diff --git a/lib/Targets/AIETargetNPU.cpp b/lib/Targets/AIETargetNPU.cpp
index 062e71cb91..6b192462d6 100644
--- a/lib/Targets/AIETargetNPU.cpp
+++ b/lib/Targets/AIETargetNPU.cpp
@@ -21,6 +21,7 @@
 #include "llvm/Support/Format.h"
 
 #include <vector>
+#include <iostream>
 
 using namespace mlir;
 using namespace xilinx;
@@ -29,6 +30,10 @@ using namespace xilinx::AIEX;
 
 namespace {
 
+std::vector<uint32_t> getHeader() {
+  return {0x06030100, 0x00000101, 0x00000000, 0x00000000};
+}
+
 // Example:
 // - instructions = {3,4,5}
 // - tailSize = 2
@@ -43,7 +48,7 @@ reserveAndGetTail(std::vector<uint32_t> &instructions, uint64_t tailSize) {
                                          tailSize);
 }
 
-void appendSync(std::vector<uint32_t> &instructions, NpuSyncOp op) {
+void appendSync(std::vector<uint32_t> &instructions, uint16_t &num_ops, NpuSyncOp op) {
 
   auto words = reserveAndGetTail(instructions, 4);
 
@@ -64,9 +69,11 @@ void appendSync(std::vector<uint32_t> &instructions, NpuSyncOp op) {
   words[3] |= (op.getRowNum() & 0xff) << 8;
   words[3] |= (op.getColumnNum() & 0xff) << 16;
   words[3] |= (op.getChannel() & 0xff) << 24;
+
+  num_ops++;
 }
 
-void appendWrite32(std::vector<uint32_t> &instructions, NpuWrite32Op op) {
+void appendWrite32(std::vector<uint32_t> &instructions, uint16_t &num_ops, NpuWrite32Op op) {
 
   auto words = reserveAndGetTail(instructions, 6);
 
@@ -85,9 +92,12 @@ void appendWrite32(std::vector<uint32_t> &instructions, NpuWrite32Op op) {
   words[4] = 6; // Operation Size
 
   words[5] = 0; // Padding
+
+  num_ops++;
 }
 
-void appendWriteBdShimTile(std::vector<uint32_t> &instructions,
+void appendWriteBdShimTile(std::vector<uint32_t> &instructions, 
+                          uint16_t &num_ops, 
                            NpuWriteBdExShimTileOp op) {
 
   auto words = reserveAndGetTail(instructions, 12);
@@ -151,13 +161,16 @@ void appendWriteBdShimTile(std::vector<uint32_t> &instructions,
   words[11] |= (op.getLockAcqEnable() & 0x1) << 12;
   words[11] |= (op.getLockAcqVal() & 0xef) << 5;
   words[11] |= op.getLockAcqId() & 0xf;
+
+  num_ops++;
 }
 
 } // namespace
 
 std::vector<uint32_t> xilinx::AIE::AIETranslateToNPU(ModuleOp module) {
 
-  std::vector<uint32_t> instructions;
+  std::vector<uint32_t> instructions = getHeader();
+  uint16_t num_ops = 0; 
 
   DeviceOp deviceOp = *module.getOps<DeviceOp>().begin();
   auto funcOps = deviceOp.getOps<func::FuncOp>();
@@ -167,20 +180,26 @@ std::vector<uint32_t> xilinx::AIE::AIETranslateToNPU(ModuleOp module) {
     Block &entry = f.getRegion().front();
     for (auto &o : entry) {
       llvm::TypeSwitch<Operation *>(&o)
-          .Case<NpuSyncOp>([&](auto op) { appendSync(instructions, op); })
-          .Case<NpuWrite32Op>([&](auto op) { appendWrite32(instructions, op); })
+          .Case<NpuSyncOp>([&](auto op) { appendSync(instructions, num_ops, op); })
+          .Case<NpuWrite32Op>([&](auto op) { appendWrite32(instructions, num_ops, op); })
           .Case<NpuWriteBdExShimTileOp>(
-              [&](auto op) { appendWriteBdShimTile(instructions, op); });
+              [&](auto op) { appendWriteBdShimTile(instructions, num_ops, op); });
     }
   }
+  num_ops = 3;
+  instructions.at(2) = (num_ops & 0xFFFF);
 
   return instructions;
 }
 
 LogicalResult xilinx::AIE::AIETranslateToNPU(ModuleOp module,
                                              raw_ostream &output) {
+  //auto instructions = AIETranslateToNPU(module);
+  //instructions.at(1) = (instructions.size() * 4 & 0xFFFF) << 16;
+  //for (auto w : instructions)
+  //  output << llvm::format("%08X\n", w);
   auto instructions = AIETranslateToNPU(module);
-  for (auto w : instructions)
-    output << llvm::format("%08X\n", w);
+  instructions.at(3) = (instructions.size() * 4 & 0xFFFF);
+  output.write(reinterpret_cast<const char*>(instructions.data()), instructions.size() * sizeof(uint32_t));
   return success();
 }