The U.S. Postal Service is the world’s largest postal service, with 485 million mail pieces processed and delivered daily. The new advanced AI system, says the company, will enable the agency to process package data 10 times faster with higher accuracy.
"AI is transforming multiple industries, enabling processes, accuracy and efficiency not possible before," says Anthony Robbins, vice president of the Federal Sector Business at Nvidia. "The U.S. Postal Service's adoption of AI demonstrates how this powerful technology can improve an excellent service that we rely on every day. Benjamin Franklin would be proud."
The new system is based on high-performance servers powered by Nvidia V100 Tensor Core GPUs and deep learning software to train multiple AI algorithms. The trained models are then deployed to Nvidia EGX edge computing systems at close to 200 Postal Service facilities throughout the U.S. to enable more efficient package data processing. The systems are being purchased by the Postal Service under contract with Hewlett Packard Enterprise.
Engineering teams from the USPS and Nvidia have been collaborating for several months to develop AI models using the following tools:
- Nvidia software including TensorRT for high-throughput, low-latency inference optimization
- automatic mixed precision in PyTorch to accelerate training while maintaining model accuracy
- NGC containers , which are GPU-optimized for streamlining software deployment
- DeepOps tools for optimizing GPU clusters
Delivery and testing of the system will start this year and it is expected to be fully operational by spring of 2020.