diff --git a/include/NeuraDialect/NeuraOps.td b/include/NeuraDialect/NeuraOps.td
index 06866449..d15bafc0 100644
--- a/include/NeuraDialect/NeuraOps.td
+++ b/include/NeuraDialect/NeuraOps.td
@@ -31,6 +31,16 @@ def Neura_FAddOp : Op<NeuraDialect, "fadd"> {
   let traits = [SameOperandsAndResultElementType];
 }
 
+// Defines a floating-point substraction operation.
+def Neura_FSubOp: Op<NeuraDialect, "fsub"> {
+  let summary = "Floating substraction operation";
+  let opName = "fsub";
+  let arguments = (ins AnyFloat:$lhs, AnyFloat:$rhs);
+  let results = (outs AnyFloat:$result);
+  // let assemblyFormat = "$lhs `,` $rhs attr-dict `:` type($result)";
+  let traits = [SameOperandsAndResultElementType];
+}
+
 // Defines a floating-point multiplication operation.
 def Neura_FMulOp : Op<NeuraDialect, "fmul"> {
   let summary = "Floating multiplication operation";
diff --git a/lib/Conversion/LlvmToNeura/LlvmToNeuraPatterns.td b/lib/Conversion/LlvmToNeura/LlvmToNeuraPatterns.td
index 7daf5b08..6004cf96 100644
--- a/lib/Conversion/LlvmToNeura/LlvmToNeuraPatterns.td
+++ b/lib/Conversion/LlvmToNeura/LlvmToNeuraPatterns.td
@@ -3,11 +3,17 @@ include "mlir/IR/PatternBase.td"
 include "mlir/Dialect/LLVMIR/LLVMOps.td"
 include "NeuraDialect/NeuraOps.td"
 
+// Floating point binary operations.
 def : Pat<
   (LLVM_FAddOp $lhs, $rhs, $_fastmath),
   (Neura_FAddOp $lhs, $rhs)
 >;
 
+def : Pat<
+  (LLVM_FSubOp $lhs, $rhs, $_fastmath),
+  (Neura_FSubOp $lhs, $rhs)
+>;
+
 def : Pat<
   (LLVM_ConstantOp $value),
   (Neura_ConstantOp $value)
diff --git a/test/neura/interpreter/interpreter.mlir b/test/neura/interpreter/interpreter.mlir
index b9efeb87..c5e163e7 100644
--- a/test/neura/interpreter/interpreter.mlir
+++ b/test/neura/interpreter/interpreter.mlir
@@ -10,4 +10,14 @@ module {
     return %2 : f32
     // CHECK: 11.0
   }
+
+  func.func @test_sub() -> f32 {
+    %arg0 = arith.constant 9.0 : f32
+    %cst = arith.constant 2.0 : f32
+    %0 = "neura.mov"(%arg0) : (f32) -> f32
+    %1 = "neura.mov"(%cst) : (f32) -> f32
+    %2 = "neura.fsub"(%0, %1) : (f32, f32) -> f32
+    return %2 : f32
+    // CHECK: 7.0
+  }
 }
diff --git a/test/neura/interpreter/lower_and_interpret_subf.mlir b/test/neura/interpreter/lower_and_interpret_subf.mlir
new file mode 100644
index 00000000..47475f70
--- /dev/null
+++ b/test/neura/interpreter/lower_and_interpret_subf.mlir
@@ -0,0 +1,41 @@
+// RUN: mlir-opt %s \
+// RUN:   --convert-scf-to-cf \
+// RUN:   --convert-math-to-llvm \
+// RUN:   --convert-arith-to-llvm \
+// RUN:   --convert-func-to-llvm \
+// RUN:   --convert-cf-to-llvm \
+// RUN:   --reconcile-unrealized-casts \
+// RUN:   -o %t-lowered-to-llvm.mlir
+
+// RUN: mlir-translate -mlir-to-llvmir \
+// RUN:   %t-lowered-to-llvm.mlir \
+// RUN:   -o %t-lower_and_interpreter.ll
+
+// RUN: llc %t-lower_and_interpreter.ll \
+// RUN:   -filetype=obj -o %t-out.o
+
+// RUN: clang++ main.cpp %t-out.o \
+// RUN:   -o %t-out.bin
+
+// RUN: %t-out.bin > %t-dumped_output.txt
+
+// RUN: mlir-neura-opt --lower-arith-to-neura --insert-mov %s \
+// RUN:   -o %t-neura.mlir
+
+// RUN: neura-interpreter %t-neura.mlir >> %t-dumped_output.txt
+// RUN: FileCheck %s < %t-dumped_output.txt
+
+// RUN: FileCheck %s -check-prefix=GOLDEN < %t-dumped_output.txt
+// GOLDEN: 7.0
+
+module {
+  func.func @test() -> f32 attributes { llvm.emit_c_interface }{
+    %arg0 = arith.constant 9.0 : f32
+    %cst = arith.constant 2.0 : f32
+    %0 = arith.subf %arg0, %cst : f32
+    // CHECK: Golden output: [[OUTPUT:[0-9]+\.[0-9]+]]
+    // CHECK: [neura-interpreter] Output: [[OUTPUT]]
+    return %0 : f32
+  }
+}
+
diff --git a/test/neura/llvm_sub.mlir b/test/neura/llvm_sub.mlir
new file mode 100644
index 00000000..0003cc34
--- /dev/null
+++ b/test/neura/llvm_sub.mlir
@@ -0,0 +1,10 @@
+// RUN: mlir-neura-opt --assign-accelerator --lower-llvm-to-neura --insert-mov %s | FileCheck %s
+
+func.func @test(%a: f32) -> f32 {
+  %b = llvm.mlir.constant(2.0 : f32) : f32
+  %res = llvm.fsub %a, %b : f32
+  // CHECK: [[LHS:%.*]] = "neura.mov"(%{{.*}}) : (f32) -> f32
+  // CHECK: [[RHS:%.*]] = "neura.mov"(%{{.*}}) : (f32) -> f32
+  // CHECK: [[RES:%.*]] = "neura.fsub"([[LHS]], [[RHS]])
+  return %res : f32
+}
\ No newline at end of file
diff --git a/tools/neura-interpreter/neura-interpreter.cpp b/tools/neura-interpreter/neura-interpreter.cpp
index 972e442c..c12b58d0 100644
--- a/tools/neura-interpreter/neura-interpreter.cpp
+++ b/tools/neura-interpreter/neura-interpreter.cpp
@@ -73,6 +73,10 @@ int main(int argc, char **argv) {
         float lhs = valueMap[faddOp.getLhs()];
         float rhs = valueMap[faddOp.getRhs()];
         valueMap[faddOp.getResult()] = lhs + rhs;
+      } else if (auto fsubOp = dyn_cast<neura::FSubOp>(op)) {
+        float lhs = valueMap[fsubOp.getLhs()];
+        float rhs = valueMap[fsubOp.getRhs()];
+        valueMap[fsubOp.getResult()] = lhs - rhs;
       } else if (auto retOp = dyn_cast<func::ReturnOp>(op)) {
         float result = valueMap[retOp.getOperand(0)];
         llvm::outs() << "[neura-interpreter] Output: " << llvm::format("%.6f", result) << "\n";