The Comparative Analysis of the Titan V and Tesla P40 on AI Algorithms
Many graphics cards, which are needed to solve the most demanding tasks in different areas, are released frequently. We have purchased a new graphics card – the Titan V and decided to compare it with the Tesla P40. Experiments on the performance of graphics cards were conducted. On the basis of the data obtained from experiments, we were able to find out which one is better. Let’s consider the characteristics of these video cards and the results of experiments.
Titan V and Tesla P40: Specifications
The NVIDIA Titan V according to the manufacturer, is the most powerful graphics card ever created for the PC, driven by the world’s most advanced architecture—NVIDIA Volta. This graphics card delivers new levels of performance. Built on the 12 nm process, and based on the GV100 graphics processor, in its GV100-400-A1 variant.
The NVIDIA Tesla P40 was designed to meet the challenges of the modern data center. This is a professional graphics card, built on the 16 nm process, and based on the GP102 graphics processor, in its GP102 variant.
Here is a comparison table.
General specs | ||
GPU | Titan V | Tesla P40 |
Architecture | NVIDIA Volta | NVIDIA Pascal™ |
GPU Name | GV100 | GP102 |
Type | Desktop | Workstation |
Release Date | December 7, 2017 | September 13, 2016 |
Technical specs | ||
CUDA Cores (Shading Units) | 5120 | 3840 |
GPU Clock Speed | 1200 MHz | 1303 MHz |
Boost Clock Speed | 1455 MHz | 1531 MHz |
Transistor Count | 21.1 Billion | 12 Billion |
Manufacturing Process Technology | 12 nm | 16 nm |
Thermal Design Power (TDP) | 250 Watts | 250 Watts |
Texture Rate (Bilinear) | 384 GigaTexels/sec | 367.4 GigaTexels/s |
Floating-point Performance | 14,899 GFLOPS | 11,758 GFLOPS |
Thermal Threshold | 91° C | 45° C |
Memory specs | ||
Memory Type | HBM2 | GDDR5X |
Total Video Memory | 12 GB | 24 GB |
Memory Interface | 3072-bit | 384-bit |
Memory Clock Speed | 850 MHz | 1251 MHz 10008 MHz effective |
Total Memory Bandwidth | 652.8 GB/s | 480.4 GB/s |
Board design | ||
Form Factor | Dual Slot | Dual Slot |
Length | 10.5 inches 267 mm |
10.5 inches 267 mm |
Outputs | 1x HDMI 3x DisplayPort |
No outputs |
Power Connectors | 1x 6-pin + 1x 8-pin | 1x 6-pin + 1x 8-pin |
Titan V and Tesla P40: Results of Experiments
Architecture | Experiment | Results | Note |
Convolutional Neural Network. This model uses three convolutional layers and two fully connected layers | 2000 iterations of training on the MNIST dataset | It took 7 seconds the Titan V and 120 seconds the Tesla P40 to complete the experiment | The fact that the Tesla P40 loaded the model approximately in a minute, whereas its counterpart did it almost instantly, influenced the results of the experiment |
GAN uses two fully connected layers in the generator/discriminator | 100,000 iterations of training on the MNIST dataset | It took 183 seconds the Titan V and 519 seconds the Tesla P40 to complete the experiment | The Titan V operated at its full capacity, and it loaded the information for processing faster than the Tesla P40 |
DCGAN uses approximately three convolutional layers in the generator and discriminator, that is why this model is more complex than GAN | 20,000 iterations of training on the MNIST dataset | It took 96 seconds the Titan V and 308 seconds the Tesla P40 to complete the experiment | The Titan V operated at its full capacity, and it loaded the information for processing faster than the Tesla P40 |
ResNet is one of the largest neural networks, which is often used to transfer the learning experience from one large dataset to another one, which is smaller and more specialized. ResNet uses thirty-two convolutional layers, plus several fully connected layers | 10,000 iterations of training on the cifar10 dataset | It took 151 seconds the Titan V and 420 seconds the Tesla P40 to complete the experiment | |
WideNet takes a huge input from statistics, not the image | 10,000 iterations of training on the tax statistics dataset | It took 12 seconds the Titan V and 57 seconds the Tesla P40 to complete the experiment |
Thus, considering all the above, we can conclude that the Titan V has better characteristics than the Tesla P40. So, the Titan V is much newer (7 December 2017 vs 13 September 2016), it has a wider memory bus (3072 vs 384 bit), more pipelines (5120 vs 3840), a higher memory bandwidth (652.8 vs 480.4 GB/s), and a finer manufacturing process technology (12 vs 16 nm). Based on the results of experiments, conducted on different models (Convolutional Neural Network, GAN, DCGAN, ResNet, WideNet), we can assert that the Titan V video card is more powerful than the Tesla P40.