diff --git a/easybuild/easyblocks/t/tensorflow.py b/easybuild/easyblocks/t/tensorflow.py
index 8e40621918d..2c62f5b5d71 100644
--- a/easybuild/easyblocks/t/tensorflow.py
+++ b/easybuild/easyblocks/t/tensorflow.py
@@ -28,6 +28,7 @@
 @author: Kenneth Hoste (HPC-UGent)
 @author: Ake Sandgren (Umea University)
 @author: Damian Alvarez (Forschungzentrum Juelich GmbH)
+@author: Alexander Grund (TU Dresden)
 """
 import glob
 import os
@@ -213,13 +214,13 @@ class EB_TensorFlow(PythonPackage):
 
     @staticmethod
     def extra_options():
-        # We only want to install mkl-dnn by default on x86_64 systems
-        with_mkl_dnn_default = get_cpu_architecture() == X86_64
         extra_vars = {
             'path_filter': [[], "List of patterns to be filtered out in paths in $CPATH and $LIBRARY_PATH", CUSTOM],
             'with_jemalloc': [None, "Make TensorFlow use jemalloc (usually enabled by default). " +
                                     "Unsupported starting at TensorFlow 1.12!", CUSTOM],
-            'with_mkl_dnn': [with_mkl_dnn_default, "Make TensorFlow use Intel MKL-DNN", CUSTOM],
+            'with_mkl_dnn': [None, "Make TensorFlow use Intel MKL-DNN / oneDNN and configure with --config=mkl "
+                                   "(enabled by default where supported for TensorFlow versions before 2.4.0)",
+                             CUSTOM],
             'with_xla': [None, "Enable XLA JIT compiler for possible runtime optimization of models", CUSTOM],
             'test_script': [None, "Script to test TensorFlow installation with", CUSTOM],
             'test_targets': [[], "List of Bazel targets which should be run during the test step", CUSTOM],
@@ -802,13 +803,33 @@ def build_step(self):
             if self._with_cuda:
                 self.target_opts.append('--config=cuda')
 
-        # if mkl-dnn is listed as a dependency it is used. Otherwise downloaded if with_mkl_dnn is true
+        # note: using --config=mkl results in a significantly different build, with a different
+        # threading model (which may lead to thread oversubscription and significant performance loss,
+        # see https://github.com/easybuilders/easybuild-easyblocks/issues/2577) and different
+        # runtime behavior w.r.t. GPU vs CPU execution of functions like tf.matmul
+        # (see https://github.com/easybuilders/easybuild-easyconfigs/issues/14120),
+        # so make sure you really know you want to use this!
+
+        # auto-enable use of MKL-DNN/oneDNN and --config=mkl when possible if with_mkl_dnn is left unspecified;
+        # only do this for TensorFlow versions older than 2.4.0, since more recent versions
+        # oneDNN is used automatically for x86_64 systems (and mkl-dnn is no longer a dependency);
+        if self.cfg['with_mkl_dnn'] is None and LooseVersion(self.version) < LooseVersion('2.4.0'):
+            cpu_arch = get_cpu_architecture()
+            if cpu_arch == X86_64:
+                # Supported on x86 since forever
+                self.cfg['with_mkl_dnn'] = True
+                self.log.info("Auto-enabled use of MKL-DNN on %s CPU architecture", cpu_arch)
+            else:
+                self.log.info("Not enabling use of MKL-DNN on %s CPU architecture", cpu_arch)
+
+        # if mkl-dnn is listed as a dependency it is used
         mkl_root = get_software_root('mkl-dnn')
         if mkl_root:
             self.target_opts.append('--config=mkl')
             env.setvar('TF_MKL_ROOT', mkl_root)
         elif self.cfg['with_mkl_dnn']:
-            # this makes TensorFlow use mkl-dnn (cfr. https://github.com/01org/mkl-dnn)
+            # this makes TensorFlow use mkl-dnn (cfr. https://github.com/01org/mkl-dnn),
+            # and download it if needed
             self.target_opts.append('--config=mkl')
 
         # Compose final command