For most of us, driving is second nature; a series of automatic decisions. To train AI for autonomous driving — to make those same decisions even a 10th of a second faster, and potentially make driving safer — requires petabytes of data.
According to the World Health Organization (WHO) (external link), approximately 1.35 million people die in road accidents every year, and another 50 million are injured. To mitigate this risk, the EU now requires the availability of self-driving vehicles by 2030. The race is on to provide the best technological path to fully autonomous driving.
“Advanced driver assistance systems react faster than drivers in critical accident scenarios,” says Robert Thiel, Head of Artificial Intelligence, Advanced Driver Assistance Systems (ADAS), at Continental. “This can be achieved by training an AI with tons of data to drive more safely than a human. Therefore, smart data management means smart vehicles and saved lives.”
Continental is a major supplier of automotive parts to nearly every car producer in the world and a leader in the autonomous driving intelligence space. Its ADAS business unit began developing intelligent sensors and data-driven traffic safety solutions more than 20 years ago, and has been working to increase the speed of development using deep learning and training artificial neural networks. The goal of Continental’s Vision Zero project is to virtually eliminate deaths by traffic accident through the development of ADAS technologies.
Continental has pushed the boundaries of automotive innovation for 150 years
Continental improved AI training time 70% using IBM Spectrum® Scale and NVIDIA DGX systems
Continental has the ability to run at least 14x more deep learning experiments per month at the same time
One of the automobile industry’s greatest challenges pertaining to autonomous driving is managing data located all over the world and using that data where it is needed. Continental’s ADAS Vision Zero initiative employs a test fleet equipped with sensors that drive 15,000 kilometers per day, generating and recording over 100 TB of data, which is then ingested, processed, selected, assessed and annotated, and used for training and validation of the system.
To detect what is happening in any given scenario and enable decisions to safely control the vehicle, the team uses NVIDIA DGX (external link) systems for training and validation. To speed development of the AI and reduce time-to-market, Continental needs high performance AI processing and access to data plus a powerful storage solution to analyze hundreds of thousands of images per second with NVIDIA GPU computing.
Continental’s ADAS solutions can support drivers in many typical driving tasks and even take control of the vehicle to avoid an accident. But as automation of its driver assistance and vehicle safety systems increased, software complexity grew, as did the number of safety requirements across multiple geographies. Continental realized it was time to scale both its technology and its teams to evolve a more globally scalable AI solution. The need for parallel data access also meant facing down a growing data management challenge.
Continental needed a powerful parallel file system to meet the high-speed demands of AI and to protect sensitive data. At the same time, it had to create a more centrally accessible place to store data and improve traceability, offering many ways for developers to securely connect.
Continental knew it was time to boost performance with scalable deep learning infrastructure and storage connected with a high-speed network. This solution would need to provide fast random access, support protocols such as Server Message Block (SMB) and Amazon Simple Storage Service (S3), and provide several different access management options.
“GPUs today are so fast that standard storage cannot keep up with the compute,” says David Enenkel, Head of IT Operations, ADAS, at Continental. “That’s why we were looking for something faster, something that really gives us the bandwidth and also the random access that we need.”
Continental established comprehensive testing and evaluated how well each of the top storage solutions measured up to its goals. To measure the performance of IBM ESS, Continental worked with IBM Business Partner SVA System Vertrieb Alexander GmbH. They found that the IBM Spectrum Storage for Data and AI with NVIDIA DGX solution provided the required performance and several other advantages. The “parallel” high performance architecture and simple-to-scale node deployment of the solution provided the AI infrastructure required, with the resilience and scalability Continental will need in the future.
The flexibility and seamless integration of IBM Storage with Kubernetes containers allowed Continental to modernize its application development without giving up on infrastructure requirements like performance, scalability or simplicity. The IBM Spectrum Scale solution ensured that its infrastructure will support the growth required, whether in the cloud or on premises. IBM has extensive experience in the automotive industry, which was also a contributing factor in Continental’s decision.
With the new solution, Continental was able to optimize for deep learning with multi-node training, enabling it to increase model accuracy for higher levels of safety without impacting time to production. Continental scaled to a cluster of DGX to handle 14 times more experiments per month, with the ability to test millions of permutations in environmental conditions — such as rain, snow, sunlight and clouds — or transients — such as cars moving too close to one another during a lane change.
With the performance improvements, flexibility and scalability of the new IBM data management solution to support an evolving AI infrastructure, Continental is on the fast track to change the future of mobility.
“We couldn’t sell any of the systems that we sell today, regarding safety requirements, without the ability to validate on huge data sets — in the range of millions of kilometers or dozens of petabytes to be processed on a regular basis, re-simulated, collected and for some kind of KPIs to be generated,” says Thiel.
“As a result of our new infrastructure, we can now run 20, 40, 80 GPUs simultaneously to really speed up our training,” says Balazs Lorand, PhD, Head of AI Competence Centre, ADAS@Budapest, at Continental. “We are proud to have solved several challenges,” he continues. With this new infrastructure, Continental achieves 14 times more deep learning experiments per month and has reduced training time from weeks to days. It has dramatically increased the efficiency of the development life cycle as it is now able to conduct more experiments and seamlessly connect its Kubernetes environment. And its solution is flexible enough to support growth in any direction — in containerized hybrid cloud environments, on premises and in multiple data centers.
Continental built a completely new infrastructure in the Frankfurt, Germany AI-ready data center at Equinix (external link), a global colocation infrastructure provider. Continental, with the support of SVA, implemented the overall integration of the storage solution in the cluster, including installation, deployment, configuration, commissioning and training for operation and administration.
This new solution includes a multimode GPU cluster, non-blocking InfiniBand network infrastructure, IBM ESS with fast Non-Volatile Memory express (NVMe) drives, NVIDIA DGX systems and NVIDIA V100 Tensor Core GPUs. Continental is also using IBM Spectrum Scale software with its Kubernetes environment for modern application development.
“To meet the demands of a five-star rating with Euro NCAP, you need to keep developing more intelligent products. So it’s very, very important to establish such data environments, and I’m really happy that we did so last year,” says Enenkel.
These improvements translate to a strong competitive position for Continental, enabling it to move forward with new, safer autonomous driving solution development more quickly than ever before.
Continental develops pioneering technologies and services for sustainable and connected mobility of people and their goods. Founded in 1871, the technology company offers safe, efficient, intelligent and affordable solutions for vehicles, machines, traffic and transportation.
The Autonomous Mobility and Safety business area develops and produces integrated active and passive driving safety technologies as well as products that support vehicle dynamics.
About SVA System Vertrieb Alexander GmbH
IBM Business Partner SVA is a leading German system integrator with 23 branch offices across Germany. SVA focuses on the combination of high quality IT products with project know-how and flexibility to achieve optimum solutions in the fields of data center infrastructure, business continuity, big data, IT security and cloud.
© Copyright IBM Corporation 2021. IBM Corporation, IBM Storage, New Orchard Road, Armonk, NY 10504
Produced in the United States of America, January 2021.
IBM, the IBM logo, ibm.com, IBM Elastic Storage, and IBM Spectrum are trademarks of International Business Machines Corp., registered in many jurisdictions worldwide. Other product and service names might be trademarks of IBM or other companies. A current list of IBM trademarks is available on the web at “Copyright and trademark information” at www.ibm.com/legal/copytrade.shtml.
This document is current as of the initial date of publication and may be changed by IBM at any time. Not all offerings are available in every country in which IBM operates.
The performance data and client examples cited are presented for illustrative purposes only. Actual performance results may vary depending on specific configurations and operating conditions. THE INFORMATION IN THIS DOCUMENT IS PROVIDED “AS IS” WITHOUT ANY WARRANTY, EXPRESS OR IMPLIED, INCLUDING WITHOUT ANY WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND ANY WARRANTY OR CONDITION OF NON-INFRINGEMENT. IBM products are warranted according to the terms and conditions of the agreements under which they are provided.
Actual available storage capacity may be reported for both uncompressed and compressed data and will vary and may be less than stated.