diff --git a/src/ARM.cpp b/src/ARM.cpp
index d71d0e1a..226b4630 100644
--- a/src/ARM.cpp
+++ b/src/ARM.cpp
@@ -48,6 +48,8 @@ ARM::ARM(u32 num)
     // well uh
     Num = num;
 
+    SetClockShift(0); // safe default
+
     for (int i = 0; i < 16; i++)
     {
         Waitstates[0][i] = 1;
@@ -432,18 +434,9 @@ s32 ARM::Execute()
             }
         }
 
-        if (Num==0)
-        {
-            s32 diff = Cycles - lastcycles;
-            NDS::RunTimingCriticalDevices(0, diff >> 1);
-            lastcycles = Cycles - (diff&1);
-        }
-        else
-        {
-            s32 diff = Cycles - lastcycles;
-            NDS::RunTimingCriticalDevices(1, diff);
-            lastcycles = Cycles;
-        }
+        s32 diff = Cycles - lastcycles;
+        NDS::RunTimingCriticalDevices(Num, diff >> ClockShift);
+        lastcycles = Cycles - (diff & ClockDiffMask);
 
         // TODO optimize this shit!!!
         if (Halted)
diff --git a/src/ARM.h b/src/ARM.h
index b778be43..f06cb7f3 100644
--- a/src/ARM.h
+++ b/src/ARM.h
@@ -38,6 +38,12 @@ public:
 
     void Reset();
 
+    void SetClockShift(u32 shift)
+    {
+        ClockShift = shift;
+        ClockDiffMask = (1<<shift) - 1;
+    }
+
     void DoSavestate(Savestate* file);
 
     void JumpTo(u32 addr, bool restorecpsr = false);
@@ -227,6 +233,11 @@ public:
 
     u32 Num;
 
+    // shift relative to system clock
+    // 0=33MHz 1=66MHz 2=133MHz
+    u32 ClockShift;
+    u32 ClockDiffMask;
+
     // waitstates:
     // 0=code16 1=code32 2=data16 3=data32
     // TODO eventually: nonsequential waitstates
diff --git a/src/DMA.cpp b/src/DMA.cpp
index f0c22b59..e88814eb 100644
--- a/src/DMA.cpp
+++ b/src/DMA.cpp
@@ -26,6 +26,25 @@
 // NOTES ON DMA SHIT
 //
 // * could use optimized code paths for common types of DMA transfers. for example, VRAM
+//   have to profile it to see if it's actually worth doing
+
+
+// DMA TIMINGS
+//
+// sequential timing:
+// * 1 cycle per read or write
+// * in 32bit mode, accessing a 16bit bus (mainRAM, palette, VRAM) incurs 1 cycle of penalty
+// * in 32bit mode, transferring from mainRAM to another bank is 1 cycle faster
+// * if source and destination are the same memory bank, there is a 1 cycle penalty
+// * transferring from mainRAM to mainRAM is a trainwreck (all accesses are made nonsequential)
+//
+// nonsequential timing:
+// * nonseq penalty is applied to the first read and write
+// * I also figure it gets nonseq penalty again when resuming, after having been interrupted by
+//   another DMA (TODO: check)
+// * applied to all accesses for mainRAM->mainRAM, resulting in timings of 16-18 cycles per unit
+//
+// TODO: GBA slot
 
 
 DMA::DMA(u32 cpu, u32 num)
diff --git a/src/NDS.cpp b/src/NDS.cpp
index 77810eef..32deeb6d 100644
--- a/src/NDS.cpp
+++ b/src/NDS.cpp
@@ -345,6 +345,9 @@ void Reset()
     SPI::Reset();
     RTC::Reset();
     Wifi::Reset();
+
+    ARM9->SetClockShift(1);
+    ARM7->SetClockShift(0);
 }
 
 void Stop()