diff --git a/Assets/dll/virtualjaguar.wbx.zst b/Assets/dll/virtualjaguar.wbx.zst
index a37f282365..6754c9057d 100644
Binary files a/Assets/dll/virtualjaguar.wbx.zst and b/Assets/dll/virtualjaguar.wbx.zst differ
diff --git a/waterbox/virtualjaguar/src/cdhle.cpp b/waterbox/virtualjaguar/src/cdhle.cpp
index 37308a65c9..7bf0ae8369 100644
--- a/waterbox/virtualjaguar/src/cdhle.cpp
+++ b/waterbox/virtualjaguar/src/cdhle.cpp
@@ -111,7 +111,7 @@ void CDHLEInit(void)
 					cd_boot_lba = startLba + i;
 					cd_boot_off = j + 32 + 4 + 4;
 					cd_byte_swapped = false;
-					cd_word_alignment = (4 - j) & 3;
+					cd_word_alignment = -j & 3;
 					foundHeader = true;
 					break;
 				}
@@ -124,7 +124,7 @@ void CDHLEInit(void)
 					cd_boot_lba = startLba + i;
 					cd_boot_off = j + 32 + 4 + 4;
 					cd_byte_swapped = true;
-					cd_word_alignment = (4 - j) & 3;
+					cd_word_alignment = -j & 3;
 					foundHeader = true;
 					break;
 				}
@@ -244,10 +244,10 @@ static void CDSendBlock(void)
 		cd_buf_rm += 2352;
 	}
 
-	// send one block of data
-	for (uint32_t i = 0; i < 64; i++)
+	// send one block of data, one long at a time
+	for (uint32_t i = 0; i < 64; i += 4)
 	{
-		JaguarWriteByte(cd_read_addr_start + i, cd_buf2352[cd_buf_pos + i], GPU);
+		GPUWriteLong(cd_read_addr_start + i, GET32(cd_buf2352, cd_buf_pos + i), GPU);
 	}
 
 	cd_read_addr_start += 64;
@@ -258,8 +258,7 @@ static void CDSendBlock(void)
 	{
 		cd_is_reading = false;
 	}
-
-	if (cd_buf_circular_size && (cd_read_addr_end - cd_read_addr_start) >= cd_buf_circular_size)
+	else if (cd_buf_circular_size && (cd_read_addr_start - cd_read_orig_addr_start) >= cd_buf_circular_size)
 	{
 		cd_read_addr_start = cd_read_orig_addr_start;
 	}
@@ -270,12 +269,15 @@ static void CDHLECallback(void)
 	RemoveCallback(CDHLECallback);
 	if (cd_is_reading)
 	{
-		if (!cd_paused)
+		if (!GPURunning())
+			fprintf(stderr, "CDHLECallback called with GPU inactive\n");
+
+		if (GPURunning() && !cd_paused)
 		{
 			CDSendBlock();
+			//GPUSetIRQLine(GPUIRQ_DSP, ASSERT_LINE);
 		}
-		//GPUSetIRQLine(GPUIRQ_DSP, ASSERT_LINE);
-		SetCallbackTime(CDHLECallback, 240 >> (cd_mode & 1));
+		SetCallbackTime(CDHLECallback, 285 >> (cd_mode & 1));
 	}
 }
 
@@ -366,14 +368,6 @@ static void CD_mode(void)
 
 static void CD_ack(void)
 {
-	if (!cd_paused)
-	{
-		while (cd_is_reading)
-		{
-			CDSendBlock();
-		}
-	}
-
 	NO_ERR();
 }
 
@@ -521,8 +515,9 @@ static void CD_read(void)
 						cd_buf_rm = bufRm;
 						cd_buf_circular_size = circBufSz ? (1 << circBufSz) : 0;
 						RemoveCallback(CDHLECallback);
-						SetCallbackTime(CDHLECallback, 240 >> (cd_mode & 1));
+						SetCallbackTime(CDHLECallback, 285 >> (cd_mode & 1));
 						JERRYWriteWord(0xF10020, 0, M68K);
+						//GPUWriteLong(0xF02100, GPUReadLong(0xF02100, M68K) | 0x20, M68K);
 						break;
 					}
 				}
@@ -542,8 +537,9 @@ static void CD_read(void)
 			cd_buf_rm = 0;
 			cd_buf_circular_size = 0;
 			RemoveCallback(CDHLECallback);
-			SetCallbackTime(CDHLECallback, 240 >> (cd_mode & 1));
+			SetCallbackTime(CDHLECallback, 285 >> (cd_mode & 1));
 			JERRYWriteWord(0xF10020, 0, M68K);
+			//GPUWriteLong(0xF02100, GPUReadLong(0xF02100, M68K) | 0x20, M68K);
 		}
 	}
 
diff --git a/waterbox/virtualjaguar/src/dsp.cpp b/waterbox/virtualjaguar/src/dsp.cpp
index 6089f49127..b0a0393906 100644
--- a/waterbox/virtualjaguar/src/dsp.cpp
+++ b/waterbox/virtualjaguar/src/dsp.cpp
@@ -784,7 +784,7 @@ static void dsp_opcode_nop(void)
 static void dsp_opcode_storeb(void)
 {
 	if (RM >= DSP_WORK_RAM_BASE && RM <= (DSP_WORK_RAM_BASE + 0x1FFF))
-		DSPWriteLong(RM, RN & 0xFF, DSP);
+		DSPWriteLong(RM & 0xFFFFFFFC, RN, DSP);
 	else
 		JaguarWriteByte(RM, RN, DSP);
 }
@@ -792,7 +792,7 @@ static void dsp_opcode_storeb(void)
 static void dsp_opcode_storew(void)
 {
 	if (RM >= DSP_WORK_RAM_BASE && RM <= (DSP_WORK_RAM_BASE + 0x1FFF))
-		DSPWriteLong(RM, RN & 0xFFFF, DSP);
+		DSPWriteLong(RM & 0xFFFFFFFC, RN, DSP);
 	else
 		JaguarWriteWord(RM, RN, DSP);
 }
@@ -805,7 +805,7 @@ static void dsp_opcode_store(void)
 static void dsp_opcode_loadb(void)
 {
 	if (RM >= DSP_WORK_RAM_BASE && RM <= (DSP_WORK_RAM_BASE + 0x1FFF))
-		RN = DSPReadLong(RM, DSP) & 0xFF;
+		RN = DSPReadLong(RM & 0xFFFFFFFC, DSP);
 	else
 		RN = JaguarReadByte(RM, DSP);
 }
@@ -813,7 +813,7 @@ static void dsp_opcode_loadb(void)
 static void dsp_opcode_loadw(void)
 {
 	if (RM >= DSP_WORK_RAM_BASE && RM <= (DSP_WORK_RAM_BASE + 0x1FFF))
-		RN = DSPReadLong(RM & 0xFFFFFFFE, DSP) & 0xFFFF;
+		RN = DSPReadLong(RM & 0xFFFFFFFC, DSP);
 	else
 		RN = JaguarReadWord(RM & 0xFFFFFFFE, DSP);
 }
diff --git a/waterbox/virtualjaguar/src/gpu.cpp b/waterbox/virtualjaguar/src/gpu.cpp
index 17d4fe389d..7cf108cc12 100644
--- a/waterbox/virtualjaguar/src/gpu.cpp
+++ b/waterbox/virtualjaguar/src/gpu.cpp
@@ -216,9 +216,9 @@ uint32_t gpu_convert_zero[32] =
 uint8_t * branch_condition_table = 0;
 #define BRANCH_CONDITION(x)	branch_condition_table[(x) + ((jaguar_flags & 7) << 5)]
 
-uint32_t GPUGetPC(void)
+bool GPURunning(void)
 {
-	return gpu_pc;
+	return GPU_RUNNING;
 }
 
 void build_branch_condition_table(void)
@@ -609,11 +609,6 @@ void GPUReset(void)
 		*((uint32_t *)(&gpu_ram_8[i])) = rand();
 }
 
-uint32_t GPUReadPC(void)
-{
-	return gpu_pc;
-}
-
 void GPUDone(void)
 {
 }
@@ -865,7 +860,7 @@ static void gpu_opcode_pack(void)
 static void gpu_opcode_storeb(void)
 {
 	if ((RM >= 0xF03000) && (RM <= 0xF03FFF))
-		GPUWriteLong(RM, RN & 0xFF, GPU);
+		GPUWriteLong(RM & 0xFFFFFFFC, RN, GPU);
 	else
 		JaguarWriteByte(RM, RN, GPU);
 }
@@ -873,7 +868,7 @@ static void gpu_opcode_storeb(void)
 static void gpu_opcode_storew(void)
 {
 	if ((RM >= 0xF03000) && (RM <= 0xF03FFF))
-		GPUWriteLong(RM & 0xFFFFFFFE, RN & 0xFFFF, GPU);
+		GPUWriteLong(RM & 0xFFFFFFFC, RN, GPU);
 	else
 		JaguarWriteWord(RM, RN, GPU);
 }
@@ -890,8 +885,7 @@ static void gpu_opcode_storep(void)
 {
 	if ((RM >= 0xF03000) && (RM <= 0xF03FFF))
 	{
-		GPUWriteLong((RM & 0xFFFFFFF8) + 0, gpu_hidata, GPU);
-		GPUWriteLong((RM & 0xFFFFFFF8) + 4, RN, GPU);
+		GPUWriteLong(RM & 0xFFFFFFFC, RN, GPU);
 	}
 	else
 	{
@@ -903,7 +897,7 @@ static void gpu_opcode_storep(void)
 static void gpu_opcode_loadb(void)
 {
 	if ((RM >= 0xF03000) && (RM <= 0xF03FFF))
-		RN = GPUReadLong(RM, GPU) & 0xFF;
+		RN = GPUReadLong(RM & 0xFFFFFFFC, GPU);
 	else
 		RN = JaguarReadByte(RM, GPU);
 }
@@ -911,7 +905,7 @@ static void gpu_opcode_loadb(void)
 static void gpu_opcode_loadw(void)
 {
 	if ((RM >= 0xF03000) && (RM <= 0xF03FFF))
-		RN = GPUReadLong(RM & 0xFFFFFFFE, GPU) & 0xFFFF;
+		RN = GPUReadLong(RM & 0xFFFFFFFC, GPU);
 	else
 		RN = JaguarReadWord(RM, GPU);
 }
@@ -925,8 +919,7 @@ static void gpu_opcode_loadp(void)
 {
 	if ((RM >= 0xF03000) && (RM <= 0xF03FFF))
 	{
-		gpu_hidata = GPUReadLong((RM & 0xFFFFFFF8) + 0, GPU);
-		RN		   = GPUReadLong((RM & 0xFFFFFFF8) + 4, GPU);
+		RN		   = GPUReadLong(RM & 0xFFFFFFFC, GPU);
 	}
 	else
 	{
diff --git a/waterbox/virtualjaguar/src/gpu.h b/waterbox/virtualjaguar/src/gpu.h
index 2d24715ffb..5c125805fe 100644
--- a/waterbox/virtualjaguar/src/gpu.h
+++ b/waterbox/virtualjaguar/src/gpu.h
@@ -25,8 +25,7 @@ void GPUWriteByte(uint32_t offset, uint8_t data, uint32_t who = UNKNOWN);
 void GPUWriteWord(uint32_t offset, uint16_t data, uint32_t who = UNKNOWN);
 void GPUWriteLong(uint32_t offset, uint32_t data, uint32_t who = UNKNOWN);
 
-uint32_t GPUGetPC(void);
-uint32_t GPUReadPC(void);
+bool GPURunning(void);
 
 // GPU interrupt numbers (from $F00100, bits 4-8)