diff --git a/Core/Src/main.c b/Core/Src/main.c
index 6a0728a..1702106 100644
--- a/Core/Src/main.c
+++ b/Core/Src/main.c
@@ -841,7 +841,7 @@ void MPU_Config(int enable_psram)
 
  HAL_MPU_ConfigRegion(&MPU_InitStruct);
 
- /* Configure the MPU as cacheable non executable SRAM */
+ /* Configure the MPU as cacheable non executable SRAM (write back, read/write allocate) */
  MPU_InitStruct.Enable = MPU_REGION_ENABLE;
  MPU_InitStruct.BaseAddress = 0x20000000;
  MPU_InitStruct.Size = MPU_REGION_SIZE_512MB;
diff --git a/damc/damc_simple_lib/AudioCApi.cpp b/damc/damc_simple_lib/AudioCApi.cpp
index 15d102e..0b38c91 100644
--- a/damc/damc_simple_lib/AudioCApi.cpp
+++ b/damc/damc_simple_lib/AudioCApi.cpp
@@ -101,7 +101,7 @@ extern "C" void BSP_AUDIO_OUT_TransferComplete_CallBack(void) {
 	BSP_AUDIO_OUT_HalfTransfer_CallBack();
 }
 
-__attribute__((used)) CircularBuffer<uint32_t, 3> usbBuffers[3];
+__attribute__((used)) CircularBuffer<uint32_t, 3, false> usbBuffers[3];
 __attribute__((used)) int32_t diff_usb[3];
 __attribute__((used)) int32_t expected_buff[3];
 __attribute__((used)) int32_t usb_buff[3];
diff --git a/damc/damc_simple_lib/CircularBuffer.h b/damc/damc_simple_lib/CircularBuffer.h
index 3656902..9301412 100644
--- a/damc/damc_simple_lib/CircularBuffer.h
+++ b/damc/damc_simple_lib/CircularBuffer.h
@@ -11,7 +11,7 @@
 #include <stm32f7xx.h>
 #include <string.h>
 
-template<typename T, int N> class CircularBuffer {
+template<typename T, int N, bool do_manage_cache> class CircularBuffer {
 public:
 	CircularBuffer();
 
@@ -38,12 +38,12 @@ template<typename T, int N> class CircularBuffer {
 	size_t in_read_offset;
 };
 
-template<typename T, int N> CircularBuffer<T, N>::CircularBuffer() {
+template<typename T, int N, bool do_manage_cache> CircularBuffer<T, N, do_manage_cache>::CircularBuffer() {
 	memset(buffer.data(), 0, buffer.size() * sizeof(buffer[0]));
 }
 
-template<typename T, int N>
-void CircularBuffer<T, N>::writeOutBuffer(uint32_t dma_read_offset, const T* data, size_t nframes) {
+template<typename T, int N, bool do_manage_cache>
+void CircularBuffer<T, N, do_manage_cache>::writeOutBuffer(uint32_t dma_read_offset, const T* data, size_t nframes) {
 	uint16_t start = out_write_offset;
 	uint16_t size = nframes;
 
@@ -70,12 +70,17 @@ void CircularBuffer<T, N>::writeOutBuffer(uint32_t dma_read_offset, const T* dat
 	}
 
 	assert(end < buffer.size());
-	SCB_CleanDCache_by_Addr((uint32_t*) buffer.data(), buffer.size() * sizeof(buffer[0]));
+	if(do_manage_cache) {
+		SCB_CleanDCache_by_Addr((uint32_t*) buffer.data(), buffer.size() * sizeof(buffer[0]));
+	} else {
+		__DSB();
+	}
+	assert(out_write_offset == start);
 	out_write_offset = end;
 }
 
-template<typename T, int N>
-size_t CircularBuffer<T, N>::readInBuffer(uint32_t dma_write_offset, T* data, size_t nframes) {
+template<typename T, int N, bool do_manage_cache>
+size_t CircularBuffer<T, N, do_manage_cache>::readInBuffer(uint32_t dma_write_offset, T* data, size_t nframes) {
 	uint16_t start = in_read_offset;
 
 	uint16_t end = dma_write_offset;
@@ -86,13 +91,15 @@ size_t CircularBuffer<T, N>::readInBuffer(uint32_t dma_write_offset, T* data, si
 
 	if(size > nframes) {
 		size = nframes;
-		end = (in_read_offset + size) % buffer.size();
+		end = (start + size) % buffer.size();
 	}
 	assert(end < buffer.size());
 
 	uint16_t total_size = 0;
 
-	SCB_InvalidateDCache_by_Addr((uint32_t*) buffer.data(), buffer.size() * sizeof(buffer[0]));
+	if(do_manage_cache) {
+		SCB_InvalidateDCache_by_Addr((uint32_t*) buffer.data(), buffer.size() * sizeof(buffer[0]));
+	}
 
 	if(end < start) {
 		// Copy between start and end of buffer
@@ -118,6 +125,10 @@ size_t CircularBuffer<T, N>::readInBuffer(uint32_t dma_write_offset, T* data, si
 		data[i] = fill_sample;
 	}
 
+	if(!do_manage_cache) {
+		__DSB();
+	}
+	assert(in_read_offset == start);
 	in_read_offset = end;
 
 	return size;
diff --git a/damc/damc_simple_lib/CodecAudio.h b/damc/damc_simple_lib/CodecAudio.h
index a39330d..4173c66 100644
--- a/damc/damc_simple_lib/CodecAudio.h
+++ b/damc/damc_simple_lib/CodecAudio.h
@@ -30,7 +30,7 @@ class CodecAudio {
 	struct CodecFrame {
 		int16_t headphone[2];
 	};
-	CircularBuffer<CodecFrame, 2> out_buffer;
-	CircularBuffer<CodecFrame, 2> in_buffer;
+	CircularBuffer<CodecFrame, 2, true> out_buffer;
+	CircularBuffer<CodecFrame, 2, true> in_buffer;
 	CodecInit codecInit;
 };