Computer vision capabilities are having a significant impact in almost every industry, from autonomous vehicles to AI-assisted medical diagnosis. Training the machine learning (ML) algorithms used for computer vision applications creates an extremely demanding workload, requiring massive quantities of data and significant computing power.
To accommodate these workloads, you can use a clustered architecture consisting of NetApp® storage systems and Fujitsu PRIMERGY servers optimized for AI. This NetApp and Fujitsu solution is designed to handle large datasets by using the processing power of GPUs alongside traditional CPUs. The combined solution of PRIMERGY servers and NetApp all-flash storage systems provides an infrastructure that delivers excellent performance and seamless scalability with industry-leading data management.
State-of-the-art NetApp AFF storage systems enable IT departments to meet enterprise storage requirements with industry-leading performance, cloud integration, and best-in-class data management. The Fujitsu PRIMERGY GX2570 server is an extremely powerful deep-learning (DL) platform that benefits from equally powerful storage and network infrastructure to deliver maximum value.
To automatically construct the system infrastructure for this solution, you can use Ansible, a DevOps-style configuration management tool developed by Red Hat. Ansible offers a variety of functional modules from NetApp and Cisco. It includes modules for the Fujitsu PRIMERGY GX2570 M5, for storage such as the NetApp AFF A800 array, and for automatic construction and configuration management of Cisco Nexus 3232C network switches. Ansible makes it easy to add GPU nodes and change the software environment on the host OS, greatly reducing the load on system administrators.
To validate the solution, NetApp and Fujitsu used one NetApp AFF A800 storage system, four Fujitsu PRIMERGY GX2570 servers, and two Cisco Nexus 3232C 100Gb Ethernet (100GbE) switches.
We validated the solution by using the MLPerf v0.6 benchmark models and testing procedure. Each MLPerf training benchmark measures the processing time required to train a model on the specified dataset to achieve the specified quality target. The following table shows the training time involved for each of the models.
Model | Training time result |
SSD | 19.54 minutes |
Mask R-CNN | 186.22 minutes |
ResNet-50 | 94.76 minutes |
Minigo | 24.97 minutes |
To learn more about the joint solution, read this technical report.
Add comment