Compute Infrastructure

NVIDIA Tesla P100 GPU on IBM Cloud: 2.8 times more performance than previous generation

Share this post:

Now that the new NVIDIA Tesla P100 GPU cards are in house, everyone is curious to know what the new bells and whistles are on the latest version, so to speak.

Our IBM engineers have been performing some in-house testing and have found that the NVIDIA Tesla P100 GPU cards can provide up to 2.8 times more performance than the previous generation NVIDIA Tesla K80 for certain deep learning workloads running on the IBM Cloud. The combination of NVIDIA Tesla P100 GPUs on the IBM Cloud reduced the corresponding training time for a deep learning image classification model by 65 percent from the NVIDIA Tesla K80 GPUs.

Increased processing capabilities

Fabulous news for those working with deep learning, AI, and high-performance data analytics—which all demand increased processing capabilities. This also makes it faster and more cost-effective to leverage deep learning techniques to train AI systems.

To conduct the benchmark, IBM engineers trained a deep learning model for image classification using two NVIDIA Tesla P100 GPU PCIe cards on IBM Cloud bare metal servers and compared the results to the same deep learning model running two NVIDIA K80 GPU PCIe cards (a total of four K80 GPU cores) on the same servers. The engineers conducted the ISLVRC image classification challenge using the VGG-16 deep neural network on the Caffe framework. The goal of the ISLVRC is to teach a deep neural network model to correctly classify images; models are trained on approximately 1.2 million images with an additional 50,000 images for validation and 100,000 images for testing.

The benchmark also found that the NVIDIA Tesla P100 GPUs on IBM Cloud can deliver greater performance for the current cost. According to the benchmark, the NVIDIA Tesla P100 GPU on IBM Cloud can process more than 116,000 images per US dollar spent – 2.5 times higher than the previous generation NVIDIA Tesla K80 GPUs on the cloud for the same test case.

The IBM benchmark results were achieved using the following criteria:

  • Two NVIDIA Tesla P100 GPU PCIe cards (a total of two P100 GPU cores) on IBM Cloud bare metal servers (with Dual Xeon E5-2690v4 processors) running the VGG-16 deep neural network on the Caffe-1.0.0-rc5 framework, CUDA version 8.0.61, NCCL version 1.3.4, cuDNN version 6.0.20, and CUDA driver version 375.51. The ILSVRC image classification challenge dataset was used. The training batch size was maximized to exploit the larger available memory capacity on the NVIDIA Tesla P100 GPU cards.  
  • Two NVIDIA Tesla K80 GPU PCIe cards (a total of four K80 GPU cores) on IBM Cloud bare metal servers (with Dual Xeon E5-2690v4 processors) running the VGG-16 deep neural network on the Caffe-1.0.0-rc5 framework, CUDA version 8.0.61, NCCL version 1.3.4, cuDNN version 6.0.20, and CUDA driver version 375.51 . The ILSVRC image classification challenge dataset was used. The training batch size was maximized to use all available memory capacity on the NVIDIA Tesla K80 GPU cards.  

 Performance may vary depending on your specific workloads and environments.

More Compute Infrastructure stories
February 12, 2019

Intel on IBM Cloud: The Latest Microarchitecture for Processor Enthusiasts

IBM Cloud is excited to announce the upcoming availability of IBM Cloud Bare Metal and Virtual Servers powered by the new Cascade Lake family of Intel® Xeon® Scalable Processors.

Continue reading

February 12, 2019

IBM POWER9 on IBM Cloud: More Choice and Flexibility, On-Premise and Off

To help clients succeed in today’s hybrid, multicloud reality, the IBM Cloud and IBM Power Systems teams are bringing the IBM POWER9 platform to the cloud with the new IBM Power Systems Virtual Server on IBM Cloud—available in select IBM Cloud data centers spring 2019.

Continue reading

February 11, 2019

IBM Cloud Internet Services is Enhancing the Edge

We are thrilled to announce two very exciting enhancements—"Range" and "Edge Functions"—to IBM Cloud Internet Services (CIS).

Continue reading