diff --git a/SYCL/Matrix/Legacy/element_wise_all_ops_half.cpp b/SYCL/Matrix/Legacy/element_wise_all_ops_half.cpp
index bd8becff4c..2cb93a42a8 100644
--- a/SYCL/Matrix/Legacy/element_wise_all_ops_half.cpp
+++ b/SYCL/Matrix/Legacy/element_wise_all_ops_half.cpp
@@ -5,7 +5,7 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
-// REQUIRES: matrix
+// REQUIRES: matrix,gpu
 
 // Only runs on DPAS because AMX implementation does not support half data type
 // yet
diff --git a/SYCL/Matrix/Legacy/joint_matrix_half.cpp b/SYCL/Matrix/Legacy/joint_matrix_half.cpp
index d88a9f0b1b..d9a1c19e18 100644
--- a/SYCL/Matrix/Legacy/joint_matrix_half.cpp
+++ b/SYCL/Matrix/Legacy/joint_matrix_half.cpp
@@ -5,7 +5,7 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
-// REQUIRES: matrix
+// REQUIRES: matrix,gpu
 
 // RUN: %clangxx -fsycl %s -o %t.out -DSYCL_EXT_ONEAPI_MATRIX_VERSION=1
 // Only run on the GPU because half is not supported on AMX hardware
diff --git a/SYCL/Matrix/XMX8/element_wise_all_ops_int8_packed.cpp b/SYCL/Matrix/XMX8/element_wise_all_ops_int8_packed.cpp
index 7124332923..696dfbf364 100644
--- a/SYCL/Matrix/XMX8/element_wise_all_ops_int8_packed.cpp
+++ b/SYCL/Matrix/XMX8/element_wise_all_ops_int8_packed.cpp
@@ -11,12 +11,7 @@
 // RUN: %CPU_RUN_PLACEHOLDER %t.out
 // RUN: %GPU_RUN_PLACEHOLDER %t.out
 
-// This test stores the matrix B that is VNNIed (packed) in a row major fashion.
-// This is expected to fail on the GPU because the implementation does not
-// support automatic transformation YET, in this case: VNNI to row major in the
-// store.
-
-// XFAIL: gpu
+// This test stores the matrix B that is VNNIed (packed).
 
 #include <iostream>
 #include <random>
diff --git a/SYCL/Matrix/element_wise_all_ops_half.cpp b/SYCL/Matrix/element_wise_all_ops_half.cpp
index e860180c03..e4b7c95a2b 100644
--- a/SYCL/Matrix/element_wise_all_ops_half.cpp
+++ b/SYCL/Matrix/element_wise_all_ops_half.cpp
@@ -5,7 +5,7 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
-// REQUIRES: matrix
+// REQUIRES: matrix,gpu
 
 // Only runs on DPAS because AMX implementation does not support half data type
 // yet
diff --git a/SYCL/Matrix/element_wise_all_ops_int8_packed.cpp b/SYCL/Matrix/element_wise_all_ops_int8_packed.cpp
index 6008079449..e7a0664688 100644
--- a/SYCL/Matrix/element_wise_all_ops_int8_packed.cpp
+++ b/SYCL/Matrix/element_wise_all_ops_int8_packed.cpp
@@ -11,10 +11,8 @@
 // RUN: %CPU_RUN_PLACEHOLDER %t.out
 // RUN: %GPU_RUN_PLACEHOLDER %t.out
 
-// This test stores the matrix B that is VNNIed (packed) in a row major fashion.
-// This is expected to fail on the GPU because the implementation does not
-// support automatic transformation YET, in this case: VNNI to row major in the
-// store.
+// This test stores the matrix B that is VNNIed (packed).
+// This is expected to fail on the GPU because some built-ins are missing still.
 
 // XFAIL: gpu
 
diff --git a/SYCL/Matrix/joint_matrix_half.cpp b/SYCL/Matrix/joint_matrix_half.cpp
index 1d131a64a8..be0bf1487c 100644
--- a/SYCL/Matrix/joint_matrix_half.cpp
+++ b/SYCL/Matrix/joint_matrix_half.cpp
@@ -5,7 +5,7 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
-// REQUIRES: matrix
+// REQUIRES: matrix,gpu
 
 // RUN: %clangxx -fsycl %s -o %t.out -DSYCL_EXT_ONEAPI_MATRIX_VERSION=4
 // Only run on the GPU because half is not supported on AMX hardware