diff --git a/src-md/Makefile b/src-md/Makefile
index a7bcc8d35..e34c78c84 100644
--- a/src-md/Makefile
+++ b/src-md/Makefile
@@ -10,7 +10,7 @@ LIBPATH = -L$(ROOTDIR)/m68k-elf/lib -L$(ROOTDIR)/m68k-elf/lib/gcc/m68k-elf/12.1.
 INCPATH = -I$. -I$(ROOTDIR)/m68k-elf/include -I$(ROOTDIR)/m68k-elf/m68k-elf/include -I../liblzss
 
 HWCCFLAGS = -m68000 -Wall -O1 -c -fomit-frame-pointer -fno-lto
-CCFLAGS = -m68000 -Wall -Os -c -fomit-frame-pointer -flto
+CCFLAGS = -m68000 -Wall -Os -c -fomit-frame-pointer -fno-lto -DUSE_SCD_IN_RAM
 ASFLAGS = -m68000 --register-prefix-optional
 LINKFLAGS = -T $(LDSCRIPTSDIR)/mars-md.ld -Wl,-Map=output.map -nostdlib -flto
 
diff --git a/src-md/crt0.s b/src-md/crt0.s
index 60576e3f4..24a7349f2 100644
--- a/src-md/crt0.s
+++ b/src-md/crt0.s
@@ -276,6 +276,10 @@ init_hardware:
         rts
 
 
+        .data
+
+| Put remaining code in data section to lower bus contention for the rom.
+
 | void write_byte(void *dst, unsigned char val)
         .global write_byte
 write_byte:
@@ -321,11 +325,6 @@ read_long:
         move.l  (a0),d0
         rts
 
-
-        .data
-
-| Put remaining code in data section to lower bus contention for the rom.
-
         .global do_main
 do_main:
         move.b  #1,0xA15107         /* set RV */
diff --git a/src-md/scd_pcm.c b/src-md/scd_pcm.c
index db08b172e..a3b32687b 100644
--- a/src-md/scd_pcm.c
+++ b/src-md/scd_pcm.c
@@ -22,19 +22,27 @@ extern unsigned int read_long(unsigned int src);
 static scd_cmd_t scd_cmds[MAX_SCD_CMDS];
 static int16_t num_scd_cmds;
 
+static void scd_delay(void) SCD_CODE_ATTR;
+static char wait_cmd_ack(void) SCD_CODE_ATTR;
+static void wait_do_cmd(char cmd) SCD_CODE_ATTR;
+
 static char wait_cmd_ack(void)
 {
     char ack = 0;
 
-    while (!ack)
+    do {
+        scd_delay();
         ack = read_byte(0xA1200F); // wait for acknowledge byte in sub comm port
+    } while (!ack);
 
     return ack;
 }
 
 static void wait_do_cmd(char cmd)
 {
-    while (read_byte(0xA1200F)) ; // wait until Sub-CPU is ready to receive command
+    while (read_byte(0xA1200F)) {
+        scd_delay(); // wait until Sub-CPU is ready to receive command
+    }
     write_byte(0xA1200E, cmd); // set main comm port to command
 }
 
@@ -222,3 +230,11 @@ int scd_flush_cmd_queue(void)
     num_scd_cmds = 0;
     return i;
 }
+
+static void scd_delay(void)
+{
+    int cnt = 5;
+    do {
+        asm __volatile("nop");
+    } while (--cnt);
+}
diff --git a/src-md/scd_pcm.h b/src-md/scd_pcm.h
index b9ecbfc76..770752727 100644
--- a/src-md/scd_pcm.h
+++ b/src-md/scd_pcm.h
@@ -1,5 +1,11 @@
 #include <stdint.h>
 
+#ifdef USE_SCD_IN_RAM
+#define SCD_CODE_ATTR __attribute__((section(".data"), aligned(16)))
+#else
+#define SCD_CODE_ATTR
+#endif
+
 // scd_init_pcm initializes the PCM driver
 void scd_init_pcm(void);
 
@@ -16,7 +22,7 @@ void scd_init_pcm(void);
 // otherwise a new memory block will be allocated from the available memory pool
 // once the driver runs out of memory, no further allocations will be possible and
 // the driver will have to be re-initialized by calling scd_init_pcm 
-void scd_upload_buf(uint16_t buf_id, const uint8_t *data, uint32_t data_len);
+void scd_upload_buf(uint16_t buf_id, const uint8_t *data, uint32_t data_len) SCD_CODE_ATTR;
 
 // scd_upload_buf starts playback on source from the start of the buffer
 // source is a virtual playback channel, one or two hardware channels can be mapped
@@ -33,12 +39,12 @@ void scd_upload_buf(uint16_t buf_id, const uint8_t *data, uint32_t data_len);
 // returned value: if the passed src_id is 255, then thew newly allocated source id,
 // if the allocation has failed, a value of 0 is returned
 // otherwise the originally passed value of src_id is returned
-uint8_t scd_play_src(uint8_t src_id, uint16_t buf_id, uint16_t freq, uint8_t pan, uint8_t vol, uint8_t autoloop);
+uint8_t scd_play_src(uint8_t src_id, uint16_t buf_id, uint16_t freq, uint8_t pan, uint8_t vol, uint8_t autoloop) SCD_CODE_ATTR;
 
 // scd_punpause_src pauses or unpauses the source
 //
 // value range for src_id: [1, 8]
-uint8_t scd_punpause_src(uint8_t src_id, uint8_t paused);
+uint8_t scd_punpause_src(uint8_t src_id, uint8_t paused) SCD_CODE_ATTR;
 
 // scd_update_src updates the frequency, panning, volume and autoloop property for the source
 //
@@ -48,44 +54,44 @@ uint8_t scd_punpause_src(uint8_t src_id, uint8_t paused);
 // values for vol: [0, 255]
 // values for autoloop: [0, 255], a boolean: the source will automatically loopf from the start after
 // reaching the end of the playback buffer
-void scd_update_src(uint8_t src_id, uint16_t freq, uint8_t pan, uint8_t vol, uint8_t autoloop);
+void scd_update_src(uint8_t src_id, uint16_t freq, uint8_t pan, uint8_t vol, uint8_t autoloop) SCD_CODE_ATTR;
 
 // scd_stop_src stops playback on the given source
 //
 // value range for src_id: [1, 8]
-void scd_stop_src(uint8_t src_id);
+void scd_stop_src(uint8_t src_id) SCD_CODE_ATTR;
 
 // scd_rewind_src sets position for the given source to the start of the playback buffer
 //
 // value range for src_id: [1, 8]
-void scd_rewind_src(uint8_t src_id);
+void scd_rewind_src(uint8_t src_id) SCD_CODE_ATTR;
 
 // scd_getpos_for_src returns playback position for the given source
 //
 // value range for src_id: [1, 8]
 //
 // returned value: current read position in PCM memory of the ricoh chip for the first channel of the source
-uint16_t scd_getpos_for_src(uint8_t src_id);
+uint16_t scd_getpos_for_src(uint8_t src_id) SCD_CODE_ATTR;
 
 // scd_clear_pcm stops playback on all channels
-void scd_clear_pcm(void);
+void scd_clear_pcm(void) SCD_CODE_ATTR;
 
 // returns playback status mask for all sources
 // if a source is active, it will have its bit set to 1 in the mask:
 // bit 0 for source id 1, bit 1 for source id 2, etc
-int scd_get_playback_status(void);
+int scd_get_playback_status(void) SCD_CODE_ATTR;
 
 // queues a scd_play_src call, always returns 0
-uint8_t scd_queue_play_src(uint8_t src_id, uint16_t buf_id, uint16_t freq, uint8_t pan, uint8_t vol, uint8_t autoloop);
+uint8_t scd_queue_play_src(uint8_t src_id, uint16_t buf_id, uint16_t freq, uint8_t pan, uint8_t vol, uint8_t autoloop) SCD_CODE_ATTR;
 
 // queues a scd_update_src call
-void scd_queue_update_src(uint8_t src_id, uint16_t freq, uint8_t pan, uint8_t vol, uint8_t autoloop);
+void scd_queue_update_src(uint8_t src_id, uint16_t freq, uint8_t pan, uint8_t vol, uint8_t autoloop) SCD_CODE_ATTR;
 
 // queues a scd_stop_src call
-void scd_queue_stop_src(uint8_t src_id);
+void scd_queue_stop_src(uint8_t src_id) SCD_CODE_ATTR;
 
 // queues a scd_clear_pcm call
-void scd_queue_clear_pcm(void);
+void scd_queue_clear_pcm(void) SCD_CODE_ATTR;
 
 // flushes the command queue
-int scd_flush_cmd_queue(void);
+int scd_flush_cmd_queue(void) SCD_CODE_ATTR;