diff --git a/fw/SummerCart64.qsf b/fw/SummerCart64.qsf
index ef2aaa0..3c41e7a 100644
--- a/fw/SummerCart64.qsf
+++ b/fw/SummerCart64.qsf
@@ -19,7 +19,7 @@
 #
 # Quartus Prime
 # Version 20.1.1 Build 720 11/11/2020 SJ Lite Edition
-# Date created = 00:33:20  February 17, 2021
+# Date created = 01:35:54  February 19, 2021
 #
 # -------------------------------------------------------------------------- #
 #
@@ -77,6 +77,7 @@ set_global_assignment -name VERILOG_FILE rtl/top.v
 set_global_assignment -name VERILOG_FILE rtl/usb/usb_ftdi_fsi.v
 set_global_assignment -name VERILOG_FILE rtl/usb/usb_pc.v
 set_global_assignment -name VERILOG_INCLUDE_FILE rtl/constants.vh
+set_global_assignment -name SLD_FILE db/signal_tap_logic_analyzer_auto_stripped.stp
 
 # Pin & Location Assignments
 # ==========================
@@ -272,11 +273,10 @@ set_global_assignment -name OUTPUT_IO_TIMING_FAR_END_VMEAS "HALF SIGNAL SWING" -
 		set_global_assignment -name PARTITION_NETLIST_TYPE SOURCE -section_id Top
 		set_global_assignment -name PARTITION_FITTER_PRESERVATION_LEVEL PLACEMENT_AND_ROUTING -section_id Top
 		set_global_assignment -name PARTITION_COLOR 16764057 -section_id Top
+		set_instance_assignment -name PARTITION_HIERARCHY root_partition -to | -section_id Top
 
 	# end DESIGN_PARTITION(Top)
 	# -------------------------
 
 # end ENTITY(top)
-# ---------------
-set_global_assignment -name SLD_FILE db/signal_tap_logic_analyzer_auto_stripped.stp
-set_instance_assignment -name PARTITION_HIERARCHY root_partition -to | -section_id Top
\ No newline at end of file
+# ---------------
\ No newline at end of file
diff --git a/fw/constraints.sdc b/fw/constraints.sdc
index 9485af5..1c52faa 100644
--- a/fw/constraints.sdc
+++ b/fw/constraints.sdc
@@ -43,14 +43,13 @@ set_false_path -from [get_ports {i_ftdi_so i_ftdi_cts}]
 set_output_delay -clock [get_clocks {sd_clk}] -max 6.0 [get_ports {io_sd_cmd io_sd_dat[*]}]
 set_output_delay -clock [get_clocks {sd_clk}] -min -2.0 [get_ports {io_sd_cmd io_sd_dat[*]}]
 
-set_input_delay -clock [get_clocks {sd_clk}] -max 14.0 [get_ports {io_sd_cmd io_sd_dat[*]}]
-set_input_delay -clock [get_clocks {sd_clk}] -min 2.5 [get_ports {io_sd_cmd io_sd_dat[*]}]
+set_input_delay -clock [get_clocks {sd_clk}] -max 15.0 [get_ports {io_sd_cmd io_sd_dat[*]}]
+set_input_delay -clock [get_clocks {sd_clk}] -min 6.5 [get_ports {io_sd_cmd io_sd_dat[*]}]
 
-set_multicycle_path -setup -start 1 -from [get_clocks $sys_clk] -to [get_clocks {sd_clk}]
 set_multicycle_path -hold -start 1 -from [get_clocks $sys_clk] -to [get_clocks {sd_clk}]
 
 set_multicycle_path -setup -end 3 -from [get_clocks {sd_clk}] -to [get_clocks $sys_clk]
-set_multicycle_path -hold -end 2 -from [get_clocks {sd_clk}] -to [get_clocks $sys_clk]
+set_multicycle_path -hold -end 1 -from [get_clocks {sd_clk}] -to [get_clocks $sys_clk]
 
 
 # N64, PI and SI timings
diff --git a/fw/rtl/sd/sd_dat.v b/fw/rtl/sd/sd_dat.v
index 10c0863..7d4bd1c 100644
--- a/fw/rtl/sd/sd_dat.v
+++ b/fw/rtl/sd/sd_dat.v
@@ -17,6 +17,7 @@ module sd_dat (
     output o_dat_write_busy,
     output reg o_dat_crc_error,
     output reg o_dat_write_error,
+    output reg o_dat_write_ok,
 
     input i_rx_fifo_overrun,
     output reg o_rx_fifo_push,
@@ -41,7 +42,7 @@ module sd_dat (
     reg [6:0] r_state;
 
     assign o_dat_busy = !r_state[STATE_IDLE];
-    assign o_dat_write_busy = r_state[STATE_SENDING] || r_state[STATE_STATUS] || r_state[STATE_BUSY];
+    assign o_dat_write_busy = !io_sd_dat[0];
 
 
     // Bit counter logic
@@ -50,13 +51,13 @@ module sd_dat (
     reg r_bit_counter_done;
 
     wire w_read_start = r_state[STATE_READ_WAIT] && !io_sd_dat[0] && i_sd_clk_strobe_rising;
-    wire w_write_start = r_state[STATE_WRITE_WAIT] && (i_tx_fifo_items == ({1'b0, i_dat_block_size} + 1)) && i_sd_clk_strobe_falling;
+    wire w_write_start = r_state[STATE_WRITE_WAIT] && (i_tx_fifo_items >= ({1'b0, i_dat_block_size} + 1)) && i_sd_clk_strobe_falling;
     wire w_status_start = r_state[STATE_SENDING] && r_bit_counter_done;
     wire w_status_done = r_state[STATE_STATUS] && r_bit_counter_done;
     wire w_block_read_done = r_state[STATE_RECEIVING] && r_bit_counter_done;
     wire w_block_write_done = r_state[STATE_SENDING] && r_bit_counter_done;
-    wire w_block_done = w_block_read_done || w_block_write_done;
     wire w_block_write_busy_done = r_state[STATE_BUSY] && io_sd_dat[0];
+    wire w_block_done = w_block_read_done || w_block_write_busy_done;
 
     wire [12:0] w_block_bit_length = i_dat_width ? ({2'b00, {1'b0, i_dat_block_size} + 1'd1, 3'b000}) : ({{1'b0, i_dat_block_size} + 1'd1, 5'b00000});
 
@@ -138,12 +139,14 @@ module sd_dat (
 
     reg [4:0] r_status;
 
+    wire w_no_write_error = r_status == STATUS_NO_ERROR;
     wire w_crc_write_error = r_status == STATUS_CRC_ERROR;
     wire w_data_write_error = r_status == STATUS_WRITE_ERROR;
 
     always @(posedge i_clk) begin
-        if (i_reset) begin
+        if (i_reset || w_block_start) begin
             o_dat_crc_error <= 1'b0;
+            o_dat_write_error <= 1'b0;
         end else begin
             if (w_block_read_done) begin
                 o_dat_crc_error <= w_crc_read_error;
@@ -151,6 +154,7 @@ module sd_dat (
             if (w_status_done) begin
                 o_dat_crc_error <= w_crc_write_error;
                 o_dat_write_error <= w_data_write_error;
+                o_dat_write_ok <= w_no_write_error;
             end
         end
     end
@@ -230,7 +234,7 @@ module sd_dat (
 
                     r_state[STATE_BUSY]: begin
                         if (w_block_write_busy_done) begin
-                            if (w_block_stop) begin
+                            if (w_block_stop || !w_no_write_error) begin
                                 r_state[STATE_IDLE] <= 1'b1;
                             end else begin
                                 r_state[STATE_WRITE_WAIT] <= 1'b1;
@@ -274,7 +278,7 @@ module sd_dat (
     wire w_tx_latch = r_state[STATE_SENDING] && i_sd_clk_strobe_falling;
     wire w_tx_data_phase = r_bit_counter >= 13'd17;
     wire w_tx_crc_phase = r_bit_counter >= 13'd1;
-    wire w_tx_fifo_pop = i_dat_width ? (r_bit_counter[2:0] == 3'd0) : (r_bit_counter[4:0] == 5'd16);
+    wire w_tx_fifo_pop = i_dat_width ? (r_bit_counter[2:0] == 3'd0) : (r_bit_counter[4:0] == 5'd17);
     wire w_tx_shift_load = r_bit_counter[2:0] == 3'd0;
 
     reg [7:0] r_tx_shift [0:3];
@@ -309,7 +313,12 @@ module sd_dat (
                             end
                         end
                     end else begin
-                        r_tx_shift[0] <= i_tx_fifo_data[{3'b000, r_bit_counter[4:3]} +: 8];
+                        case (r_bit_counter[4:3])
+                            2'b10: r_tx_shift[0] <= i_tx_fifo_data[31:24];
+                            2'b01: r_tx_shift[0] <= i_tx_fifo_data[23:16];
+                            2'b00: r_tx_shift[0] <= i_tx_fifo_data[15:8];
+                            2'b11: r_tx_shift[0] <= i_tx_fifo_data[7:0];
+                        endcase
                     end
                 end
             end else if (w_tx_crc_phase) begin
diff --git a/fw/rtl/sd/sd_interface.v b/fw/rtl/sd/sd_interface.v
index 222495f..27ea16a 100644
--- a/fw/rtl/sd/sd_interface.v
+++ b/fw/rtl/sd/sd_interface.v
@@ -159,6 +159,7 @@ module sd_interface (
     wire w_dat_write_busy;
     wire w_dat_crc_error;
     wire w_dat_write_error;
+    wire w_dat_write_ok;
 
     sd_dat sd_dat_inst (
         .i_clk(i_clk),
@@ -179,6 +180,7 @@ module sd_interface (
         .o_dat_write_busy(w_dat_write_busy),
         .o_dat_crc_error(w_dat_crc_error),
         .o_dat_write_error(w_dat_write_error),
+        .o_dat_write_ok(w_dat_write_ok),
 
         .i_rx_fifo_overrun(w_rx_fifo_overrun),
         .o_rx_fifo_push(w_rx_fifo_push),
@@ -267,6 +269,7 @@ module sd_interface (
         .i_dat_write_busy(w_dat_write_busy),
         .i_dat_crc_error(w_dat_crc_error),
         .i_dat_write_error(w_dat_write_error),
+        .i_dat_write_ok(w_dat_write_ok),
 
         .o_rx_fifo_flush(w_rx_fifo_flush),
         .o_rx_fifo_pop(w_rx_fifo_regs_pop),
diff --git a/fw/rtl/sd/sd_regs.v b/fw/rtl/sd/sd_regs.v
index 02afded..d8a7339 100644
--- a/fw/rtl/sd/sd_regs.v
+++ b/fw/rtl/sd/sd_regs.v
@@ -25,6 +25,7 @@ module sd_regs (
     input i_dat_write_busy,
     input i_dat_crc_error,
     input i_dat_write_error,
+    input i_dat_write_ok,
 
     output reg o_rx_fifo_flush,
     output reg o_rx_fifo_pop,
@@ -204,7 +205,8 @@ module sd_regs (
 
                     SD_REG_DAT: begin
                         o_data <= {
-                            4'd0,
+                            3'd0,
+                            i_dat_write_ok,
                             i_dat_write_error,
                             i_dat_write_busy,
                             i_tx_fifo_items,
diff --git a/fw/rtl/top.v b/fw/rtl/top.v
index 634c801..4860aee 100644
--- a/fw/rtl/top.v
+++ b/fw/rtl/top.v
@@ -539,7 +539,7 @@ module top (
 
     // LED
 
-    wire w_led_trigger = (w_n64_request && !w_n64_busy) || (w_pc_request && !w_pc_busy);
+    wire w_led_trigger = (w_n64_request && !w_n64_busy) || (w_pc_request && !w_pc_busy) || (w_sd_dma_request && !w_sd_dma_busy);
 
     cart_led cart_led_inst (
         .i_clk(w_sys_clk),
diff --git a/sw/bootloader/src/fatfs/diskio.c b/sw/bootloader/src/fatfs/diskio.c
index b4479c3..95e0805 100644
--- a/sw/bootloader/src/fatfs/diskio.c
+++ b/sw/bootloader/src/fatfs/diskio.c
@@ -53,7 +53,7 @@ DRESULT disk_read(BYTE pdrv, BYTE *buff, LBA_t sector, UINT count) {
 
 #if !FF_FS_READONLY
 
-DRESULT disk_write (BYTE pdrv, const BYTE* buff, LBA_t sector, UINT count) {
+DRESULT disk_write(BYTE pdrv, const BYTE* buff, LBA_t sector, UINT count) {
     sc64_sd_err_t error;
 
     if (pdrv > 0) {
@@ -79,5 +79,15 @@ DRESULT disk_write (BYTE pdrv, const BYTE* buff, LBA_t sector, UINT count) {
 #endif
 
 DRESULT disk_ioctl(BYTE pdrv, BYTE cmd, void *buff) {
-    return RES_PARERR;
+    switch (cmd) {
+        case CTRL_SYNC: {
+            sc64_sd_err_t error = sc64_sd_dat_busy_wait();
+            if (error != E_OK) {
+                return RES_ERROR;
+            }
+            break;
+        }
+    }
+
+    return RES_OK;
 }
diff --git a/sw/bootloader/src/loader/loader.c b/sw/bootloader/src/loader/loader.c
index 91b3fbd..77e2a20 100644
--- a/sw/bootloader/src/loader/loader.c
+++ b/sw/bootloader/src/loader/loader.c
@@ -97,24 +97,6 @@ void loader_display_logo(void) {
     display_show(display);
 }
 
-void loader_display_message(const char *message) {
-    if (!loader_initialized) {
-        loader_init();
-    }
-
-    display_context_t display;
-
-    display = loader_get_display(true);
-
-    graphics_fill_screen(display, 0);
-
-    loader_draw_version_and_logo(display);
-
-    graphics_draw_text(display, x_offset, y_offset, message);
-
-    display_show(display);
-}
-
 void loader_display_error_and_halt(menu_load_error_t error, const char *message) {
     if (!loader_initialized) {
         loader_init();
diff --git a/sw/bootloader/src/loader/loader.h b/sw/bootloader/src/loader/loader.h
index 9cb5e90..eafe5aa 100644
--- a/sw/bootloader/src/loader/loader.h
+++ b/sw/bootloader/src/loader/loader.h
@@ -7,7 +7,6 @@
 
 void loader_cleanup(void);
 void loader_display_logo(void);
-void loader_display_message(const char *message);
 void loader_display_error_and_halt(menu_load_error_t error, const char *message);
 
 
diff --git a/sw/bootloader/src/sc64/sc64_regs.h b/sw/bootloader/src/sc64/sc64_regs.h
index 03d7d53..06c543e 100644
--- a/sw/bootloader/src/sc64/sc64_regs.h
+++ b/sw/bootloader/src/sc64/sc64_regs.h
@@ -146,6 +146,9 @@ typedef struct sc64_sd_registers_s {
 #define SC64_SD_CMD_INDEX(i)                ((i) & 0x3F)
 
 
+#define SC64_SD_DAT_WRITE_OK                (1 << 28)
+#define SC64_SD_DAT_WRITE_ERROR             (1 << 27)
+#define SC64_SD_DAT_WRITE_BUSY              (1 << 26)
 #define SC64_SD_DAT_TX_FIFO_ITEMS_GET(dat)  (((dat) >> 17) & 0x1FF)
 #define SC64_SD_DAT_TX_FIFO_BYTES_GET(dat)  (SC64_SD_DAT_TX_FIFO_ITEMS_GET(dat) * 4)
 #define SC64_SD_DAT_TX_FIFO_FULL            (1 << 16)
@@ -184,6 +187,7 @@ typedef struct sc64_sd_registers_s {
 
 #define SC64_SD_DMA_BANK_ADDR(b, a)         ((((b) & 0xF) << 28) | ((a) & 0x3FFFFFC))
 
+
 #define SC64_SD_DMA_LEN_GET(len)            (((len) & 0x7FFF) * 4)
 
 #define SC64_SD_DMA_LEN(l)                  ((((l) / 4) - 1) & 0x7FFF)
diff --git a/sw/bootloader/src/sc64/sc64_sd.c b/sw/bootloader/src/sc64/sc64_sd.c
index 5c413d6..742feae 100644
--- a/sw/bootloader/src/sc64/sc64_sd.c
+++ b/sw/bootloader/src/sc64/sc64_sd.c
@@ -157,20 +157,24 @@ static sc64_sd_err_t sc64_sd_cmd_send(uint8_t index, uint32_t arg, sc64_sd_cmd_f
 }
 
 static void sc64_sd_dat_prepare(size_t num_blocks, size_t block_size, sc64_sd_dat_direction_t direction) {
-    platform_pi_io_write(&SC64_SD->DAT, (
+    uint32_t reg = (
         SC64_SD_DAT_NUM_BLOCKS(num_blocks) |
         SC64_SD_DAT_BLOCK_SIZE(block_size) |
         ((direction == DAT_DIR_TX) ? SC64_SD_DAT_DIRECTION : 0) |
         SC64_SD_DAT_START
-    ));
+    );
+
+    platform_pi_io_write(&SC64_SD->DAT, reg);
 }
 
 static void sc64_sd_dat_abort(void) {
-    platform_pi_io_write(&SC64_SD->DAT, (
+    uint32_t reg = (
         SC64_SD_DAT_TX_FIFO_FLUSH |
         SC64_SD_DAT_RX_FIFO_FLUSH |
         SC64_SD_DAT_STOP
-    ));
+    );
+
+    platform_pi_io_write(&SC64_SD->DAT, reg);
 }
 
 static sc64_sd_err_t sc64_sd_dat_read(size_t block_size, void *buffer) {
@@ -212,7 +216,7 @@ static sc64_sd_err_t sc64_sd_dat_read(size_t block_size, void *buffer) {
 static sc64_sd_err_t sc64_sd_dat_write(size_t block_size, void *buffer) {
     int timeout;
     uint32_t reg;
-
+    
     timeout = 1000000;
     do {
         reg = platform_pi_io_read(&SC64_SD->DAT);
@@ -221,11 +225,21 @@ static sc64_sd_err_t sc64_sd_dat_write(size_t block_size, void *buffer) {
         }
     } while ((reg & SC64_SD_DAT_BUSY) && (--timeout));
 
+    if (timeout == 0) {
+        sc64_sd_dat_abort();
+
+        return E_TIMEOUT;
+    }
+
     if (!(reg & SC64_SD_DAT_BUSY)) {
         if (reg & SC64_SD_DAT_CRC_ERROR) {
             return E_CRC_ERROR;
         }
 
+        if (reg & SC64_SD_DAT_WRITE_ERROR) {
+            return E_WRITE_ERROR;
+        }
+
         if (reg & SC64_SD_DAT_TX_FIFO_UNDERRUN) {
             platform_pi_io_write(&SC64_SD->DAT, SC64_SD_DAT_TX_FIFO_FLUSH);
 
@@ -233,14 +247,8 @@ static sc64_sd_err_t sc64_sd_dat_write(size_t block_size, void *buffer) {
         }
     }
 
-    if (timeout == 0) {
-        sc64_sd_dat_abort();
-
-        return E_TIMEOUT;
-    }
-    
+    platform_cache_writeback(buffer, block_size);
     platform_pi_dma_write(buffer, &SC64_SD->FIFO, block_size);
-    platform_cache_invalidate(buffer, block_size);
 
     return E_OK;
 }
@@ -420,8 +428,11 @@ sc64_sd_err_t sc64_sd_sectors_read(uint32_t starting_sector, size_t count, uint8
 
 sc64_sd_err_t sc64_sd_sectors_write(uint32_t starting_sector, size_t count, uint8_t *buffer) {
     sc64_sd_err_t error;
+    sc64_sd_err_t write_error;
     uint32_t response;
     uint32_t current_sector;
+    size_t sectors_remaining;
+    size_t num_blocks;
 
     error = sc64_sd_sectors_parameters_check(count, buffer, true);
     if (error != E_OK) {
@@ -433,23 +444,51 @@ sc64_sd_err_t sc64_sd_sectors_write(uint32_t starting_sector, size_t count, uint
         current_sector *= SD_BLOCK_SIZE;
     }
 
-    for (size_t i = 0; i < count; i++) {
-        sc64_sd_dat_prepare(1, SD_BLOCK_SIZE, DAT_DIR_TX);
+    sectors_remaining = count;
 
-        error = sc64_sd_cmd_send(24, current_sector, NO_FLAGS, &response);
+    while (sectors_remaining) {
+        num_blocks = (sectors_remaining > SC64_SD_DAT_NUM_BLOCKS_MAX) ? SC64_SD_DAT_NUM_BLOCKS_MAX : sectors_remaining;
+
+        sc64_sd_dat_prepare(num_blocks, SD_BLOCK_SIZE, DAT_DIR_TX);
+
+        error = sc64_sd_cmd_send(25, current_sector, NO_FLAGS, &response);
         if (error != E_OK) {
             sc64_sd_dat_abort();
 
             return error;
         }
 
-        error = sc64_sd_dat_write(SD_BLOCK_SIZE, buffer);
+        for (size_t i = 0; i < num_blocks; i++) {
+            write_error = sc64_sd_dat_write(SD_BLOCK_SIZE, buffer);
+            if (write_error != E_OK) {
+                break;
+            }
+
+            buffer += SD_BLOCK_SIZE;
+        }
+
+        error = sc64_sd_dat_busy_wait();
+        if (error != E_OK) {
+            sc64_sd_dat_abort();
+            return error;
+        }
+
+        error = sc64_sd_cmd_send(12, 0, NO_FLAGS, &response);
         if (error != E_OK) {
             return error;
         }
 
-        buffer += SD_BLOCK_SIZE;
-        current_sector += sd_card_type_block ? 1 : SD_BLOCK_SIZE;
+        error = sc64_sd_dat_busy_wait();
+        if (error != E_OK) {
+            return error;
+        }
+
+        if (write_error != E_OK) {
+            return write_error;
+        }
+
+        current_sector += num_blocks * (sd_card_type_block ? 1 : SD_BLOCK_SIZE);
+        sectors_remaining -= num_blocks;
     }
 
     return E_OK;
@@ -534,79 +573,95 @@ sc64_sd_err_t sc64_sd_sectors_read_dma(uint32_t starting_sector, size_t count, u
     return E_OK;
 }
 
-sc64_sd_err_t sc64_sd_sectors_write_dma(uint32_t starting_sector, size_t count, uint8_t bank, uint32_t address) {
-    size_t sectors_left;
-    uint32_t current_sector;
-    uint32_t current_address;
-    uint32_t num_blocks;
-    uint32_t reg;
-    sc64_sd_err_t error;
-    uint32_t response;
+// sc64_sd_err_t sc64_sd_sectors_write_dma(uint32_t starting_sector, size_t count, uint8_t bank, uint32_t address) {
+//     size_t sectors_left;
+//     uint32_t current_sector;
+//     uint32_t current_address;
+//     uint32_t num_blocks;
+//     uint32_t reg;
+//     sc64_sd_err_t error;
+//     uint32_t response;
+//     int timeout;
+
+//     error = sc64_sd_sectors_parameters_check(count, NULL, false);
+//     if (error != E_OK) {
+//         return error;
+//     }
+
+//     sectors_left = count;
+//     current_sector = starting_sector;
+//     if (!sd_card_type_block) {
+//         current_sector *= SD_BLOCK_SIZE;
+//     }
+//     current_address = address;
+
+//     do {
+//         num_blocks = (sectors_left > SC64_SD_DAT_NUM_BLOCKS_MAX) ? SC64_SD_DAT_NUM_BLOCKS_MAX : sectors_left;
+
+//         sc64_sd_dat_prepare(num_blocks, SD_BLOCK_SIZE, DAT_DIR_TX);
+//         sc64_sd_dma_prepare(num_blocks, SD_BLOCK_SIZE, DMA_DIR_CARD, bank, current_address);
+
+//         error = sc64_sd_cmd_send(25, current_sector, NO_FLAGS, &response);
+//         if (error != E_OK) {
+//             sc64_sd_dma_abort();
+//             sc64_sd_dat_abort();
+
+//             return error;
+//         }
+
+//         timeout = 1000000;
+//         do {
+//             reg = platform_pi_io_read(&SC64_SD->DAT);
+//         } while ((reg & SC64_SD_DAT_BUSY) && (--timeout));
+
+//         error = sc64_sd_cmd_send(12, 0, NO_FLAGS, &response);
+//         if (error != E_OK) {
+//             sc64_sd_dma_abort();
+//             sc64_sd_dat_abort();
+
+//             return error;
+//         }
+
+//         if (reg & SC64_SD_DAT_CRC_ERROR) {
+//             sc64_sd_dma_abort();
+
+//             return E_CRC_ERROR;
+//         }
+
+//         if (reg & SC64_SD_DAT_TX_FIFO_UNDERRUN) {
+//             sc64_sd_dma_abort();
+//             platform_pi_io_write(&SC64_SD->DAT, SC64_SD_DAT_TX_FIFO_FLUSH);
+
+//             return E_FIFO_ERROR;
+//         }
+
+//         if (timeout == 0) {
+//             sc64_sd_dma_abort();
+//             sc64_sd_dat_abort();
+
+//             return E_TIMEOUT;
+//         }
+
+//         sectors_left -= num_blocks;
+//         current_sector += num_blocks * (sd_card_type_block ? 1 : SD_BLOCK_SIZE);
+//         current_address += num_blocks * SD_BLOCK_SIZE;
+//     } while (sectors_left > 0);
+
+//     return E_OK;
+// }
+
+sc64_sd_err_t sc64_sd_dat_busy_wait(void) {
     int timeout;
+    uint32_t reg;
 
-    error = sc64_sd_sectors_parameters_check(count, NULL, false);
-    if (error != E_OK) {
-        return error;
-    }
-
-    sectors_left = count;
-    current_sector = starting_sector;
-    if (!sd_card_type_block) {
-        current_sector *= SD_BLOCK_SIZE;
-    }
-    current_address = address;
-
+    timeout = 1000000;
     do {
-        num_blocks = (sectors_left > SC64_SD_DAT_NUM_BLOCKS_MAX) ? SC64_SD_DAT_NUM_BLOCKS_MAX : sectors_left;
+        reg = platform_pi_io_read(&SC64_SD->DAT);
+    } while ((reg & (SC64_SD_DAT_WRITE_BUSY | SC64_SD_DAT_BUSY)) && (--timeout));
 
-        sc64_sd_dat_prepare(num_blocks, SD_BLOCK_SIZE, DAT_DIR_TX);
-        sc64_sd_dma_prepare(num_blocks, SD_BLOCK_SIZE, DMA_DIR_CARD, bank, current_address);
-
-        error = sc64_sd_cmd_send(25, current_sector, NO_FLAGS, &response);
-        if (error != E_OK) {
-            sc64_sd_dma_abort();
-            sc64_sd_dat_abort();
-
-            return error;
-        }
-
-        timeout = 1000000;
-        do {
-            reg = platform_pi_io_read(&SC64_SD->DAT);
-        } while ((reg & SC64_SD_DAT_BUSY) && (--timeout));
-
-        error = sc64_sd_cmd_send(12, 0, NO_FLAGS, &response);
-        if (error != E_OK) {
-            sc64_sd_dma_abort();
-            sc64_sd_dat_abort();
-
-            return error;
-        }
-
-        if (reg & SC64_SD_DAT_CRC_ERROR) {
-            sc64_sd_dma_abort();
-
-            return E_CRC_ERROR;
-        }
-
-        if (reg & SC64_SD_DAT_TX_FIFO_UNDERRUN) {
-            sc64_sd_dma_abort();
-            platform_pi_io_write(&SC64_SD->DAT, SC64_SD_DAT_TX_FIFO_FLUSH);
-
-            return E_FIFO_ERROR;
-        }
-
-        if (timeout == 0) {
-            sc64_sd_dma_abort();
-            sc64_sd_dat_abort();
-
-            return E_TIMEOUT;
-        }
-
-        sectors_left -= num_blocks;
-        current_sector += num_blocks * (sd_card_type_block ? 1 : SD_BLOCK_SIZE);
-        current_address += num_blocks * SD_BLOCK_SIZE;
-    } while (sectors_left > 0);
+    if (timeout == 0) {
+        return E_TIMEOUT;
+    }
 
     return E_OK;
 }
diff --git a/sw/bootloader/src/sc64/sc64_sd.h b/sw/bootloader/src/sc64/sc64_sd.h
index ece6052..395746b 100644
--- a/sw/bootloader/src/sc64/sc64_sd.h
+++ b/sw/bootloader/src/sc64/sc64_sd.h
@@ -12,6 +12,7 @@ typedef enum sc64_sd_err_e {
     E_BAD_INDEX,
     E_PAR_ERROR,
     E_FIFO_ERROR,
+    E_WRITE_ERROR,
     E_NO_INIT,
 } sc64_sd_err_t;
 
@@ -22,7 +23,8 @@ bool sc64_sd_status_get(void);
 sc64_sd_err_t sc64_sd_sectors_read(uint32_t starting_sector, size_t count, uint8_t *buffer);
 sc64_sd_err_t sc64_sd_sectors_write(uint32_t starting_sector, size_t count, uint8_t *buffer);
 sc64_sd_err_t sc64_sd_sectors_read_dma(uint32_t starting_sector, size_t count, uint8_t bank, uint32_t address);
-sc64_sd_err_t sc64_sd_sectors_write_dma(uint32_t starting_sector, size_t count, uint8_t bank, uint32_t address);
+// sc64_sd_err_t sc64_sd_sectors_write_dma(uint32_t starting_sector, size_t count, uint8_t bank, uint32_t address);
+sc64_sd_err_t sc64_sd_dat_busy_wait(void);
 
 
 #endif
diff --git a/sw/bootloader/src/sc64/sc64_sd_fs.c b/sw/bootloader/src/sc64/sc64_sd_fs.c
index e4fc0e3..3a274ae 100644
--- a/sw/bootloader/src/sc64/sc64_sd_fs.c
+++ b/sw/bootloader/src/sc64/sc64_sd_fs.c
@@ -12,6 +12,8 @@
 static uint8_t current_bank = SC64_BANK_INVALID;
 static uint32_t current_offset = 0;
 static uint8_t save_buffer[128 * 1024] __attribute__((aligned(16)));
+static bool fs_initialized = false;
+static FATFS fatfs;
 
 
 static DRESULT sc64_sd_fs_load_with_dma(BYTE pdrv, FSIZE_t offset, LBA_t sector, UINT count) {
@@ -38,13 +40,9 @@ static DRESULT sc64_sd_fs_load_with_dma(BYTE pdrv, FSIZE_t offset, LBA_t sector,
 }
 
 
-static bool fs_initialized = false;
-static FATFS fatfs;
-
-
 sc64_sd_fs_error_t sc64_sd_fs_init(void) {
     FRESULT fresult;
-    
+
     fresult = f_mount(&fatfs, "", 1);
     if (fresult != FR_OK) {
         switch (fresult) {