Nvidia's Supply Chain and AI Integration

Jun 8, 2024

Nvidia's Supply Chain and AI Integration

Overview

  • Nvidia operates one of the largest and most complex supply chains.
  • Example: A single Nvidia Hopper DGX has nearly 40,000 parts from over 150 suppliers and weighs nearly 350 pounds.
  • Nvidia's supercomputers connect thousands of GPUs with hundreds of miles of high-speed optical cables.
  • Rely on hundreds of partners to deliver thousands of components to a dozen factories to build nearly 3,000 products.
  • Single disruptions can impact commitments.

AI Planner

  • Developed by Nvidia’s business operations team to manage supply chain data.
  • Built on Nvidia Inference Microservices (NIM).
    • NIM is a container with pre-trained models and CUDA acceleration libraries.
    • Easy to download, deploy, and operate on-premises or in the cloud.

Components of AI Planner

  • Uses multiple NIMs for different functions:
    • LLM NIM: Understands planner’s intention and directs other models.
    • Nemo Retriever RAG NIM: Connects LLM to proprietary data.
    • KOP NIM: For logistics optimization.
      • Nvidia KOP (Combinatorial Optimization Library) is GPU-accelerated and holds 23 world record benchmarks.

Operations Research

  • Concerns efficient decision-making for highly complex logistics problems.
  • Examples include traveling salesman, vehicle routing, supply chain optimization.
  • Industries spend billions on routing operations; efficiency gains can save millions.
  • Modeling and optimization are complex due to numerous variables, constraints, and real-time changes.

KOP: Optimization Brain

  • Acts as a supercomputer for decision-making.
  • Uses GPU-accelerated heuristics and evolutionary algorithms.
  • Calculates optimal routes in seconds and dynamically recalculates based on changes.
  • Allows analysis of thousands of scenarios in real-time using natural language.

Example Scenario

  • AI planner reallocates products based on changing conditions.
  • Scenario: Reallocating RTX 6000 to customers with highest demand given supply and production changes.
    • Traditional replanning: Overnight duration.
    • AI Planner: Takes seconds, providing agility and speed.