diff --git a/examples/cifar10_dist_npu_eval/README.md b/examples/cifar10_dist_npu_eval/README.md
new file mode 100644
index 00000000..56c7da1e
--- /dev/null
+++ b/examples/cifar10_dist_npu_eval/README.md
@@ -0,0 +1,13 @@
+# CIFAR-10 Evaluation Example
+
+## Single process evaluation
+
+```bash
+python cifar10_npu_eval.py
+```
+
+## Multiple processes evaluation with torch.distributed
+
+```bash
+python cifar10_eval_torch_npu_dist.py
+```
diff --git a/examples/cifar10_dist_npu_eval/cifar10_eval_torch_npu_dist.py b/examples/cifar10_dist_npu_eval/cifar10_eval_torch_npu_dist.py
new file mode 100644
index 00000000..d38233d5
--- /dev/null
+++ b/examples/cifar10_dist_npu_eval/cifar10_eval_torch_npu_dist.py
@@ -0,0 +1,63 @@
+import os
+import torch
+import torchvision as tv
+import tqdm
+from torch.utils.data import DataLoader, DistributedSampler
+
+from mmeval import Accuracy
+
+
+def get_eval_dataloader(rank=0, num_replicas=1):
+    dataset = tv.datasets.CIFAR10(
+        root='./',
+        train=False,
+        download=True,
+        transform=tv.transforms.ToTensor())
+    dist_sampler = DistributedSampler(
+        dataset, num_replicas=num_replicas, rank=rank)
+    data_loader = DataLoader(dataset, batch_size=1, sampler=dist_sampler)
+    return data_loader, len(dataset)
+
+
+def get_model(pretrained_model_fpath=None):
+    model = tv.models.resnet18(num_classes=10)
+    if pretrained_model_fpath is not None:
+        model.load_state_dict(torch.load(pretrained_model_fpath))
+    return model.eval()
+
+
+def eval_fn(rank, process_num):
+    master_addr = 'localhost'
+    master_port = 12345
+
+    os.environ['MASTER_ADDR'] = master_addr
+    os.environ['MASTER_PORT'] = str(master_port)
+
+    torch.distributed.init_process_group(
+        backend='hccl',
+        init_method='env://',
+        world_size=process_num,
+        rank=rank)
+
+    num_npus = torch.npu.device_count()
+    torch.npu.set_device(rank % num_npus)
+
+    eval_dataloader, total_num_samples = get_eval_dataloader(rank, process_num)
+    model = get_model().npu()
+    accuracy = Accuracy(topk=(1, 3), dist_backend='npu_dist')
+
+    with torch.no_grad():
+        for images, labels in tqdm.tqdm(eval_dataloader, disable=(rank != 0)):
+            images = images.npu()
+            labels = labels.npu()
+            predicted_score = model(images)
+            accuracy.add(predictions=predicted_score, labels=labels)
+
+    print(accuracy.compute(size=total_num_samples))
+    accuracy.reset()
+
+
+if __name__ == '__main__':
+    process_num = 8
+    torch.multiprocessing.spawn(
+        eval_fn, nprocs=process_num, args=(process_num, ))
diff --git a/examples/cifar10_dist_npu_eval/cifar10_npu_eval.py b/examples/cifar10_dist_npu_eval/cifar10_npu_eval.py
new file mode 100644
index 00000000..4d244cdb
--- /dev/null
+++ b/examples/cifar10_dist_npu_eval/cifar10_npu_eval.py
@@ -0,0 +1,37 @@
+import torch
+import torchvision as tv
+import tqdm
+from torch.utils.data import DataLoader
+
+from mmeval import Accuracy
+
+
+def get_eval_dataloader():
+    dataset = tv.datasets.CIFAR10(
+        root='./',
+        train=False,
+        download=True,
+        transform=tv.transforms.ToTensor())
+    return DataLoader(dataset, batch_size=1)
+
+
+def get_model(pretrained_model_fpath=None):
+    model = tv.models.resnet18(num_classes=10)
+    if pretrained_model_fpath is not None:
+        model.load_state_dict(torch.load(pretrained_model_fpath))
+    return model.eval()
+
+
+eval_dataloader = get_eval_dataloader()
+model = get_model().npu()
+accuracy = Accuracy(topk=(1, 3))
+
+with torch.no_grad():
+    for images, labels in tqdm.tqdm(eval_dataloader):
+        images = images.npu()
+        labels = labels.npu()
+        predicted_score = model(images)
+        accuracy.add(predictions=predicted_score, labels=labels)
+
+print(accuracy.compute())
+accuracy.reset()
diff --git a/mmeval/core/dist.py b/mmeval/core/dist.py
index 2a3e2f6d..1d501aab 100644
--- a/mmeval/core/dist.py
+++ b/mmeval/core/dist.py
@@ -2,9 +2,9 @@
 
 from typing import List, Optional, no_type_check
 
-from .dist_backends import (BaseDistBackend, MPI4PyDist, NonDist, OneFlowDist,
-                            PaddleDist, TFHorovodDist, TorchCPUDist,
-                            TorchCUDADist)
+from .dist_backends import (BaseDistBackend, MPI4PyDist, NonDist, NPUDist,
+                            OneFlowDist, PaddleDist, TFHorovodDist,
+                            TorchCPUDist, TorchCUDADist)
 
 _DIST_BACKENDS = {
     'non_dist': NonDist,
@@ -14,6 +14,7 @@
     'torch_cpu': TorchCPUDist,
     'torch_cuda': TorchCUDADist,
     'paddle_dist': PaddleDist,
+    'npu_dist': NPUDist
 }
 
 _DEFAULT_BACKEND = 'non_dist'
diff --git a/mmeval/core/dist_backends/__init__.py b/mmeval/core/dist_backends/__init__.py
index d35b7130..02d7e878 100644
--- a/mmeval/core/dist_backends/__init__.py
+++ b/mmeval/core/dist_backends/__init__.py
@@ -3,6 +3,7 @@
 from .base_backend import BaseDistBackend, TensorBaseDistBackend
 from .mpi4py import MPI4PyDist
 from .non_dist import NonDist
+from .npu_dist import NPUDist
 from .oneflow_dist import OneFlowDist
 from .paddle_dist import PaddleDist
 from .tf_horovod import TFHorovodDist
@@ -12,5 +13,5 @@
 __all__ = [
     'BaseDistBackend', 'TensorBaseDistBackend', 'MPI4PyDist', 'NonDist',
     'OneFlowDist', 'TFHorovodDist', 'TorchCPUDist', 'TorchCUDADist',
-    'PaddleDist'
+    'PaddleDist', 'NPUDist'
 ]
diff --git a/mmeval/core/dist_backends/npu_dist.py b/mmeval/core/dist_backends/npu_dist.py
new file mode 100644
index 00000000..0c4086a4
--- /dev/null
+++ b/mmeval/core/dist_backends/npu_dist.py
@@ -0,0 +1,59 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+
+from typing import TYPE_CHECKING, Any, Tuple, TypeVar, Union
+
+from mmeval.utils import try_import
+from .torch_cpu import TorchCPUDist
+
+if TYPE_CHECKING:
+    import torch
+    import torch_npu
+else:
+    torch = try_import('torch')
+    torch_npu = try_import('torch_npu')
+
+Tensor = TypeVar('Tensor', bound='torch.Tensor')
+
+
+class NPUDist(TorchCPUDist):
+    """A distributed communication backend for Ascend NPU."""
+
+    def __init__(self) -> None:
+        super().__init__()
+        if torch_npu is None:
+            raise ImportError(f'For availability of {self.__class__.__name__},'
+                              ' please install ascend pytorch first.')
+        if not torch.distributed.is_hccl_available():
+            raise RuntimeError(
+                f'For availability of {self.__class__.__name__},'
+                ' make sure torch.distributed.is_hccl_available().')
+
+    def _object_to_tensor(self, obj: Any) -> Tuple[Tensor, Tensor]:
+        """Convert the given object to a npu tensor via `pickle.dumps`.
+
+        Args:
+            obj (any): Any pickle-able python object.
+
+        Returns:
+            tuple: A tuple of the tensor converted from given object and the
+            tensor size.
+        """
+        # Add type annotation make mypy happy
+        obj_tensor: Tensor
+        obj_size_tensor: Tensor
+        obj_tensor, obj_size_tensor = super()._object_to_tensor(obj)
+        return obj_tensor.npu(), obj_size_tensor.npu()
+
+    def _tensor_to_object(self, tensor: Tensor,
+                          tensor_size: Union[int, Tensor]) -> Any:
+        """Convert the given npu tensor to a object via `pickle.loads`.
+
+        Args:
+            tenosr (Tensor): A npu tensor.
+            tensor_size (int or Tensor): The tensor size of the given Tensor to
+            be convert object.
+
+        Returns:
+            Any: The object converted from the given npu tensor.
+        """
+        return super()._tensor_to_object(tensor.detach().cpu(), tensor_size)
diff --git a/mmeval/metrics/accuracy.py b/mmeval/metrics/accuracy.py
index 48c1149b..8ab0674e 100644
--- a/mmeval/metrics/accuracy.py
+++ b/mmeval/metrics/accuracy.py
@@ -25,6 +25,8 @@
     jax = try_import('jax')
     flow = try_import('oneflow')
 
+torch = try_import('torch')
+
 
 @overload
 @dispatch
diff --git a/mmeval/metrics/average_precision.py b/mmeval/metrics/average_precision.py
index 62efcc0d..b09918a1 100644
--- a/mmeval/metrics/average_precision.py
+++ b/mmeval/metrics/average_precision.py
@@ -16,6 +16,8 @@
     torch = try_import('torch')
     flow = try_import('oneflow')
 
+torch_npu = try_import('torch_npu')
+
 NUMPY_IMPL_HINTS = Tuple[Union[np.ndarray, np.number], Union[np.ndarray,
                                                              np.number]]
 TORCH_IMPL_HINTS = Tuple['torch.Tensor', 'torch.Tensor']
diff --git a/mmeval/metrics/end_point_error.py b/mmeval/metrics/end_point_error.py
index 07fcb77c..928609b6 100644
--- a/mmeval/metrics/end_point_error.py
+++ b/mmeval/metrics/end_point_error.py
@@ -15,6 +15,8 @@
     torch = try_import('torch')
     flow = try_import('oneflow')
 
+torch_npu = try_import('torch_npu')
+
 
 class EndPointError(BaseMetric):
     """EndPointError evaluation metric.
diff --git a/mmeval/metrics/f1_score.py b/mmeval/metrics/f1_score.py
index 75401591..9718eb3f 100644
--- a/mmeval/metrics/f1_score.py
+++ b/mmeval/metrics/f1_score.py
@@ -14,6 +14,8 @@
     torch = try_import('torch')
     flow = try_import('oneflow')
 
+torch_npu = try_import('torch_npu')
+
 
 class F1Score(BaseMetric):
     """Compute F1 scores.
diff --git a/mmeval/metrics/mean_iou.py b/mmeval/metrics/mean_iou.py
index be17b987..c431cc17 100644
--- a/mmeval/metrics/mean_iou.py
+++ b/mmeval/metrics/mean_iou.py
@@ -24,6 +24,8 @@
     tf = try_import('tensorflow')
     flow = try_import('oneflow')
 
+torch_npu = try_import('torch_npu')
+
 
 class MeanIoU(BaseMetric):
     """MeanIoU evaluation metric.
diff --git a/mmeval/metrics/perplexity.py b/mmeval/metrics/perplexity.py
index 00d2eadc..1443ffba 100644
--- a/mmeval/metrics/perplexity.py
+++ b/mmeval/metrics/perplexity.py
@@ -21,6 +21,8 @@
     tf = try_import('tensorflow')
     flow = try_import('oneflow')
 
+torch_npu = try_import('torch_npu')
+
 
 def softmax(x: np.ndarray) -> np.ndarray:
     """Compute the softmax function.
diff --git a/mmeval/metrics/precision_recall_f1score.py b/mmeval/metrics/precision_recall_f1score.py
index 3ca661b7..c08df603 100644
--- a/mmeval/metrics/precision_recall_f1score.py
+++ b/mmeval/metrics/precision_recall_f1score.py
@@ -21,6 +21,8 @@
     flow = try_import('oneflow')
     of_F = try_import('oneflow.nn.functional')
 
+torch_npu = try_import('torch_npu')
+
 NUMPY_IMPL_HINTS = Tuple[Union[np.ndarray, np.number], Union[np.ndarray,
                                                              np.number]]
 TORCH_IMPL_HINTS = Tuple['torch.Tensor', 'torch.Tensor']
diff --git a/tests/test_core/test_dist_backends/test_npu_dist_single_node.py b/tests/test_core/test_dist_backends/test_npu_dist_single_node.py
new file mode 100644
index 00000000..45ed92de
--- /dev/null
+++ b/tests/test_core/test_dist_backends/test_npu_dist_single_node.py
@@ -0,0 +1,143 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+
+import os
+import pytest
+
+from mmeval.core.dist_backends.npu_dist import NPUDist
+
+# check if current process is launch via mpirun
+if os.environ.get('OMPI_COMM_WORLD_SIZE', '0') != '0':
+    pytest.skip(allow_module_level=True)
+
+torch_npu = pytest.importorskip('torch_npu')
+
+torch = pytest.importorskip('torch')
+torch_dist = pytest.importorskip('torch.distributed')
+mp = pytest.importorskip('torch.multiprocessing')
+
+
+def _init_torch_dist(rank, world_size, comm_backend, port):
+    os.environ['MASTER_ADDR'] = 'localhost'
+    os.environ['MASTER_PORT'] = str(port)
+
+    torch_dist.init_process_group(
+        backend=comm_backend,
+        init_method='env://',
+        world_size=world_size,
+        rank=rank)
+
+    if comm_backend == 'hccl':
+        num_gpus = torch.npu.device_count()
+        torch.npu.set_device(rank % num_gpus)
+
+
+def _create_obj_list(world_size):
+    obj_list = []
+    for idx in range(world_size):
+        obj = dict()
+        obj['rank'] = idx
+        obj['world_size'] = world_size
+        obj['data'] = [i for i in range(idx)]
+        obj_list.append(obj)
+    return obj_list
+
+
+def _torch_dist_all_gather_fn(rank, world_size, comm_backend, port):
+    _init_torch_dist(rank, world_size, comm_backend, port)
+    dist_comm = NPUDist()
+
+    assert dist_comm.is_initialized
+    assert dist_comm.rank == rank
+    assert dist_comm.world_size == world_size
+
+    obj_list = _create_obj_list(world_size)
+    local_obj = obj_list[rank]
+    print(f'rank {rank}, local_obj {local_obj}')
+
+    gather_obj_list = dist_comm.all_gather_object(local_obj)
+    print(f'rank {rank}, gather_obj_list {gather_obj_list}')
+
+    assert gather_obj_list == obj_list
+
+
+def _torch_dist_broadcast_fn(rank, world_size, comm_backend, port):
+    _init_torch_dist(rank, world_size, comm_backend, port)
+    dist_comm = NPUDist()
+
+    assert dist_comm.is_initialized
+    assert dist_comm.rank == rank
+    assert dist_comm.world_size == world_size
+
+    rank_0_obj = {'rank': 0}
+
+    if rank == 0:
+        obj = rank_0_obj
+    else:
+        obj = None
+
+    print(f'rank {rank}, obj {obj}')
+    broadcast_obj = dist_comm.broadcast_object(obj, src=0)
+    print(f'rank {rank}, broadcast_obj {broadcast_obj}')
+
+    assert broadcast_obj == rank_0_obj
+
+
+@pytest.mark.skipif(
+    not torch_dist.is_hccl_available(),
+    reason='HCCL backend is not available.')
+@pytest.mark.skipif(
+    torch.npu.device_count() < 1,
+    reason='NPU device count must greater than 0.')
+@pytest.mark.parametrize(
+    argnames=['process_num', 'comm_port'],
+    argvalues=[
+        pytest.param(
+            1,
+            2347,
+            marks=pytest.mark.skipif(
+                torch.npu.device_count() < 1,
+                reason='npu device count must greater than 0.')),
+        pytest.param(
+            2,
+            2347,
+            marks=pytest.mark.skipif(
+                torch.npu.device_count() < 2,
+                reason='NPU device count must greater than 2.'))
+    ])
+def test_hccl_all_gather_object(process_num, comm_port):
+    comm_backend = 'hccl'
+    mp.spawn(
+        _torch_dist_all_gather_fn,
+        nprocs=process_num,
+        args=(process_num, comm_backend, comm_port))
+
+
+@pytest.mark.skipif(
+    not torch_dist.is_hccl_available(),
+    reason='HCCL backend is not available.')
+@pytest.mark.parametrize(
+    argnames=['process_num', 'comm_port'],
+    argvalues=[
+        pytest.param(
+            1,
+            2350,
+            marks=pytest.mark.skipif(
+                torch.npu.device_count() < 1,
+                reason='npu device count must greater than 0.')),
+        pytest.param(
+            2,
+            2350,
+            marks=pytest.mark.skipif(
+                torch.npu.device_count() < 2,
+                reason='NPU device count must greater than 2.'))
+    ])
+def test_hccl_broadcast_object(process_num, comm_port):
+    comm_backend = 'hccl'
+    mp.spawn(
+        _torch_dist_broadcast_fn,
+        nprocs=process_num,
+        args=(process_num, comm_backend, comm_port))
+
+
+if __name__ == '__main__':
+    pytest.main([__file__, '-v', '--capture=no'])