Posted on 18 February, 2019
The NVIDIA® DGX-2™ arrived at vScaler's internal lab and they couldn't wait to test it and integrate their cloud system with it!
Paving the way for modern AI, the NVIDIA DGX-2 is recognised as 'the World's most powerful Deep Learning system' with unprecedented levels of compute, the platform is targeted at deep learning computing boasting the processing power of its equally magnificent predecessor, the NVIDIA® DGX-1™. As AI is getting increasingly complex, the DGX-2 packs 16 of the world's most advanced GPUs and now model complexity and size are no longer constrained by the limits of traditional architecture.
System Specs: NVIDIA DGX-2
- 16 x fully connected NVIDIA® Tesla® V100 32GB GPUs
- 2 high-speed Ethernet networking cards
- 1.5TB system memory
- 30TB NVMe SSDs
- Latest Generation CPUs for faster, more resilient boot and storage management
vScaler
vScaler is a private Cloud platform built on Open Source technology that enables the creation of a secure, scalable, cost-effective, flexible IT infrastructure. Find out more about vScaler here.
The integration of vScaler was seamless thanks to a preconfigured image that the team had been using for their DeepOps integration flashing the system with that. This enabled all the tools needed to access the NVIDIA GPU container bank with Kubernetes and similar optimisation options.
TensorFlow Benchmarking for ResNet Models
In order to assess the performance of the DGX-2, vScaler used the ResNet Model which is a popular industry benchmarking platform for assessing training and inference performance. vScaler selected two common models; ResNet-50 and ResNet-152, which are 50 and 152 layer Residual Network models respectively. To guarantee each GPU was fully utilised, each model was run using different batch sizes and the GPU count was tested 3 times over 20 epochs, with vScaler recording the average result. To understand the results better, you can view a full list here!
Results
When analysing the results of the benchmarks for each model there are some important trends to be observed, the most significant of which is the near linear scaling of performance as the GPU count increases. For example, consider the case of ResNet-50 training with a batch size of 256. With a single GPU we were able to achieve almost 899 images / second, and with 16 GPUs this increased to 12497 images/second. This is close to a 14x speedup, which represents an ~87% efficiency. The same scaling performance can also be observed in the ResNet-152 results, which shows a 86% efficiency when scaling from 1 to 16 V100 GPUs.
To understand more about the NVIDIA® DGX-2 and vScaler, use the resources below:
If you would like to test the NVIDIA® DGX-2 yourself, then get in touch with us and we can book you into our Boston Labs today!
To understand more about the NVIDIA® DGX-2 and vScaler, use the resources below: