diff --git a/.github/automation/build_acl.sh b/.github/automation/build_acl.sh
index 0c45cc5a291..7ed588618ff 100755
--- a/.github/automation/build_acl.sh
+++ b/.github/automation/build_acl.sh
@@ -28,7 +28,7 @@ source ${SCRIPT_DIR}/common_aarch64.sh
 
 ACL_CONFIG=${ACL_CONFIG:-"Release"}
 ACL_ROOT_DIR=${ACL_ROOT_DIR:-"${PWD}/ComputeLibrary"}
-ACL_VERSION=${ACL_VERSION:-v24.09}
+ACL_VERSION=${ACL_VERSION:-v24.11.1}
 ACL_ARCH=${ACL_ARCH:-"armv8.2-a"}
 ACL_REPO="https://github.com/ARM-software/ComputeLibrary.git"
 
diff --git a/README.md b/README.md
index 6b5c384a069..7030985f013 100644
--- a/README.md
+++ b/README.md
@@ -173,7 +173,7 @@ On a CPU based on Arm AArch64 architecture, oneDNN CPU engine can be built with
 machine learning applications and provides AArch64 optimized implementations
 of core functions. This functionality currently requires that ACL is downloaded
 and built separately. See [Build from Source] section of the Developer Guide for
-details. oneDNN only supports Compute Library versions 24.08.1 or later.
+details. oneDNN only supports Compute Library versions 24.11.1 or later.
 
 [Arm Compute Library (ACL)]: https://github.com/arm-software/ComputeLibrary
 
diff --git a/cmake/ACL.cmake b/cmake/ACL.cmake
index b185f7ba340..3a08f779a39 100644
--- a/cmake/ACL.cmake
+++ b/cmake/ACL.cmake
@@ -31,7 +31,7 @@ endif()
 
 find_package(ACL REQUIRED)
 
-set(ACL_MINIMUM_VERSION "24.08.1")
+set(ACL_MINIMUM_VERSION "24.11.1")
 
 if(ACL_FOUND)
     file(GLOB_RECURSE ACL_VERSION_FILE ${ACL_INCLUDE_DIR}/*/arm_compute_version.embed)
diff --git a/src/cpu/aarch64/matmul/acl_matmul.cpp b/src/cpu/aarch64/matmul/acl_matmul.cpp
index 64d594c6a9d..0bb8c9a6a4b 100644
--- a/src/cpu/aarch64/matmul/acl_matmul.cpp
+++ b/src/cpu/aarch64/matmul/acl_matmul.cpp
@@ -15,6 +15,7 @@
 *******************************************************************************/
 
 #include "cpu/aarch64/matmul/acl_matmul.hpp"
+#include <mutex>
 
 namespace dnnl {
 namespace impl {
@@ -171,12 +172,20 @@ status_t acl_matmul_t::pd_t::init(engine_t *engine) {
 
 template <bool IsFixedFormat>
 status_t acl_matmul_t::execute_forward(const exec_ctx_t &ctx) const {
-
     status_t status = status::success;
     auto src_base = CTX_IN_MEM(const data_t *, DNNL_ARG_SRC);
     auto wei_base = CTX_IN_MEM(const data_t *, DNNL_ARG_WEIGHTS);
 
-    auto amp = pd()->amp_;
+    const auto &amp = pd()->amp_;
+
+    std::unique_lock<std::mutex> locker {mtx_, std::defer_lock};
+
+    // Some of the underlying kernels used by ACL still require some state and
+    // are not safe to be called in parallel with different execution contexts.
+    // Eventually when all kernels are truly stateless, this guard can be
+    // removed.
+    if (!acl_obj_->asm_gemm.has_stateless_impl()) { locker.lock(); }
+
     bool is_transA = amp.is_transA;
     bool is_transB = amp.is_transB;
     bool do_transC = amp.do_transC;
diff --git a/src/cpu/aarch64/matmul/acl_matmul.hpp b/src/cpu/aarch64/matmul/acl_matmul.hpp
index 30641a746a7..4cb84e873e7 100644
--- a/src/cpu/aarch64/matmul/acl_matmul.hpp
+++ b/src/cpu/aarch64/matmul/acl_matmul.hpp
@@ -17,6 +17,7 @@
 #ifndef ACL_MATMUL_HPP
 #define ACL_MATMUL_HPP
 
+#include <mutex>
 #include "common/utils.hpp"
 #include "cpu/aarch64/acl_post_ops.hpp"
 #include "cpu/aarch64/matmul/acl_matmul_utils.hpp"
@@ -71,6 +72,7 @@ struct acl_matmul_t : public primitive_t {
     const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); }
 
     std::unique_ptr<acl_matmul_obj_t> acl_obj_;
+    mutable std::mutex mtx_;
 }; // acl_matmul_t
 
 } // namespace matmul