Robust and flexible platform supporting large-scale GPU cluster for AI-driven autonomous vehicles boosts Turing’s speed-to-market by 75%.
Digital Realty, a global provider of cloud-and carrier-neutral data centre, colocation and interconnection solutions, has announced that Japanese Artificial Intelligence company Turing, has established its cutting-edge computation platform for full driving automation at Digital Realty’s NRT10 data centre in Japan.
Turing, a pioneer in developing full driving automation technology, is creating an end-to-end self-driving system powered by AI to control all driving functions, including steering, acceleration and braking, using only vehicle camera data.
The company’s advancements include ‘Heron,’ a multimodal generative AI for sophisticated decision-making using text and visual data, and ‘Terra,’ a generative world model that simulates realistic driving scenarios as videos. Turing’s flagship project, ‘Tokyo30’, aims to achieve Level 5 full driving automation for 30 minutes or more in Tokyo’s urban areas by the end of 2025.
To support this ambitious initiative, Turing has developed the Gaggle Cluster, a powerful computation platform equipped with 96 NVIDIA H100 GPUs. NRT10, a certified NVIDIA DGX-ready data centre, offers a high-density power supply and advanced infrastructure that maximises GPU performance and efficiency.
Operated by MC Digital Realty, Digital Realty’s joint venture with Mitsubishi Corporation in Japan, NRT10 is part of PlatformDIGITAL, Digital Realty’s global data centre platform. The facility delivers several key benefits for Turing:
• Accelerated time-to-market: Turing’s AI development timeline has been reduced from one year to just three months compared to an on-premise build.
• Ready-to-deploy, high-density colocation solution: NRT10’s modular design allows rapid configuration and deployment, reducing infrastructure requirements from 20 racks to just eight while enhancing efficiency and scalability.
• Maximised GPU performance: NRT10’s advanced cooling and power systems ensure optimal GPU efficiency and reliability.
• Flexible expansion: The campus-style configuration of NRT10 supports easy scalability as Turing’s computational needs evolve.
Yu Yamaguchi, CTO at Turing, said: “As the global race to develop full driving automation systems accelerates, securing computing resources quickly is crucial. Digital Realty’s high-performance and flexible data centre enabled us to swiftly deploy a powerful GPU cluster that delivers maximum performance in a short period of time, further accelerating our AI development for full driving automation.”
Serene Nah, Managing Director and Head of Asia Pacific at Digital Realty, said: “Digital Realty is excited to be at the forefront of this transformative journey, leveraging our extensive AI experience and expertise to support Turing’s groundbreaking AI innovation. By providing a high-performance and flexible data centre environment, we are enabling Turing to rapidly deploy and scale their critical AI computations. This collaboration not only accelerates the development of Turing’s full driving automation technology but also sets a new standard for the efficient and effective deployment of AI solutions. We are committed to helping Turing and other innovative companies achieve their goals, shaping the future of how the world operates through our advanced AI infrastructures and forward-thinking data centre solutions.”
NRT10’s capabilities and Digital Realty’s operational expertise position the facility as a cornerstone for supporting next-generation AI workloads, empowering organisations like Turing to advance AI-driven innovation and accelerate speed-to-market for revolutionary technologies.
We asked Yu Yamaguchi, CTO at Turing, further questions to find out more about the project.
How did the infrastructure at Digital Realty’s NRT10 data centre align with the specific needs of Turing’s AI-driven full driving automation projects?
Developing autonomous driving AI requires an enormous amount of computing resources, and the newly introduced NVIDIA DGX H100 system has strict installation requirements. It must be deployed in a data centre certified as an NVIDIA DGX-Ready Data Center. In Japan, there is a shortage of data centres that have obtained this certification. Among them, MC Digital Realty’s NRT10 stood out for its high scalability and the successful execution of its expansion plan.
Can you elaborate on the role of the Gaggle Cluster and its NVIDIA H100 GPUs in advancing the development of Turing’s ‘Tokyo30’ project and achieving Level 5 automation goals?
We are currently building a massive proprietary driving dataset focused on the Tokyo area. Gaggle-Cluster leverages this ever-expanding dataset to develop autonomous driving AI models suited to Japan’s complex road environments. Not only do these models require enormous computing resources for training, but completing training as rapidly as possible is also critical to facilitate further testing. Gaggle-Cluster offers a level of cost efficiency and computing efficiency that cannot be achieved through cloud services, making it an essential engine for autonomous driving model development.
What were the key factors that led you to choose a colocation solution at NRT10 over building on-premises infrastructure, and how has this decision impacted your development timelines?
AI development infrastructure, including GPUs, is evolving rapidly, and the requirements for data centres are becoming more stringent by the day. It is challenging for a startup with limited funds and resources to build and maintain a data centre requiring such complex and specialised expertise. By choosing colocation this time, we were able to build this cluster in about three months and have it running smoothly without any issues.
How does Turing leverage AI models like ‘Heron’ and ‘Terra’ within the NRT10 data centre to simulate complex driving scenarios and enhance decision-making capabilities?
Achieving full autonomy in driving requires the development of autonomous driving AI that can handle edge cases rarely encountered in everyday driving. To address these edge cases, we are leveraging the AI model ‘Heron,’ which has a deep understanding of Japanese culture and driving conditions, as well as the world model ‘Terra’, which can generate driving scenarios for edge cases that cannot be captured in the real world.
As Turing scales its AI workloads, how does the flexibility and scalability offered by NRT10’s campus-style configuration support your evolving computational requirements?
A high-speed communication network between GPU nodes is essential for scaling up computing capacity. Because this network relies on cutting-edge specifications, wiring costs are very high, and any communication latency directly impacts computational efficiency. With a campus-style data centre, we can secure scalability while minimising wiring complexity and costs, which makes it especially appealing for us as we plan to expand.
What role do data centre partners like Digital Realty play in accelerating AI innovation, and how critical are these partnerships in the competitive race to achieve full driving automation?
The race to develop autonomous driving AI is accelerating. In circumstances where even a single-year delay can lead to a critical technological gap, a data center that can consistently introduce and operate on the latest device platforms is considered an essential partner for maintaining our competitiveness.
