Why is Tesla P100 12 GB ~1/10 as fast as Apple M1?

I am finetuning the pretrained AlexNet to classify images of galaxies. I have a dataset of about 14,000 images, and on my base M1 MacBook Air I can train and test over 10 epochs in roughly 35-38 minutes.

I am trying to repeat this procedure remotely on my university computing resources, where I can access a Tesla P100 12GB GPU. I expect this to be much faster than my MacBook Air. However, when I train remotely, it seems to be approximately 1/10 as fast as my MacBook Air. (I don’t have a precise number because I’ve only asked the scheduler for “jobs” that are 2 hours long, and I get booted off before training finishes.) I’m pretty sure this indicates a problem somewhere.

I am trying to figure out whether there is a problem in my PyTorch code, or whether I’m somehow incorrectly using my university resources. Therefore I am sharing my code below in the hopes that someone can tell me whether it has any errors.

I exported all the code from my Jupyter Notebook (run remotely) as Markdown, and pasted it below. A lot of it’s not really relevant, but I thought I’d include all of it. You can see where I confirm that PyTorch sees the GPU, and I also use the .to(device) method to move the model to the GPU.

Note that my code is a sort of hodgepodge from Sebastian Raschka’s 2022 PyTorch book, and the PyTorch tutorials.

To be clear, the code seems to work when I run it remotely, but it’s just agonizingly slow.

Thanks for any help.

My code:

import torch
import torchvision.transforms as transforms
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import os
import pathlib
from PIL import Image
imgdir_path = pathlib.Path('/home/jsa378/NA10/14K/50kpc')
file_list = sorted([str(path) for path in
'Tesla P100-PCIE-12GB'
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
namorph = pd.read_table("/home/jsa378/NA10/NairAbrahamMorphology.cat")
0        3.0
1       -5.0
2       -2.0
3        4.0
4        5.0
14029    2.0
14030    2.0
14031    1.0
14032    0.0
14033    3.0
Name: TType, Length: 14034, dtype: float64
labels = torch.zeros(len(namorph.TType), dtype=torch.int64)
for i in range(len(namorph.TType)):
    if namorph.TType[i] == -5.0:
        labels[i] = 0 # Elliptical
    elif -3 <= namorph.TType[i] and namorph.TType[i] <= 0:
        labels[i] = 1 # Lenticular
    elif 1 <= namorph.TType[i] and namorph.TType[i] <= 9:
        labels[i] = 2 # Spiral
    elif 10 == namorph.TType[i] or 99 == namorph.TType[i]:
        labels[i] = 3 # Irr+Misc
tensor([2, 0, 1,  ..., 2, 1, 2])
class ImageDataset():#(Dataset):
    def __init__(self, file_list, labels):
        self.file_list = file_list
        self.labels = labels
    def __getitem__(self, index):
        file = self.file_list[index]
        label = self.labels[index]
        return file, label
    def __len__(self):
        return len(self.labels)
image_dataset = ImageDataset(file_list, labels)
img_height, img_width = 256, 256
size = [224, 224]

transform = transforms.Compose([
    transforms.Resize((img_height, img_width)),
class ImageDataset():#(Dataset):
    def __init__(self, file_list, labels, transform=None):
        self.file_list = file_list
        self.labels = labels
        self.transform = transform
    def __getitem__(self, index):
        img = Image.open(self.file_list[index])
        if self.transform is not None:
            img = self.transform(img)
        label = self.labels[index]
        return img, label
    def __len__(self):
        return len(self.labels)
image_dataset = ImageDataset(file_list, labels, transform)
# import ssl
# ssl._create_default_https_context = ssl._create_unverified_context
# model = torch.hub.load('pytorch/vision:v0.10.0', 'alexnet', pretrained=True)
# model.to(device)
# model.eval()
# if torch.cuda.is_available():
#     input_batch = input_batch.to('cuda')
#     model.to('cuda')
from __future__ import print_function
from __future__ import division
import torch.nn as nn
import torch.optim as optim
import torchvision
from torchvision import datasets, models, transforms
import time
import copy
print("PyTorch Version: ",torch.__version__)
print("Torchvision Version: ",torchvision.__version__)
PyTorch Version:  1.10.0
Torchvision Version:  0.11.1
data_dir = imgdir_path
model_name = "alexnet"
num_classes = 4
batch_size = 8
num_epochs = 10
feature_extract = False
def train_model(model, dataloaders, criterion, optimizer, num_epochs=25, is_inception=False):
    since = time.time()
    val_acc_history = []
    best_model_wts = copy.deepcopy(model.state_dict())
    best_acc = 0.0
    for epoch in range(num_epochs):
        print('Epoch {}/{}'.format(epoch, num_epochs - 1))
        print('-' * 10)
        # Each epoch has a training and validation phase
        for phase in ['train', 'val']:
            if phase == 'train':
                model.train() # Set model to training mode
                model.eval() # Set model to evaluate mode
            running_loss = 0.0
            running_corrects = 0
            # Iterate over data.
            for inputs, labels in dataloaders[phase]:
                inputs = inputs.to(device)
                labels = labels.to(device)
                # zero the parameter gradients
                # forward
                # track history if only in train
                with torch.set_grad_enabled(phase == 'train'):
                    # Get model outputs and calculate loss
                    # Special case for inception because in training it has an
                    # auxiliary output. In train mode we calculate the loss by
                    # summing the final output and the auxiliary output but in
                    # testing we only consider the final output.
                    if is_inception and phase == 'train':
                        # From https://discuss.pytorch.org/t/how-to-optimize-inception-model-with-auxiliary-classifiers/7958
                        outputs, aux_outputs = model(inputs)
                        loss1 = criterion(outputs, labels)
                        loss2 = criterion(aux_outputs, labels)
                        loss = loss1 + 0.4*loss2
                        outputs = model(inputs)
                        loss = criterion(outputs, labels)
                    _, preds = torch.max(outputs, 1)
                    # backward + optimize only if in training phase
                    if phase == 'train':
                # statistics
                running_loss += loss.item() * inputs.size(0)
                running_corrects += torch.sum(preds == labels.data)
            epoch_loss = running_loss / len(dataloaders[phase].dataset)
            epoch_acc = running_corrects.double() / len(dataloaders[phase].dataset)
            print('{} Loss: {:.4f} Acc: {:.4f}'.format(phase, epoch_loss, epoch_acc))
            # deep copy the model
            if phase == 'val' and epoch_acc > best_acc:
                best_acc = epoch_acc
                best_model_wts = copy.deepcopy(model.state_dict())
            if phase == 'val':
    time_elapsed = time.time() - since
    print('Training complete in {:.0f}m {:.0f}s'.format(time_elapsed // 60, time_elapsed % 60))
    print('Best val Acc: {:4f}'.format(best_acc))
    # load best model weights
    return model, val_acc_history
def set_parameter_requires_grad(model, feature_extracting):
    if feature_extracting:
        for param in model.parameters():
            param.requires_grad = False
# model.classifier[6] = nn.Linear(4096,num_classes)
# model.eval()
def initialize_model(model_name, num_classes, feature_extract, use_pretrained=True):
    # Initialize these variables which will be set in this if statement.
    # Each of these variables is model specific.
    model_ft = None
    input_size = 0
    if model_name == "alexnet":
        """ Alexnet
        model_ft = models.alexnet(pretrained=use_pretrained)
        set_parameter_requires_grad(model_ft, feature_extract)
        num_ftrs = model_ft.classifier[6].in_features
        model_ft.classifier[6] = nn.Linear(num_ftrs,num_classes)
        input_size = 224
        print("Invalid model name, exiting...")
    return model_ft, input_size
model_ft, input_size = initialize_model(model_name, num_classes, feature_extract=False, use_pretrained=True)
  (features): Sequential(
    (0): Conv2d(3, 64, kernel_size=(11, 11), stride=(4, 4), padding=(2, 2))
    (1): ReLU(inplace=True)
    (2): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
    (3): Conv2d(64, 192, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
    (4): ReLU(inplace=True)
    (5): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
    (6): Conv2d(192, 384, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (7): ReLU(inplace=True)
    (8): Conv2d(384, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (9): ReLU(inplace=True)
    (10): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (11): ReLU(inplace=True)
    (12): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
  (avgpool): AdaptiveAvgPool2d(output_size=(6, 6))
  (classifier): Sequential(
    (0): Dropout(p=0.5, inplace=False)
    (1): Linear(in_features=9216, out_features=4096, bias=True)
    (2): ReLU(inplace=True)
    (3): Dropout(p=0.5, inplace=False)
    (4): Linear(in_features=4096, out_features=4096, bias=True)
    (5): ReLU(inplace=True)
    (6): Linear(in_features=4096, out_features=4, bias=True)
model_ft = model_ft.to(device)
params_to_update = model_ft.parameters()
print("Params to learn:")
if feature_extract:
    params_to_update = []
    for name,param in model_ft.named_parameters():
        if param.requires_grad == True:
    for name,param in model_ft.named_parameters():
        if param.requires_grad == True:

optimizer_ft = optim.SGD(params_to_update, lr=0.001, momentum=0.9)
Params to learn:
train_size = 12000
test_size = len(image_dataset) - train_size
from torch.utils.data import Subset
train_dataset = Subset(image_dataset, torch.arange(train_size))
test_dataset = Subset(image_dataset, torch.arange(train_size, len(image_dataset)))
print(len(train_dataset), len(test_dataset))
12000 2034
from torch.utils.data import DataLoader
train_dataloader = DataLoader(train_dataset, batch_size=64, shuffle=True)
test_dataloader = DataLoader(test_dataset, batch_size=64, shuffle=True)
dataloaders_dict = {
    "train": train_dataloader,
    # "test": test_dataloader # I think the code I'm using uses 'val' instead of 'test',
    # so I'm renaming this to 'val' below
    "val": test_dataloader
criterion = nn.CrossEntropyLoss()
model_ft, hist = train_model(model_ft, dataloaders_dict, criterion, optimizer_ft, num_epochs=num_epochs, is_inception=(model_name=="inception"))
Epoch 0/9
train Loss: 0.7081 Acc: 0.7117
val Loss: 0.5314 Acc: 0.7940

Epoch 1/9

1 Like

It’s hard to tell where the bottleneck in your code is and if the GPU executed the training slow or if it’s stale due to another bottleneck e.g. in the data loading pipeline.
E.g. if your university cluster uses a mounted network drive to load the data, I would expect to see a lot of idle time for the GPU as you won’t be able to feed the samples fast enough to it.
You could either profile the code using e.g. Nsight systems and check where the bottleneck is and/or also execute a pure GPU workload by removing parts of the code such as the data loading.


Thanks for your reply. I am investigating the possibility that my error is in how I’m using my university cluster. I may need to add something to my job script copying the data set to faster storage, closer to the GPU.

Thanks for your help. If I remember, I’ll update this thread once I’ve (hopefully) straightened this out.


After altering my job script to extract the images to “node-local storage”, training over 10 epochs using the Tesla P100 12GB took 17 minutes, which is about twice a fast as the MacBook Air M1. This seems much more reasonable.

Thanks for your help.

1 Like