diff --git a/src/DMA.cpp b/src/DMA.cpp
index 80cd592c..56ec1564 100644
--- a/src/DMA.cpp
+++ b/src/DMA.cpp
@@ -187,16 +187,20 @@ void DMA::Start()
 
     // TODO eventually: not stop if we're running code in ITCM
 
-    Running = 2;
+    Running = 3;
 
     // safety measure
     MRAMBurstTable = DMATiming::MRAMDummy;
 
     InProgress = true;
     NDS.StopCPU(CPU, 1<<Num);
+
+    if (Num == 0) NDS.DMAs[(CPU*4)+1].ResetBurst();
+    if (Num <= 1) NDS.DMAs[(CPU*4)+2].ResetBurst();
+    if (Num <= 2) NDS.DMAs[(CPU*4)+3].ResetBurst();
 }
 
-u32 DMA::UnitTimings9_16(bool burststart)
+u32 DMA::UnitTimings9_16(u8 burststart)
 {
     u32 src_id = CurSrcAddr >> 14;
     u32 dst_id = CurDstAddr >> 14;
@@ -213,11 +217,13 @@ u32 DMA::UnitTimings9_16(bool burststart)
     if (src_rgn == Mem9_MainRAM)
     {
         if (dst_rgn == Mem9_MainRAM)
-            return 16;
+        {
+            return (burststart == 2) ? 11 : 16;
+        }
 
         if (SrcAddrInc > 0)
         {
-            if (burststart || MRAMBurstTable[MRAMBurstCount] == 0)
+            if ((burststart == 2) || MRAMBurstTable[MRAMBurstCount] == 0)
             {
                 MRAMBurstCount = 0;
 
@@ -239,14 +245,14 @@ u32 DMA::UnitTimings9_16(bool burststart)
         {
             // TODO: not quite right for GBA slot
             return (((CurSrcAddr & 0x1F) == 0x1E) ? 7 : 8) +
-                   (burststart ? dst_n : dst_s);
+                   ((burststart == 2) ? dst_n : dst_s);
         }
     }
     else if (dst_rgn == Mem9_MainRAM)
     {
         if (DstAddrInc > 0)
         {
-            if (burststart || MRAMBurstTable[MRAMBurstCount] == 0)
+            if ((burststart == 2) || MRAMBurstTable[MRAMBurstCount] == 0)
             {
                 MRAMBurstCount = 0;
 
@@ -266,23 +272,26 @@ u32 DMA::UnitTimings9_16(bool burststart)
         }
         else
         {
-            return (burststart ? src_n : src_s) + 7;
+            return ((burststart == 2) ? src_n : src_s) + 7;
         }
     }
     else if (src_rgn & dst_rgn)
     {
-        return src_n + dst_n + 1;
+        if (burststart != 1)
+            return src_n + dst_n + (src_n == 1 || burststart <= 0);
+        else
+            return src_n + dst_n + (src_n != 1);
     }
     else
     {
-        if (burststart)
-            return src_n + dst_n;
+        if (burststart == 2)
+            return src_n + dst_n + (src_n == 1);
         else
             return src_s + dst_s;
     }
 }
 
-u32 DMA::UnitTimings9_32(bool burststart)
+u32 DMA::UnitTimings9_32(u8 burststart)
 {
     u32 src_id = CurSrcAddr >> 14;
     u32 dst_id = CurDstAddr >> 14;
@@ -299,11 +308,11 @@ u32 DMA::UnitTimings9_32(bool burststart)
     if (src_rgn == Mem9_MainRAM)
     {
         if (dst_rgn == Mem9_MainRAM)
-            return 18;
+            return (burststart == 2) ? 13 : 18;
 
         if (SrcAddrInc > 0)
         {
-            if (burststart || MRAMBurstTable[MRAMBurstCount] == 0)
+            if ((burststart == 2) || MRAMBurstTable[MRAMBurstCount] == 0)
             {
                 MRAMBurstCount = 0;
 
@@ -327,14 +336,14 @@ u32 DMA::UnitTimings9_32(bool burststart)
         {
             // TODO: not quite right for GBA slot
             return (((CurSrcAddr & 0x1F) == 0x1C) ? (dst_n==2 ? 7:8) : 9) +
-                   (burststart ? dst_n : dst_s);
+                   ((burststart == 2) ? dst_n : dst_s);
         }
     }
     else if (dst_rgn == Mem9_MainRAM)
     {
         if (DstAddrInc > 0)
         {
-            if (burststart || MRAMBurstTable[MRAMBurstCount] == 0)
+            if ((burststart == 2) || MRAMBurstTable[MRAMBurstCount] == 0)
             {
                 MRAMBurstCount = 0;
 
@@ -356,17 +365,20 @@ u32 DMA::UnitTimings9_32(bool burststart)
         }
         else
         {
-            return (burststart ? src_n : src_s) + 8;
+            return ((burststart == 2) ? src_n : src_s) + 8;
         }
     }
     else if (src_rgn & dst_rgn)
     {
-        return src_n + dst_n + 1;
+        if (burststart != 1)
+            return src_n + dst_n + (src_n == 1 || burststart <= 0);
+        else
+            return src_n + dst_n + (src_n != 1);
     }
     else
     {
-        if (burststart)
-            return src_n + dst_n;
+        if (burststart == 2)
+            return src_n + dst_n + (src_n == 1);
         else
             return src_s + dst_s;
     }
@@ -557,15 +569,17 @@ void DMA::Run9()
     Executing = true;
 
     // add NS penalty for first accesses in burst
-    bool burststart = (Running == 2);
-    Running = 1;
+    int burststart = Running-1;
+    Running = 2;
+
+    NDS.ARM9Timestamp = (NDS.ARM9Timestamp + ((1<<NDS.ARM9ClockShift)-1)) & ~((1<<NDS.ARM9ClockShift)-1);
 
     if (!(Cnt & (1<<26)))
     {
         while (IterCount > 0 && !Stall)
         {
             NDS.ARM9Timestamp += (UnitTimings9_16(burststart) << NDS.ARM9ClockShift);
-            burststart = false;
+            burststart -= 1;
 
             NDS.ARM9Write16(CurDstAddr, NDS.ARM9Read16(CurSrcAddr));
 
@@ -582,7 +596,7 @@ void DMA::Run9()
         while (IterCount > 0 && !Stall)
         {
             NDS.ARM9Timestamp += (UnitTimings9_32(burststart) << NDS.ARM9ClockShift);
-            burststart = false;
+            burststart -= 1;
 
             NDS.ARM9Write32(CurDstAddr, NDS.ARM9Read32(CurSrcAddr));
 
@@ -595,6 +609,8 @@ void DMA::Run9()
         }
     }
 
+    if (burststart == 1) Running = 1;  
+
     Executing = false;
     Stall = false;
 
diff --git a/src/DMA.h b/src/DMA.h
index 354f4495..64d5647f 100644
--- a/src/DMA.h
+++ b/src/DMA.h
@@ -40,8 +40,8 @@ public:
     void WriteCnt(u32 val);
     void Start();
 
-    u32 UnitTimings9_16(bool burststart);
-    u32 UnitTimings9_32(bool burststart);
+    u32 UnitTimings9_16(u8 burststart);
+    u32 UnitTimings9_32(u8 burststart);
     u32 UnitTimings7_16(bool burststart);
     u32 UnitTimings7_32(bool burststart);
 
@@ -73,6 +73,11 @@ public:
         if (Executing) Stall = true;
     }
 
+    void ResetBurst()
+    {
+        if (Running > 0) Running = (CPU ? 2 : 3);
+    }
+
     u32 SrcAddr {};
     u32 DstAddr {};
     u32 Cnt {};
diff --git a/src/DMA_Timings.cpp b/src/DMA_Timings.cpp
index a51fedfb..02539a62 100644
--- a/src/DMA_Timings.cpp
+++ b/src/DMA_Timings.cpp
@@ -48,7 +48,7 @@ extern const std::array<u8, 256> MRAMDummy = {0};
 extern const std::array<u8, 256> MRAMRead16Bursts[] =
 {
     // main RAM to regular 16bit or 32bit bus (similar)
-    {7, 3, 2, 2, 2, 2, 2, 2, 2, 2,
+    {6, 3, 2, 2, 2, 2, 2, 2, 2, 2,
      2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
      2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
      2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
@@ -60,7 +60,7 @@ extern const std::array<u8, 256> MRAMRead16Bursts[] =
      2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
      2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
      2, 2, 2, 2, 2, 2, 2, 2, 2,
-     7, 3, 2, 2, 2, 2, 2, 2, 2, 2,
+     6, 3, 2, 2, 2, 2, 2, 2, 2, 2,
      2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
      2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
      2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
@@ -72,7 +72,7 @@ extern const std::array<u8, 256> MRAMRead16Bursts[] =
      2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
      2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
      2, 2, 2, 2, 2, 2, 2, 2, 2,
-     7, 3,
+     6, 3,
      0},
     // main RAM to GBA/wifi, seq=4
     {8, 6, 5, 5, 5, 5, 5, 5, 5, 5,
@@ -181,7 +181,7 @@ extern const std::array<u8, 256> MRAMRead32Bursts[] =
 extern const std::array<u8, 256> MRAMWrite16Bursts[] =
 {
     // regular 16bit or 32bit bus to main RAM (similar)
-    {8, 2, 2, 2, 2, 2, 2, 2, 2, 2,
+    {5, 2, 2, 2, 2, 2, 2, 2, 2, 2,
      2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
      2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
      2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
@@ -212,7 +212,7 @@ extern const std::array<u8, 256> MRAMWrite16Bursts[] =
 extern const std::array<u8, 256> MRAMWrite32Bursts[4] =
 {
     // regular 16bit bus to main RAM
-    {9, 4, 4, 4, 4, 4, 4, 4, 4, 4,
+    {6, 4, 4, 4, 4, 4, 4, 4, 4, 4,
      4, 4, 4, 4, 4, 4, 4, 4, 4, 4,
      4, 4, 4, 4, 4, 4, 4, 4, 4, 4,
      4, 4, 4, 4, 4, 4, 4, 4, 4, 4,
@@ -220,7 +220,7 @@ extern const std::array<u8, 256> MRAMWrite32Bursts[4] =
      4, 4, 4, 4, 4, 4, 4, 4, 4, 4,
      0},
     // regular 32bit bus to main RAM
-    {9, 3, 3, 3, 3, 3, 3, 3, 3, 3,
+    {6, 3, 3, 3, 3, 3, 3, 3, 3, 3,
      3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
      3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
      3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
diff --git a/src/NDS.h b/src/NDS.h
index e23b1f27..c1f0ff88 100644
--- a/src/NDS.h
+++ b/src/NDS.h
@@ -309,6 +309,7 @@ public: // TODO: Encapsulate the rest of these members
     GBACart::GBACartSlot GBACartSlot;
     melonDS::GPU GPU;
     melonDS::AREngine AREngine;
+    DMA DMAs[8];
 
 #ifdef JIT_ENABLED
     bool IsJITEnabled(){return EnableJIT;};
@@ -494,7 +495,6 @@ private:
     u16 WifiWaitCnt;
     u8 TimerCheckMask[2];
     u64 TimerTimestamp[2];
-    DMA DMAs[8];
     u32 DMA9Fill[4];
     u16 IPCSync9, IPCSync7;
     u16 IPCFIFOCnt9, IPCFIFOCnt7;