diff --git a/CHANGELOG.md b/CHANGELOG.md
index 81fbefca31453..fee8f94f226b9 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -317,6 +317,10 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 - Fixed a bug where an infinite recursion would be triggered when using the `BaseFinetuning` callback on a model that contains a `ModuleDict` ([#8170](https://github.com/PyTorchLightning/pytorch-lightning/pull/8170))
 
+
+- Fixed `log_gpu_memory` metrics not being added to `logging` when nothing else is logged ([#8174](https://github.com/PyTorchLightning/pytorch-lightning/pull/8174))
+
+
 ## [1.3.7] - 2021-06-22
 
 - Fixed a bug where skipping an optimizer while using amp causes amp to trigger an assertion error ([#7975](https://github.com/PyTorchLightning/pytorch-lightning/pull/7975))
diff --git a/pytorch_lightning/trainer/connectors/logger_connector/logger_connector.py b/pytorch_lightning/trainer/connectors/logger_connector/logger_connector.py
index 6aae89273f4fc..5e75d51fc33a6 100644
--- a/pytorch_lightning/trainer/connectors/logger_connector/logger_connector.py
+++ b/pytorch_lightning/trainer/connectors/logger_connector/logger_connector.py
@@ -38,6 +38,7 @@ def __init__(self, trainer: 'pl.Trainer', log_gpu_memory: Optional[str] = None)
         self._progress_bar_metrics: Dict[str, float] = {}
         self._logged_metrics: Dict[str, _METRIC] = {}
         self._callback_metrics: Dict[str, _METRIC] = {}
+        self._gpus_metrics: Dict[str, str] = {}
         self._epoch_end_reached = False
         self._current_fx: Optional[str] = None
         self._batch_idx: Optional[int] = None
@@ -94,11 +95,6 @@ def log_metrics(self, metrics: Dict[str, _METRIC], step: Optional[int] = None) -
         if self.trainer.logger is None or not metrics:
             return
 
-        # add gpu memory
-        if self.trainer._device_type == DeviceType.GPU and self.log_gpu_memory:
-            mem_map = memory.get_memory_profile(self.log_gpu_memory)
-            metrics.update(mem_map)
-
         # turn all tensors to scalars
         scalar_metrics = metrics_to_scalars(metrics)
 
@@ -213,6 +209,8 @@ def update_train_step_metrics(self) -> None:
         if self.trainer.fit_loop.should_accumulate() and self.trainer.lightning_module.automatic_optimization:
             return
 
+        self._log_gpus_metrics()
+
         # when metrics should be logged
         assert not self._epoch_end_reached
         if self.should_update_logs or self.trainer.fast_dev_run:
@@ -226,6 +224,12 @@ def update_train_epoch_metrics(self) -> None:
         # reset result collection for next epoch
         self.trainer._results.reset(metrics=True)
 
+    def _log_gpus_metrics(self):
+        for key, mem in self.gpus_metrics.items():
+            gpu_id = int(key.split('/')[0].split(':')[1])
+            if gpu_id in self.trainer.accelerator_connector.parallel_device_ids:
+                self.trainer.lightning_module.log(key, mem, prog_bar=False, logger=True, on_step=True, on_epoch=False)
+
     """
     Utilities and properties
     """
@@ -276,6 +280,13 @@ def metrics(self) -> Dict[MetricSource, Dict[str, _METRIC]]:
         on_step = not self._epoch_end_reached
         return self.trainer._results.metrics(on_step)
 
+    @property
+    def gpus_metrics(self) -> Dict[str, str]:
+        if self.trainer._device_type == DeviceType.GPU and self.log_gpu_memory:
+            mem_map = memory.get_memory_profile(self.log_gpu_memory)
+            self._gpus_metrics.update(mem_map)
+        return self._gpus_metrics
+
     @property
     def callback_metrics(self) -> Dict[str, _METRIC]:
         if self.trainer._results:
diff --git a/tests/trainer/logging_/test_train_loop_logging.py b/tests/trainer/logging_/test_train_loop_logging.py
index 8ca51b2dee3ef..61e7a2a37ebc6 100644
--- a/tests/trainer/logging_/test_train_loop_logging.py
+++ b/tests/trainer/logging_/test_train_loop_logging.py
@@ -712,3 +712,21 @@ def training_step(self, *args):
     model = TestModel()
     with pytest.raises(MisconfigurationException, match=r'reduce_fx={min,max,mean,sum}\)` are currently supported'):
         trainer.fit(model)
+
+
+@RunIf(min_gpus=2)
+def test_log_gpu_memory_without_logging_on_step(tmpdir):
+
+    model = BoringModel()
+    trainer = Trainer(
+        default_root_dir=tmpdir,
+        max_epochs=1,
+        limit_train_batches=1,
+        limit_val_batches=0,
+        log_gpu_memory='all',
+        log_every_n_steps=1,
+        gpus=[1]
+    )
+    trainer.fit(model)
+
+    assert 'gpu_id: 1/memory.used (MB)' in trainer.logged_metrics