Energy consumption comparison of dense networks in GPU

Energy consumption comparison in machine learning platforms

By Rodrigo Ingelmo, Artelnics. 05 September 2022.

TensorFlow and Neural Designer are popular machine learning platforms developed by Google and Artelnics, respectively.

Although all that frameworks are based on neural networks, they present important differences in functionality, usability, performance, consumption etc.

This post compares the energy consumption of TensorFlow and Neural Designer using the GPU for an approximation benchmark.

As we will see, Neural Designer consumes 42% less than its competitor machine learning platform.

In this article, we provide all the steps that you need to reproduce the result using the free trial of Neural Designer.

Contents:

Introduction

Two of the most important features in machine learning platforms are their training speed and the total amount of energy consumed during this process.

In most cases, modelling huge data sets is very expensive in computational terms which leads to a high economic cost of the neural network training and to a high environmental impact.

Hence, here in Artelnics, we have been recently working hard on this issue, achieving incredible unexpected astonishing results.

Thus, the aim of this article is to measure the GPU energy consumption of TensorFlow and Neural Designer for a benchmark application. Also, a couple of instructions are given so as to enable anyone to repeat this one or a similar benchmark and check by their own the amazing results obtained when Neural Designer is used.

The following table summarizes the technical features of these tools that might impact their GPU performance.

The above table shows that TensorFlow is programmed in C++ and Python, whereas Neural Designer is entirely programmed in C++.

Interpreted languages like Python have some advantages over compiled languages like C ++, such as their ease of use.

However, the performance of Python is, in general, lower than that of C++. Indeed, Python takes significant time to interpret sentences during the program's execution.

On the other hand, TensorFlow use automatic differentiation, while Neural Designer uses analytical differentiation.

As before, automatic differentiation has some advantages over analytical differentiation. Indeed, it simplifies obtaining the gradient for new architectures or loss indices.

However, the performance of automatic differentiation is, in general, lower than that of analytical differentiation: The first derives the gradient during the program's execution, while the second has that formula pre-calculated.

Next, we measure the energy consumption for a benchmark problem on a reference computer using TensorFlow and Neural Designer. The results produced by these platforms are then compared.

Benchmark application

The first step is to choose a benchmark application that is general enough to draw conclusions about the performance of the machine learning platforms. As previously stated, we will train a neural network that approximates a set of input-target samples.

In this regard, an approximation application is defined by a data set, a neural network and an associated training strategy. The next table uniquely defines these three components.

Data set	Benchmark: Rosenbrock Inputs number: 1000 Targets number: 1 Samples number: 1000000 File size: 22 Gb (download)
Neural network	Layers number: 2 Layer 1: -Type: Perceptron (Dense) -Inputs number: 1000 -Neurons number: 1000 -Activation function: Hyperbolic tangent (tanh) Layer 2: -Type: Perceptron (Dense) -Inputs number: 1000 -Neurons number: 1 -Activation function: Linear Initialization: Random uniform [-1,1]
Training strategy	Loss index: -Error: Mean Squared Error (MSE) -Regularization: None Optimization algorithm: -Algorithm: Adaptive Moment Estimation (Adam) -Batch size: 1000 -Maximum epochs: 20000

Once the TensorFlow and Neural Designer applications have been created, we need to run them.

Reference computer

The next step is to choose the computer in which the neural network will be trained with TensorFlow and Neural Designer. The table below shows the features of the computer used for this instance.

Once the computer has been selected, we install TensorFlow (2.1.0) and Neural Designer (5.9.9) on it.

Below, the TensorFlow code used is shown.

import tensorflow as tf
import pandas as pd
import time
import numpy as np
from tensorflow.keras.utils import Sequence

#read data float32


filename = "C:/Users/Usuario/Downloads/rosenbrock.csv"
df_test = pd.read_csv(filename, nrows=100)
float_cols = [c for c in df_test if df_test[c].dtype == "float64"]
float32_cols = {c: np.float32 for c in float_cols}
data = pd.read_csv(filename, engine='c', dtype=float32_cols)

    
x = data.iloc[:,:-1].values
y = data.iloc[:,[-1]].values

initializer = tf.keras.initializers.RandomUniform(minval=-1., maxval=1.)


#build model

model = tf.keras.models.Sequential([tf.keras.layers.Dense(1000,activation = 'tanh', kernel_initializer = initializer, bias_initializer=initializer),
tf.keras.layers.Dense(1, activation = 'linear', kernel_initializer = initializer, bias_initializer=initializer)])	   	


#compile model

model.compile(optimizer='adam', loss = 'mean_squared_error')

		
#train model 

class DataGenerator(Sequence):
                def __init__(self, x_set, y_set, batch_size):
                self.x, self.y = x_set, y_set
                self.batch_size = batch_size

                def __len__(self):
                return int(np.ceil(len(self.x) / float(self.batch_size)))

                def __getitem__(self, idx):
        batch_x = self.x[idx * self.batch_size:(idx + 1) * self.batch_size]
        batch_y = self.y[idx * self.batch_size:(idx + 1) * self.batch_size]
                return batch_x, batch_y

train_gen = DataGenerator(x, y, 1000)

start_time = time.time()
with tf.device('/gpu:0'):
    history = model.fit(train_gen, epochs=20000)
print("Training time: ", round(time.time() - start_time), " seconds")

Reference electricity consumption meter

In this section, it is described the device used for the energy consumption measurements so as to the reader will be able to reproduce the results obtained in the following section with maximum accuracy.

Results

The last step is to run the benchmark application on the selected machine with TensorFlow and Neural Designer and to compare the energy consumed by those platforms during the training processes.

The next figure shows the training time with TensorFlow.

As we can see, TensorFlow takes 30:14:30 to train the neural network for 20000 epochs (5.44 seconds/epoch). The final mean squared error is 0.0003. And the overall energy consumption of the training process is 4.5 kWh, as shown below.

Finally, the following figure shows the training time with Neural Designer.

Neural Designer takes 21:03:43 to train the neural network for 20000 epochs (3.79 seconds/epoch). During that time, it reaches a mean squared error of 0.023. And the overall energy consumption of the training process is 2.6 kWh, as shown below.

The following table summarizes the the most important metrics that the two machine learning platforms yielded .

Now, the following chart depicts the enery consumed using TensorFlow and Neural Designer graphically in this case.

As we can see, the energy consumption of Neural Designer for this application is 42 % lower than that using TensorFlow.

Conclusions

Neural Designer is entirely written in C ++, uses analytical differentiation, and has been optimized to minimize the number of operations during training.

As a result, its energy consumption during the training process using Neural Designer is 42 % lower than that using TensorFlow.

To reproduce these results, download the free trial of Neural Designer and follow the steps described in this article.

Do you need help? Contact us | FAQ | Privacy | © 2022, Artificial Intelligence Techniques, Ltd.