diff --git a/README.md b/README.md
index e9c7d492..ff721a95 100644
--- a/README.md
+++ b/README.md
@@ -63,9 +63,11 @@ pip install -U torch torchaudio --index-url https://download.pytorch.org/whl/cu1
 pip install -U so-vits-svc-fork
 ```
 
-- If you are using an AMD GPU on Linux, replace `--index-url https://download.pytorch.org/whl/cu117` with `--index-url https://download.pytorch.org/whl/rocm5.4.2`.
 - If no GPU is available, simply remove `pip install -U torch torchaudio --index-url https://download.pytorch.org/whl/cu117`.
-- If `fairseq` raises an error that [`Microsoft C++ Build Tools`](https://visualstudio.microsoft.com/visual-cpp-build-tools/) is not installed or that some dll is missing, please (re)install it.
+- If you are using an AMD GPU on Linux, replace `--index-url https://download.pytorch.org/whl/cu117` with `--index-url https://download.pytorch.org/whl/rocm5.4.2`. AMD GPUs are not supported on Windows (#120).
+- If `fairseq` raises an error:
+  - If it prompts [`Microsoft C++ Build Tools`](https://visualstudio.microsoft.com/visual-cpp-build-tools/) is not installed. please install it.
+  - If it prompts that some dll is missing, reinstalling `Microsoft Visual C++ 2022` and `Windows SDK` may help.
 
 ### Update
 
@@ -132,14 +134,14 @@ Place your dataset like `dataset_raw/{speaker_id}/**/{wav_file}.{any_format}` (s
 svc pre-resample
 svc pre-config
 svc pre-hubert
-svc train
+svc train -t
 ```
 
 #### Notes
 
 - Dataset audio duration per file should be <~ 10s or VRAM will run out.
 - To change the f0 inference method to CREPE, replace `svc pre-hubert` with `svc pre-hubert -fm crepe`. You may need to reduce `--n-jobs` due to performance issues.
-- It is recommended to change the batch_size in `config.json` before the `train` command to match the VRAM capacity. As tested, the default requires about 14 GB.
+- It is recommended to change the batch_size in `config.json` before the `train` command to match the VRAM capacity. The default value is optimized for Tesla T4 (16GB VRAM), but training is possible without that much VRAM.
 - Silence removal and volume normalization are automatically performed (as in the upstream repo) and are not required.
 
 ### Further help
diff --git a/src/so_vits_svc_fork/__main__.py b/src/so_vits_svc_fork/__main__.py
index 7f41ef44..e0335996 100644
--- a/src/so_vits_svc_fork/__main__.py
+++ b/src/so_vits_svc_fork/__main__.py
@@ -86,13 +86,32 @@ def cli():
     help="path to output dir",
     default=Path("./logs/44k"),
 )
-def train(config_path: Path, model_path: Path):
+@click.option(
+    "-t/-nt",
+    "--tensorboard/--no-tensorboard",
+    default=False,
+    type=bool,
+    help="launch tensorboard",
+)
+def train(config_path: Path, model_path: Path, tensorboard: bool = False):
     """Train model
     If D_0.pth or G_0.pth not found, automatically download from hub."""
     from .train import train
 
     config_path = Path(config_path)
     model_path = Path(model_path)
+
+    if tensorboard:
+        import webbrowser
+
+        from tensorboard import program
+
+        getLogger("tensorboard").setLevel(30)
+        tb = program.TensorBoard()
+        tb.configure(argv=[None, "--logdir", model_path.as_posix()])
+        url = tb.launch()
+        webbrowser.open(url)
+
     train(config_path=config_path, model_path=model_path)