semaphore_info.sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO;
 
 	// VRAM allocation
-	m_attrib_ring_info.create(VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT, VK_ATTRIB_RING_BUFFER_SIZE_M * 0x100000, "attrib buffer", 0x400000, VK_TRUE);
-	m_fragment_env_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, "fragment env buffer");
-	m_vertex_env_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, "vertex env buffer");
-	m_fragment_texture_params_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, "fragment texture params buffer");
-	m_vertex_layout_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, "vertex layout buffer", 0x10000, VK_TRUE);
-	m_fragment_constants_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, "fragment constants buffer");
-	m_transform_constants_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_TRANSFORM_CONSTANTS_BUFFER_SIZE_M * 0x100000, "transform constants buffer");
-	m_index_buffer_ring_info.create(VK_BUFFER_USAGE_INDEX_BUFFER_BIT, VK_INDEX_RING_BUFFER_SIZE_M * 0x100000, "index buffer");
-	m_texture_upload_buffer_ring_info.create(VK_BUFFER_USAGE_TRANSFER_SRC_BIT, VK_TEXTURE_UPLOAD_RING_BUFFER_SIZE_M * 0x100000, "texture upload buffer", 32 * 0x100000);
-	m_raster_env_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, "raster env buffer");
-	m_instancing_buffer_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_TRANSFORM_CONSTANTS_BUFFER_SIZE_M * 0x100000, "instancing data buffer");
+	m_attrib_ring_info.create(VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT, VK_ATTRIB_RING_BUFFER_SIZE_M * 0x100000, vk::heap_pool_default, "attrib buffer", 0x400000, VK_TRUE);
+	m_fragment_env_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, vk::heap_pool_low_latency, "fragment env buffer");
+	m_vertex_env_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, vk::heap_pool_low_latency, "vertex env buffer");
+	m_fragment_texture_params_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, vk::heap_pool_low_latency, "fragment texture params buffer");
+	m_vertex_layout_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, vk::heap_pool_low_latency, "vertex layout buffer", 0x10000, VK_TRUE);
+	m_fragment_constants_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, vk::heap_pool_low_latency, "fragment constants buffer");
+	m_transform_constants_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_TRANSFORM_CONSTANTS_BUFFER_SIZE_M * 0x100000, vk::heap_pool_low_latency, "transform constants buffer");
+	m_index_buffer_ring_info.create(VK_BUFFER_USAGE_INDEX_BUFFER_BIT, VK_INDEX_RING_BUFFER_SIZE_M * 0x100000, vk::heap_pool_default, "index buffer");
+	m_texture_upload_buffer_ring_info.create(VK_BUFFER_USAGE_TRANSFER_SRC_BIT, VK_TEXTURE_UPLOAD_RING_BUFFER_SIZE_M * 0x100000, vk::heap_pool_default, "texture upload buffer", 32 * 0x100000);
+	m_raster_env_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, vk::heap_pool_low_latency, "raster env buffer");
+	m_instancing_buffer_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_TRANSFORM_CONSTANTS_BUFFER_SIZE_M * 0x100000, vk::heap_pool_low_latency, "instancing data buffer");
 
 	vk::data_heap_manager::register_ring_buffers
 	({
@@ -540,8 +540,8 @@ VKGSRender::VKGSRender(utils::serial* ar) noexcept : GSRender(ar)
 
 	if (shadermode == shader_mode::async_with_interpreter || shadermode == shader_mode::interpreter_only)
 	{
-		m_vertex_instructions_buffer.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, 64 * 0x100000, "vertex instructions buffer", 512 * 16);
-		m_fragment_instructions_buffer.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, 64 * 0x100000, "fragment instructions buffer", 2048);
+		m_vertex_instructions_buffer.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, 64 * 0x100000, vk::heap_pool_low_latency, "vertex instructions buffer", 512 * 16);
+		m_fragment_instructions_buffer.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, 64 * 0x100000, vk::heap_pool_low_latency, "fragment instructions buffer", 2048);
 
 		vk::data_heap_manager::register_ring_buffers
 		({
diff --git a/rpcs3/Emu/RSX/VK/VKOverlays.cpp b/rpcs3/Emu/RSX/VK/VKOverlays.cpp
index e377e1efdf..d0abe33485 100644
--- a/rpcs3/Emu/RSX/VK/VKOverlays.cpp
+++ b/rpcs3/Emu/RSX/VK/VKOverlays.cpp
@@ -42,12 +42,12 @@ namespace vk
 	{
 		if (!m_vao.heap)
 		{
-			m_vao.create(VK_BUFFER_USAGE_VERTEX_BUFFER_BIT, 1 * 0x100000, "overlays VAO", 128);
+			m_vao.create(VK_BUFFER_USAGE_VERTEX_BUFFER_BIT, 1 * 0x100000, vk::heap_pool_default, "overlays VAO", 128);
 		}
 
 		if (!m_ubo.heap && m_num_uniform_buffers > 0)
 		{
-			m_ubo.create(VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT, 8 * 0x100000, "overlays UBO", 128);
+			m_ubo.create(VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT, 8 * 0x100000, vk::heap_pool_default, "overlays UBO", 128);
 		}
 	}
 
diff --git a/rpcs3/Emu/RSX/VK/vkutils/data_heap.cpp b/rpcs3/Emu/RSX/VK/vkutils/data_heap.cpp
index 1d7fab0cab..cef0662e4a 100644
--- a/rpcs3/Emu/RSX/VK/vkutils/data_heap.cpp
+++ b/rpcs3/Emu/RSX/VK/vkutils/data_heap.cpp
@@ -13,14 +13,20 @@ namespace vk
 {
 	data_heap g_upload_heap;
 
-	void data_heap::create(VkBufferUsageFlags usage, usz size, const char* name, usz guard, VkBool32 notify)
+	void data_heap::create(VkBufferUsageFlags usage, usz size, rsx::flags32_t flags, const char* name, usz guard, VkBool32 notify)
 	{
 		::data_heap::init(size, name, guard);
 
 		const auto& memory_map = g_render_device->get_memory_mapping();
 
+		if (flags & heap_pool_low_latency)
+		{
+			// Prefer uploading to BAR if low latency is desired.
+			m_prefer_writethrough = memory_map.device_bar_total_bytes > (2048ull * 0x100000);
+		}
+
 		VkFlags memory_flags = VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
-		auto memory_index = memory_map.host_visible_coherent;
+		auto memory_index = m_prefer_writethrough ? memory_map.device_bar : memory_map.host_visible_coherent;
 
 		if (!(get_heap_compatible_buffer_types() & usage))
 		{
@@ -75,7 +81,7 @@ namespace vk
 		const auto& memory_map = g_render_device->get_memory_mapping();
 
 		VkFlags memory_flags = VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
-		auto memory_index = memory_map.host_visible_coherent;
+		auto memory_index = m_prefer_writethrough ? memory_map.device_bar : memory_map.host_visible_coherent;
 
 		// Update heap information and reset the allocator
 		::data_heap::init(aligned_new_size, m_name, m_min_guard_size);
@@ -163,7 +169,7 @@ namespace vk
 	{
 		if (!g_upload_heap.heap)
 		{
-			g_upload_heap.create(VK_BUFFER_USAGE_TRANSFER_SRC_BIT, 64 * 0x100000, "auxilliary upload heap", 0x100000);
+			g_upload_heap.create(VK_BUFFER_USAGE_TRANSFER_SRC_BIT, 64 * 0x100000, vk::heap_pool_default, "auxilliary upload heap", 0x100000);
 		}
 
 		return &g_upload_heap;
diff --git a/rpcs3/Emu/RSX/VK/vkutils/data_heap.h b/rpcs3/Emu/RSX/VK/vkutils/data_heap.h
index 8d8c6e3be3..e804bcdc59 100644
--- a/rpcs3/Emu/RSX/VK/vkutils/data_heap.h
+++ b/rpcs3/Emu/RSX/VK/vkutils/data_heap.h
@@ -11,6 +11,12 @@
 
 namespace vk
 {
+	enum data_heap_pool_flags
+	{
+		heap_pool_default = 0,
+		heap_pool_low_latency = 1,
+	};
+
 	class data_heap : public ::data_heap
 	{
 	private:
@@ -19,6 +25,7 @@ namespace vk
 		void* _ptr = nullptr;
 
 		bool notify_on_grow = false;
+		bool m_prefer_writethrough = false;
 
 		std::unique_ptr<buffer> shadow;
 		std::vector<VkBufferCopy> dirty_ranges;
@@ -33,7 +40,7 @@ namespace vk
 		// Avoid mapping/unmapping to keep these drivers from stalling
 		// NOTE2: HOST_CACHED flag does not keep the mapped ptr around in the driver either
 
-		void create(VkBufferUsageFlags usage, usz size, const char* name, usz guard = 0x10000, VkBool32 notify = VK_FALSE);
+		void create(VkBufferUsageFlags usage, usz size, rsx::flags32_t flags, const char* name, usz guard = 0x10000, VkBool32 notify = VK_FALSE);
 		void destroy();
 
 		void* map(usz offset, usz size);