Turing (Tokyo) has completed the construction of the Gaggle Cluster, a dedicated computing platform for fully autonomous driving, and has begun operation.
The construction of the computing platform was made possible with support from the Docomo Group, including technical support from NTTPC Communications and investment from NTT Docomo Ventures.
NTTPC provided the computing platform in the form of a GPU private cloud that utilizes interconnect technology for high-speed communication between GPU nodes, storage design know-how, and know-how as a cloud operator in monitoring and operating networks, data centers, and equipment.
The Gaggle Cluster is Turing’s dedicated computing platform equipped with 96 NVIDIA H100 GPUs. The network uses NVIDIA InfiniBand/NDR400 to minimize the communication speed constraints between servers that were a bottleneck when using multiple GPUs simultaneously for large-scale AI learning. In addition, the adoption of All-Flash distributed storage maximizes performance in distributed learning, and the entire cluster is optimized for large-scale deep learning tasks as a “single computer.”

Turing has developed its own autonomous driving AI “TD-1” by utilizing Gaggle Cluster and has begun road testing.

TD-1 is an autonomous driving AI for Turing’s “Tokyo 30” project, and is a Transformer model that takes only camera images as input and outputs surrounding maps, vehicle/pedestrian recognition, and even its own driving operations from a single model.

※Translating Japanese articles into English with AI