Recorded talk

Talk to Your Supply Chain Data Using NIM

enplatform captions

Summary

Nvidia faces extremely large, interdependent supply-chain decision problems (a single Hopper DGX can contain ~40,000 parts from 150+ suppliers across many factories) that require fast replanning when disruptions occur. Their business operations team built an AI planner assembled from NIMs (Nvidia Inference Microservices): an LLM Nim for intent, a Nemo retriever RAG Nim for proprietary data access, and a Nvidia coop Nim for GPU-accelerated combinatorial optimization. Nvidia coop (a GPU-accelerated operations-research library with 23 world-record benchmarks) computes optimal routes in seconds and dynamically recalculates when constraints or requirements change. The result is real-time, natural-language driven scenario analysis and product reallocation that reduces replanning from overnight to seconds, enabling faster operational response to supply changes.

Key technical points

Scale and complexity: a single Nvidia Hopper DGX contains nearly 40,000 parts from over 150 suppliers and weighs ~350 lb; Nvidia products and supercomputers involve tens of thousands of GPUs and hundreds of miles of optical cable.
Supply-chain fragility: hundreds of partners, thousands of components, and multiple factories create a system where a single disruption can materially impact delivery commitments.
AI planner concept: an agent built from NIMs (Nvidia Inference Microservices) enables users to 'talk to' supply-chain data via natural language.
NIM definition and deployment: a NIM is a containerized package with pre-trained models and CUDA acceleration libraries that is easy to download, deploy, and operate on premises or in the cloud.
Modular assembly: the AI planner composes multiple NIMs — an LLM Nim to interpret planner intent, a Nemo retriever RAG Nim to link the LLM to proprietary data, and a Nvidia coop Nim for logistics optimization.
Nvidia coop description: a GPU-accelerated combinatorial optimization library for operations research that holds 23 world-record benchmarks.
Operations-research scope: coop targets classic hard problems (traveling salesman, vehicle routing, supply-chain optimization) characterized by many variables, constraints, and interdependencies.
Search-space difficulty: realistic logistics/search spaces are extremely large and change in real time, making accurate modeling and optimization very complex.
Optimization methods: coop leverages GPU acceleration with heuristics and evolutionary algorithms to act as a supercomputer for decision-making.
Latency and dynamism: coop can calculate optimal routes in seconds and dynamically recalculate when requirements or constraints change.
Natural-language control: the planner allows real-time analysis of thousands of scenarios driven by natural language inputs.
Example scenario: the planner can reallocate an RTX 6000 product to customers given changes such as a 40% increase in Taiwan supply, production decreases in Hong Kong, and a 20% increase in global demand while acknowledging constraints.
Outcome vs. legacy process: replanning this scale of supply-chain optimization would normally take at least overnight; the coop-powered AI planner completes it in seconds.
Integration pattern: intent understanding (LLM) → retrieval of proprietary data (RAG/Nemo) → combinatorial optimization (coop NIM) is the described production pipeline.

Engineering relevance

The talk describes a modular, containerized architecture for supply-chain automation: NIMs provide packaged, CUDA-accelerated model services that can be deployed on-premises or in cloud environments. Key engineering elements are LLMs for intent parsing, a retrieval-augmented Nemo Nim to connect to proprietary datasets, and Nvidia coop as a GPU-accelerated combinatorial optimizer that runs heuristics/evolutionary algorithms to produce near-real-time optimal routes. Performance claims include computing optimal routes in seconds and holding 23 world-record benchmarks for operations-research tasks; the system supports dynamic recalculation when constraints change. Engineers should note the pipeline pattern (intent → data retrieval → optimization), the need for low-latency GPU resources for coop, and integration points between natural-language interfaces and deterministic optimization components.

Japan relevance

For engineering teams in Japan working on manufacturing, logistics, or large-scale supply-chain systems, this approach offers a blueprint for combining LLM-driven intent, RAG-based proprietary data access, and GPU-accelerated combinatorial optimization to enable fast, constraint-aware replanning. The NIM container model supports on-premises deployment, which can align with Japanese requirements for local data control and factory-proximate compute. The ability to turn complex what-if changes (e.g., regional supply shifts, production drops, and demand surges) into seconds-scale reallocation decisions can improve responsiveness for just-in-time manufacturing and multi-supplier coordination common in Japanese supply chains.

Full transcript

Transcript

platform captions

[Music] Nvidia operates one of the largest and most complex Supply chains in the world a single Nvidia Hopper dgx has nearly 40,000 parts from over 150 suppliers and weighs nearly 350 lb and the supercomputers we build connect tens of thousands of Nvidia gpus with hundreds of miles of high-speed optical cables we rely on hundreds of Partners to deliver thousands of different components to a dozen factories to build nearly 3,000 products a single disruption to our supply chain can impact our ability to meet our commitments to help us manufacture and deliver AI Factories at extraordinary scale nvidia's business operations team is developing an AI planner that lets us P to our supply chain data the AI planner is an agent built on Nim Nvidia inference microservices a Nim is a container with pre-trained models and Cuda acceleration libraries that is easy to download deploy and operate on premises or in the cloud our supply chain team assembles their AI planner with an llm Nim to understand planner's intention and direct the other models Nemo retriever rag Nim to connect the llm to proprietary data and and a kop Nim for Logistics optimization Nvidia coop is a state-of-the-art GPU accelerated combinatorial optimization library for operations research and holds an incredible 23 world record benchmarks operations research is the discipline of efficient decision-making for highly complex Logistics problems such as the traveling salesman vehicle routing or supply chain optimization problems Industries spend billions in rooting operations so large efficiency gains can easily save hundreds of millions in costs but accurately modeling and optimizing these scenarios is extremely complex these are intricate networks with numerous variables constraints and interdependencies that change in real time the resulting search space is many orders of magnitude larger than the number of atoms in the universe kop acts as a supercomputer for this type of decision- making using GP pu accelerated heris sixs and evolutionary algorithms it calculates an optimal route in seconds and dynamically recalculates when requirements change with kop as the optimization brain behind our AI planner our operations team can analyze thousands of possible scenarios in real time using natural language inputs to talk to our supply chain data let's look at an example where the AI planner automatically reallocates a product to customers based on changing conditions please show me what the new RTX 6000 allocation would be for attend customers with the highest demand if there's 40% increase in Taiwan Supply 3we production Falls in Hong Kong and 20% increase in global demand acknowledging constraints solving here's how this will impact those customers normally replanning such a complex large-scale supply chain would have taken at least overnight but with our cop powered AI planner it takes just seconds giving our team the Speed and Agility to respond to an Ever shifting supply chain