diff --git a/.github/workflows/aipu-build-and-test.yml b/.github/workflows/aipu-build-and-test.yml
index 9fd20ee20..1d0314fbe 100644
--- a/.github/workflows/aipu-build-and-test.yml
+++ b/.github/workflows/aipu-build-and-test.yml
@@ -52,7 +52,7 @@ jobs:
           source ~/env_setup.sh
           export FLAGTREE_BACKEND=aipu
           cd python
-          MAX_JOBS=64 python3.10 -m pip install . --no-build-isolation -v
+          MAX_JOBS=16 python3.10 -m pip install . --no-build-isolation -v
 
       - name: FlagTree Test on AIPU
         shell: bash
diff --git a/include/CMakeLists.txt b/include/CMakeLists.txt
index 109c292fe..72181b98f 100644
--- a/include/CMakeLists.txt
+++ b/include/CMakeLists.txt
@@ -1 +1 @@
-add_subdirectory(triton)
+add_subdirectory(triton)
\ No newline at end of file
diff --git a/include/flagtree/Common/UnifiedHardwareBase.h b/include/flagtree/Common/UnifiedHardwareBase.h
new file mode 100644
index 000000000..03e8aba38
--- /dev/null
+++ b/include/flagtree/Common/UnifiedHardwareBase.h
@@ -0,0 +1,27 @@
+#ifndef UNIFIED_HARDWARE_BASE_H
+#define UNIFIED_HARDWARE_BASE_H
+
+#include <optional> 
+
+namespace mlir {
+namespace flagtree {
+//this is the unified hardware abstraction for hardware
+//to determined if these abstraction is specified, using std::optional is needed
+//using in passes: if(uh_flagtree->xxx()){...}
+
+class UnifiedHardware{
+
+public:
+    virtual ~UnifiedHardware() = default;
+
+    //DMA
+    virtual std::optional<int> getAllocSpaceForDMATag() const {
+        return std::nullopt;
+    }
+
+};
+
+} // namespace flagtree
+} // namespace mlir
+
+#endif // UNIFIED_HARDWARE_BASE_H
diff --git a/include/triton/Dialect/Triton/IR/TritonAttrDefs.td b/include/triton/Dialect/Triton/IR/TritonAttrDefs.td
index 571d2b55b..898ca1559 100644
--- a/include/triton/Dialect/Triton/IR/TritonAttrDefs.td
+++ b/include/triton/Dialect/Triton/IR/TritonAttrDefs.td
@@ -4,6 +4,7 @@
 include "mlir/IR/EnumAttr.td"
 include "mlir/IR/AttrTypeBase.td"
 
+
 // Attributes for LoadOp and StoreOp
 def TT_CacheModifierAttr : I32EnumAttr<
     "CacheModifier", "",
diff --git a/python/src/ir.cc b/python/src/ir.cc
index ee35ce834..c1f092ed7 100644
--- a/python/src/ir.cc
+++ b/python/src/ir.cc
@@ -674,6 +674,13 @@ void init_triton_ir(py::module &&m) {
         values));
   });
 
+  m.def("make_attr_i64", [](const std::vector<int64_t> &values, MLIRContext &context) {
+    return mlir::cast<Attribute>(DenseIntElementsAttr::get(
+        RankedTensorType::get({static_cast<int64_t>(values.size())},
+                              IntegerType::get(&context, 64)),
+        values));
+  });
+
   m.def(
       "parse_mlir_module",
       [](const std::string &inputFilename, MLIRContext &context) {
diff --git a/third_party/aipu/backend/compiler.py b/third_party/aipu/backend/compiler.py
index 5a2c16b84..5596d21a4 100644
--- a/third_party/aipu/backend/compiler.py
+++ b/third_party/aipu/backend/compiler.py
@@ -7,6 +7,9 @@
 from mlir.passmanager import PassManager
 from mlir.ir import Context, Module
 
+import ctypes
+from triton._C.libtriton.aipu import UnifiedHardwareAIPU
+
 from dataclasses import dataclass
 import functools
 import hashlib
@@ -96,6 +99,12 @@ def make_ttir(mod, metadata, opt):
 
     @staticmethod
     def make_linalg(mod, metadata, opt):
+        uh_aipu = UnifiedHardwareAIPU()
+        uh_aipu_ptr = ctypes.cast(id(uh_aipu), ctypes.c_void_p)
+        ptr_value = int(uh_aipu_ptr.value)
+        ptr_attr = ir.make_attr_i64([ptr_value], mod.context)
+        #mod.set_attr("uh_aipu_ptr", ptr_attr)  # TODO
+
         pm = ir.pass_manager(mod.context)
         pm.enable_debug()
         # Add pass here.
diff --git a/third_party/aipu/triton_aipu.cc b/third_party/aipu/triton_aipu.cc
index b617232df..47014e011 100644
--- a/third_party/aipu/triton_aipu.cc
+++ b/third_party/aipu/triton_aipu.cc
@@ -22,6 +22,8 @@
 #include "triton-shared/Conversion/TritonToLinalg/TritonToLinalg.h"
 #include "triton-shared/Conversion/TritonToLinalgExperimental/TritonToLinalgExperimental.h"
 
+#include "unified_hardware_aipu.h"
+
 #include <pybind11/functional.h>
 #include <pybind11/pybind11.h>
 #include <pybind11/stl.h>
@@ -69,6 +71,7 @@ void init_triton_aipu(py::module &&m) {
   init_triton_aipu_common(m.def_submodule("common"));
   auto passes = m.def_submodule("passes");
   init_triton_aipu_passes_convert(passes.def_submodule("convert"));
+
   // load dialects
   m.def("load_dialects", [](mlir::MLIRContext &context) {
     using namespace mlir;
@@ -88,4 +91,6 @@ void init_triton_aipu(py::module &&m) {
     context.loadAllAvailableDialects();
   });
   // register passes here
+  py::class_<mlir::aipu::UnifiedHardwareAIPU>(m, "UnifiedHardwareAIPU")
+        .def(py::init<>());    // flagtree
 }
diff --git a/third_party/aipu/unified_hardware_aipu.h b/third_party/aipu/unified_hardware_aipu.h
new file mode 100644
index 000000000..f8f272ecc
--- /dev/null
+++ b/third_party/aipu/unified_hardware_aipu.h
@@ -0,0 +1,22 @@
+#ifndef UNIFIED_HARDWARE_AIPU_H
+#define UNIFIED_HARDWARE_AIPU_H
+
+#include <optional> 
+
+#include "flagtree/Common/UnifiedHardwareBase.h"
+
+namespace mlir {
+namespace aipu {
+
+class UnifiedHardwareAIPU final : public mlir::flagtree::UnifiedHardware {
+    
+    //DMA
+    std::optional<int> getAllocSpaceForDMATag() const override{
+        return std::optional<int>(11);
+    }
+};
+
+} // namespace aipu
+} // namespace mlir
+
+#endif // UNIFIED_HARDWARE_AIPU_H
\ No newline at end of file