Usually, located at /usr/local/cuda/bin

## Non-Visual Profiler

```
$ nvprof python train_mnist.py
```

I prefer to use --print-gpu-trace.

```
$ nvprof --print-gpu-trace python train_mnist.py
```

## Visual Profiler

On GPU machine, run

```
$ nvprof -o prof.nvvp python train_mnist.py
```

Copy `prof.nvvp` into your local machine

```
$ scp your_gpu_machine:/path/to/prof.nvvp .
```

Then, run nvvp (nvidia visual profiler) on your local machine:

```
$ nvvp prof.nvvp
```

It works more comfortably than X11 forwarding or something.