diff --git a/sw/n64/src/exception.S b/sw/n64/src/exception.S
index ab3d16e..c7caa23 100644
--- a/sw/n64/src/exception.S
+++ b/sw/n64/src/exception.S
@@ -1,7 +1,7 @@
 #include "vr4300.h"
 
 
-#define WATCHDOG_TIMEOUT            (5 * (93750000UL / 2))
+#define WATCHDOG_TIMEOUT            (3 * (93750000UL / 2))
 
 #define VECTOR_LOCATION             (0xA0000000UL)
 #define VECTOR_SIZE                 (0x80)
diff --git a/sw/n64/src/exception.c b/sw/n64/src/exception.c
index 927f722..c965c05 100644
--- a/sw/n64/src/exception.c
+++ b/sw/n64/src/exception.c
@@ -276,7 +276,7 @@ void exception_fatal_handler (uint32_t exception_code, uint32_t interrupt_mask,
     if (exception_code == EXCEPTION_INTERRUPT) {
         if (interrupt_mask & INTERRUPT_MASK_TIMER) {
             exception_disable_watchdog();
-            exception_print("Still loading after 5 second limit...\n\n");
+            exception_print("Still loading after 3 second limit...\n\n");
             return;
         }
     } else if (exception_code == EXCEPTION_SYSCALL) {
diff --git a/sw/n64/src/io.h b/sw/n64/src/io.h
index 8e8eb4e..e6897cf 100644
--- a/sw/n64/src/io.h
+++ b/sw/n64/src/io.h
@@ -230,8 +230,8 @@ typedef struct {
     io32_t DATA[2];
     io32_t VERSION;
     io32_t __padding[4092];
-    io32_t CPU_RAM[3968];
-    io32_t BUFFER[128];
+    io32_t CPU_RAM[3840];
+    io32_t BUFFER[256];
 } sc64_regs_t;
 
 #define SC64_BASE                   (0x1FFF0000UL)
diff --git a/sw/n64/src/sc64.c b/sw/n64/src/sc64.c
index cef551c..34761cb 100644
--- a/sw/n64/src/sc64.c
+++ b/sw/n64/src/sc64.c
@@ -63,7 +63,7 @@ void sc64_init (void) {
 }
 
 static uint32_t sc64_wait_drive_ready (drive_id_t drive) {
-    uint32_t args[2] = { drive, 0 };
+    uint32_t args[2] = { (drive & 0xFF), 0 };
     uint32_t result[2];
     do {
         sc64_perform_cmd(SC64_CMD_DRIVE_BUSY, args, result);
@@ -72,46 +72,44 @@ static uint32_t sc64_wait_drive_ready (drive_id_t drive) {
 }
 
 bool sc64_storage_init (drive_id_t drive) {
-    uint32_t args[2] = { drive, 0 };
+    uint32_t args[2] = { (drive & 0xFF), 0 };
     if (sc64_perform_cmd(SC64_CMD_DRIVE_INIT, args, NULL)) {
         return true;
     }
-    if (sc64_wait_drive_ready(drive) != 0) {
+    if (sc64_wait_drive_ready(drive)) {
         return true;
     }
     return false;
 }
 
-static bool sc64_drive_start_read_and_wait (drive_id_t drive, uint32_t sector) {
-    uint32_t args[2] = { drive, sector };
-    if (sc64_perform_cmd(SC64_CMD_DRIVE_READ, args, NULL)) {
-        return true;
-    }
-    if (sc64_wait_drive_ready(drive) != 0) {
-        return true;
-    }
-    return false;
-}
-
-static bool sc64_drive_start_write_and_wait (drive_id_t drive, uint32_t sector) {
-    uint32_t args[2] = { drive, sector };
-    if (sc64_perform_cmd(SC64_CMD_DRIVE_WRITE, args, NULL)) {
-        return true;
-    }
-    if (sc64_wait_drive_ready(drive) != 0) {
+static bool sc64_drive_start_rw (drive_id_t drive, bool write, uint32_t sector, uint32_t offset) {
+    uint32_t args[2] = { (((offset & 0xFFFFFF) << 8) | (drive & 0xFF)), sector };
+    if (sc64_perform_cmd(write ? SC64_CMD_DRIVE_WRITE : SC64_CMD_DRIVE_READ, args, NULL)) {
         return true;
     }
     return false;
 }
 
 bool sc64_storage_read (drive_id_t drive, void *buffer, uint32_t sector, uint32_t count) {
-    io32_t *src = SC64->BUFFER;
+    io32_t *src;
     uint8_t *dst = (uint8_t *) (buffer);
+    uint32_t current_offset = 0;
+    uint32_t next_offset = SECTOR_SIZE;
 
+    if (sc64_drive_start_rw(drive, false, sector++, 0)) {
+        return true;
+    }
     while (count > 0) {
-        if (sc64_drive_start_read_and_wait(drive, sector)) {
+        if (sc64_wait_drive_ready(drive)) {
             return true;
         }
+        if (count > 1) {
+            if (sc64_drive_start_rw(drive, false, sector++, next_offset)) {
+                return true;
+            }
+            next_offset = next_offset ? 0 : SECTOR_SIZE;
+        }
+        src = &SC64->BUFFER[current_offset / sizeof(io32_t)];
         for (int i = 0; i < (SECTOR_SIZE / sizeof(uint32_t)); i++) {
             uint32_t data = pi_io_read(src + i);
             *dst++ = ((data >> 24) & 0xFF);
@@ -119,7 +117,7 @@ bool sc64_storage_read (drive_id_t drive, void *buffer, uint32_t sector, uint32_
             *dst++ = ((data >> 8) & 0xFF);
             *dst++ = ((data >> 0) & 0xFF);
         }
-        sector += 1;
+        current_offset = current_offset ? 0 : SECTOR_SIZE;
         count -= 1;
     }
 
@@ -139,7 +137,10 @@ bool sc64_storage_write (drive_id_t drive, const void *buffer, uint32_t sector,
             data |= ((*src++) << 0);
             pi_io_write((dst + i), data);
         }
-        if (sc64_drive_start_write_and_wait(drive, sector)) {
+        if (sc64_drive_start_rw(drive, true, sector, 0)) {
+            return true;
+        }
+        if (sc64_wait_drive_ready(drive)) {
             return true;
         }
         sector += 1;
diff --git a/sw/pc/sc64.py b/sw/pc/sc64.py
index d06ecbf..7567550 100644
--- a/sw/pc/sc64.py
+++ b/sw/pc/sc64.py
@@ -542,22 +542,24 @@ class SC64:
 
     def __debug_process_fsd_read(self, data: bytes) -> None:
         sector = int.from_bytes(data[0:4], byteorder='little')
+        offset = int.from_bytes(data[4:8], byteorder='little')
         if (self.__fsd_file):
             self.__fsd_file.seek(sector * 512)
-            self.__write_cmd("T", 0, 512)
+            self.__write_cmd("T", offset, 512)
             self.__write(self.__fsd_file.read(512))
         else:
-            self.__write_cmd("T", 0, 0)
+            self.__write_cmd("T", offset, 0)
 
 
     def __debug_process_fsd_write(self, data: bytes) -> None:
         sector = int.from_bytes(data[0:4], byteorder='little')
+        offset = int.from_bytes(data[4:8], byteorder='little')
         if (self.__fsd_file):
             self.__fsd_file.seek(sector * 512)
-            self.__write_cmd("F", 0, 512)
+            self.__write_cmd("F", offset, 512)
             self.__fsd_file.write(self.__read(512))
         else:
-            self.__write_cmd("F", 0, 0)
+            self.__write_cmd("F", offset, 0)
 
 
     def __debug_process_dd_block(self, data: bytes) -> None:
diff --git a/sw/riscv/SC64.ld b/sw/riscv/SC64.ld
index 1573836..69d4601 100644
--- a/sw/riscv/SC64.ld
+++ b/sw/riscv/SC64.ld
@@ -1,6 +1,6 @@
 MEMORY {
-    ram (rwx)   : org = 0x00000000, len = 16k - 512
-    buffer (rw) : org = 0x00003E00, len = 512
+    ram (rwx)   : org = 0x00000000, len = 15k
+    buffer (rw) : org = 0x00003C00, len = 1k
     rom (rx)    : org = 0x10010000, len = 26k
 }
 
diff --git a/sw/riscv/src/cfg.c b/sw/riscv/src/cfg.c
index cfa0d44..afb635c 100644
--- a/sw/riscv/src/cfg.c
+++ b/sw/riscv/src/cfg.c
@@ -310,7 +310,8 @@ void process_cfg (void) {
                     event.id = EVENT_ID_FSD_READ;
                     event.trigger = CALLBACK_BUFFER_WRITE;
                     event.callback = set_usb_drive_not_busy;
-                    if (usb_put_event(&event, &args[1], sizeof(args[1]))) {
+                    uint32_t data[2] = { args[1], (args[0] >> 8) };
+                    if (usb_put_event(&event, data, sizeof(data))) {
                         p.usb_drive_busy = true;
                     } else {
                         return;
@@ -328,7 +329,8 @@ void process_cfg (void) {
                     event.id = EVENT_ID_FSD_WRITE;
                     event.trigger = CALLBACK_BUFFER_READ;
                     event.callback = set_usb_drive_not_busy;
-                    if (usb_put_event(&event, &args[1], sizeof(args[1]))) {
+                    uint32_t data[2] = { args[1], (args[0] >> 8) };
+                    if (usb_put_event(&event, data, sizeof(data))) {
                         p.usb_drive_busy = true;
                     } else {
                         return;
diff --git a/sw/riscv/src/sys.h b/sw/riscv/src/sys.h
index 4d84f9d..380aafa 100644
--- a/sw/riscv/src/sys.h
+++ b/sw/riscv/src/sys.h
@@ -22,7 +22,7 @@ typedef volatile uint32_t           io32_t;
 
 
 #define RAM_BASE                    (0x00000000UL)
-#define RAMBUFFER_SIZE              (512)
+#define RAMBUFFER_SIZE              (1024)
 #define RAM_SIZE                    ((16 * 1024) - RAMBUFFER_SIZE)
 #define RAMBUFFER_BASE              (RAM_BASE + RAM_SIZE)
 #define RAM                         (*((io32_t *) RAM_BASE))