diff --git a/i2c.c b/i2c.c
index ca561b7..c22d479 100644
--- a/i2c.c
+++ b/i2c.c
@@ -27,7 +27,9 @@
 #include <avr/io.h>
 #include <avr/interrupt.h>
 #include <util/twi.h>
+#include "delay.h"
 #include "pinio.h"
+#include "memory_barrier.h"
 
 
 #if defined I2C_MASTER_MODE && defined I2C_SLAVE_MODE
@@ -66,11 +68,7 @@
 // Address of the device that is communicated with.
 uint8_t i2c_address;
 // State of TWI component of MCU.
-volatile uint8_t i2c_state;
-// Index into the send/receive buffer.
-uint8_t i2c_index;
-// Count of bytes to be sent.
-uint8_t i2c_byte_count;
+volatile uint8_t i2c_state = I2C_MODE_FREE;
 
 #ifdef I2C_EEPROM_SUPPORT
   // For SAWSARP mode (see ENHA in i2c.h).
@@ -84,10 +82,17 @@ uint8_t i2c_byte_count;
 #ifdef I2C_SLAVE_MODE
   uint8_t i2c_in_buffer[I2C_SLAVE_RX_BUFFER_SIZE];
   uint8_t i2c_out_buffer[I2C_SLAVE_TX_BUFFER_SIZE];
-#else
-  uint8_t *i2c_buffer;
 #endif /* I2C_SLAVE_MODE */
 
+// Ringbuffer logic for buffer 'send'.
+#define BUFSIZE I2C_BUFFER_SIZE
+
+volatile uint8_t sendhead = 0;
+volatile uint8_t sendtail = 0;
+volatile uint8_t sendbuf[BUFSIZE];
+
+#include "ringbuffer.h"
+
 
 /**
   Inititalise the I2C/TWI subsystem.
@@ -129,21 +134,46 @@ void i2c_init(uint8_t address) {
 }
 
 /**
-  Send a data block to a slave device.
+  Send a byte to the I2C partner.
+
+  \param address    I2C address of the communications partner.
+
+  \param data       The byte to be buffered/sent.
+
+  \param last_byte  Wether this is the last byte of a transaction.
+
+  Unlike many other protocols (serial, SPI), I2C has an explicite transmission
+  start and transmission end. Transmission start is detected automatically,
+  but end of the transmission has to be told by the invoking code.
+
+  Data is buffered, so this returns quickly for small amounts of data. Large
+  amounts don't get lost, but this function has to wait until sufficient
+  previous data was sent.
+
+  TODO: for now this function assumes that the buffer drains between distinct
+        transmissions. This means, last_byte is ignored and transmission is
+        ended as soon as the buffer drains.
 */
-void i2c_send(uint8_t address, uint8_t* block, uint8_t tx_len) {
+void i2c_write(uint8_t address, uint8_t data, uint8_t last_byte) {
 
-  i2c_address = address;
-  i2c_buffer = block;
-  i2c_index = 0;
-  i2c_byte_count = tx_len;
+  if ( ! (i2c_state & I2C_MODE_BUSY)) {
+    // No transmission ongoing, start one.
+    i2c_address = address;
 
-  // Just send.
-  i2c_state = I2C_MODE_SAWP;
+    // Just send.
+    i2c_state = I2C_MODE_SAWP;
 
-  // Start transmission.
-  TWCR = (1<<TWINT)|(0<<TWEA)|(1<<TWSTA)|(0<<TWSTO)|(1<<TWEN)|(1<<TWIE);
-  i2c_state |= I2C_MODE_BUSY;
+    // Start transmission.
+    TWCR = (1<<TWINT)|(0<<TWEA)|(1<<TWSTA)|(0<<TWSTO)|(1<<TWEN)|(1<<TWIE);
+    i2c_state |= I2C_MODE_BUSY;
+  }
+
+  while ( ! buf_canwrite(send))
+    delay_us(10);
+
+  ATOMIC_START
+    buf_push(send, data);
+  ATOMIC_END
 }
 
 /**
@@ -206,8 +236,8 @@ ISR(TWI_vect) {
       break;
     case TW_MT_SLA_ACK:
       // SLA+W was sent, then ACK received.
-      if ((i2c_state & I2C_MODE_MASK) == I2C_MODE_SAWP) {
-        TWDR = i2c_buffer[i2c_index++];
+      if ((i2c_state & I2C_MODE_MASK) == I2C_MODE_SAWP && buf_canread(send)) {
+        buf_pop(send, TWDR);
         TWCR = (1<<TWINT)|(I2C_MODE<<TWEA)|(0<<TWSTA)|(0<<TWSTO)|(1<<TWEN)|(1<<TWIE);
       }
       #ifdef I2C_EEPROM_SUPPORT
@@ -226,13 +256,14 @@ ISR(TWI_vect) {
     case TW_MT_DATA_ACK:
       // A byte was sent, got ACK.
       if ((i2c_state & I2C_MODE_MASK) == I2C_MODE_SAWP) {
-        if (i2c_index == i2c_byte_count) {
-          // Last byte, send stop condition.
-          TWCR = (1<<TWINT)|(I2C_MODE<<TWEA)|(0<<TWSTA)|(1<<TWSTO)|(1<<TWEN)|(0<<TWIE);
-        } else {
+        if (buf_canread(send)) {
           // Send the next byte.
-          TWDR = i2c_buffer[i2c_index++];
+          buf_pop(send, TWDR);
           TWCR = (1<<TWINT)|(I2C_MODE<<TWEA)|(0<<TWSTA)|(0<<TWSTO)|(1<<TWEN)|(1<<TWIE);
+        } else {
+          // Last byte, send stop condition.
+          i2c_state = I2C_MODE_FREE;
+          TWCR = (1<<TWINT)|(I2C_MODE<<TWEA)|(0<<TWSTA)|(1<<TWSTO)|(1<<TWEN)|(0<<TWIE);
         }
       }
       #ifdef I2C_EEPROM_SUPPORT
@@ -261,13 +292,15 @@ ISR(TWI_vect) {
       // It looks like there is another master on the bus.
       i2c_state |= I2C_ERROR_LOW_PRIO;
       // Setup all again.
-      i2c_index = 0;
+      sendtail = sendhead;
       #ifdef I2C_EEPROM_SUPPORT
         i2c_page_index = 0;
       #endif
       // Try to resend when the bus became free.
       TWCR = (1<<TWINT)|(I2C_MODE<<TWEA)|(1<<TWSTA)|(0<<TWSTO)|(1<<TWEN)|(1<<TWIE);
       break;
+#if 0
+// Note: we currently have no receive buffer, so we can't receive anything.
     case TW_MR_SLA_ACK:
       // SLA+R was sent, got АСК, then received a byte.
       if (i2c_index + 1 == i2c_byte_count) {
@@ -302,6 +335,7 @@ ISR(TWI_vect) {
       // Send stop condition.
       TWCR = (1<<TWINT)|(I2C_MODE<<TWEA)|(0<<TWSTA)|(1<<TWSTO)|(1<<TWEN)|(0<<TWIE);
       break;
+#endif
 
     #ifdef I2C_SLAVE_MODE
 
diff --git a/i2c.h b/i2c.h
index e484b80..d5d50a7 100644
--- a/i2c.h
+++ b/i2c.h
@@ -17,7 +17,15 @@
 // Comment out if there are external pullups.
 #define I2C_ENABLE_PULLUPS
 
-#define I2C_BUFFER_SIZE             4
+/** \def I2C_BUFFER_SIZE
+
+  Size of send buffer. MUST be a \f$2^n\f$ value, maximum is 512.
+
+  A larger buffer allows to store more display data immediately, so it can
+  speed operations up. An exhaused buffer doesn't mean data gets lost, writing
+  to the buffer then waits until sufficient previous data is sent.
+*/
+#define I2C_BUFFER_SIZE             128
 
 #ifdef I2C_SLAVE_MODE
   #define I2C_SLAVE_RX_BUFFER_SIZE  1
@@ -31,7 +39,7 @@
 
 
 void i2c_init(uint8_t address);
-void i2c_send(uint8_t address, uint8_t* block, uint8_t tx_len);
+void i2c_write(uint8_t address, uint8_t data, uint8_t last_byte);
 
 #endif /* I2C */
 
diff --git a/i2c_test.c b/i2c_test.c
index 15d7f57..055ef25 100644
--- a/i2c_test.c
+++ b/i2c_test.c
@@ -28,6 +28,8 @@ static uint8_t display_init[] = {
   0xA8, 0x1F,       // 1/32 duty.
   0x40 | 0x00,      // Start line (reset).
   0x20, 0x02,       // Page addressing mode (reset).
+  0x22, 0x00, 0x03, // Start and end page in horiz./vert. addressing mode[1].
+  0x21, 0x00, 0x7F, // Start and end column in horiz./vert. addressing mode.
   0xA0 | 0x00,      // No segment remap (reset).
   0xC0 | 0x00,      // Normal com pins mapping (reset).
   0xDA, 0x02,       // Sequental without remap com pins.
@@ -39,6 +41,10 @@ static uint8_t display_init[] = {
   0xA4,             // Resume display.
   0xAF              // Display on.
 };
+// [1] Do not set this to 0x00..0x07 on a 32 pixel high display, or vertical
+//     addressing mode will mess up. 32 pixel high displays have only 4 pages
+//     (0..3), still addressing logic accepts, but can't deal with the 0..7
+//     meant for 64 pixel high displays.
 
 typedef struct {
   uint8_t columns;
@@ -153,12 +159,37 @@ SYMBOL font_8x4[] = {
 
 
 static void i2c_test(void) {
-  static char* message = "Welcome to Teacup";
-  static uint8_t block[128];
-  uint8_t* pointer = block + 1;
+  uint16_t i;
+  const char* message = "Welcome to Teacup";
 
-  i2c_send(DISPLAY_I2C_ADDRESS, display_init, sizeof(display_init));
-  delay_ms(250);
+  for (i = 0; i < sizeof(display_init); i++) {
+    i2c_write(DISPLAY_I2C_ADDRESS, display_init[i], 0);
+  }
+  delay_ms(500);
+
+  /**
+    Clear the screen. As this display supports many sophisticated commands,
+    but not a simple 'clear', we have to overwrite the entire memory with
+    zeros, byte by byte.
+  */
+  // Set horizontal adressing mode.
+  i2c_write(DISPLAY_I2C_ADDRESS, 0x00, 0);
+  i2c_write(DISPLAY_I2C_ADDRESS, 0x20, 0);
+  i2c_write(DISPLAY_I2C_ADDRESS, 0x00, 0);
+  delay_ms(100);
+
+  // Write 512 zeros.
+  i2c_write(DISPLAY_I2C_ADDRESS, 0x40, 0);
+  for (i = 0; i < 512; i++) {
+    i2c_write(DISPLAY_I2C_ADDRESS, 0x00, 0);
+  }
+  delay_ms(2000);
+
+  // Return to page adressing mode.
+  i2c_write(DISPLAY_I2C_ADDRESS, 0x00, 0);
+  i2c_write(DISPLAY_I2C_ADDRESS, 0x20, 0);
+  i2c_write(DISPLAY_I2C_ADDRESS, 0x02, 0);
+  delay_ms(100);
 
   /**
     Setup cursor on display.
@@ -166,24 +197,30 @@ static void i2c_test(void) {
     "Welcome to Teacup" is 64 pixel columns wide, entire display is
     128 columns, so we offset by 32 columns to get it to the center.
   */
-  block[0] = 0x00;
+  i2c_write(DISPLAY_I2C_ADDRESS, 0x00, 0);
   // Line 1.
-  block[1] = 0xB0 | 1;
+  i2c_write(DISPLAY_I2C_ADDRESS, 0xB0 | 1, 0);
   // Column 32.
-  block[2] = 0x00 | (32 & 0x0F);
-  block[3] = 0x10 | ((32 >> 4) & 0x0F);
-  i2c_send(DISPLAY_I2C_ADDRESS, block, 4);
-  delay_ms(50);
+  i2c_write(DISPLAY_I2C_ADDRESS, 0x00 | (32 & 0x0F), 0);
+  i2c_write(DISPLAY_I2C_ADDRESS, 0x10 | ((32 >> 4) & 0x0F), 0);
+  delay_ms(100);
 
-  // Render text to bitmap.
+  // Render text to bitmap to display.
+  i2c_write(DISPLAY_I2C_ADDRESS, 0x40, 0);
   while (*message) {
     SYMBOL symbol = font_8x4[(uint8_t)*message - 0x20];
-    memcpy(pointer, symbol.data, symbol.columns);
-    pointer += symbol.columns + FONT_SYMBOLS_SPACE;
+
+    // Send the character bitmap.
+    for (i = 0; i < symbol.columns; i++) {
+      i2c_write(DISPLAY_I2C_ADDRESS, symbol.data[i], 0);
+    }
+    // Send space between characters.
+    for (i = 0; i < FONT_SYMBOLS_SPACE; i++) {
+      i2c_write(DISPLAY_I2C_ADDRESS, 0x00, 0);
+    }
+
     message++;
   }
-  block[0] = 0x40; // Data marker.
-  i2c_send(DISPLAY_I2C_ADDRESS, block, pointer - block);
 }
 
 #endif /* I2C_TEST */