3.4 The NVIDIA Solution Portfolio: Domain Platforms & Blueprints
Key Takeaways
- Blueprint objective 1.6 asks you to explain the purpose and use case of various NVIDIA solutions, so the exam tests breadth of portfolio recognition rather than depth in any one product.
- The portfolio stacks in four layers: silicon and systems, platform software, domain frameworks, and delivery vehicles such as NGC, NIM, and NVIDIA Blueprints.
- Domain frameworks are named for their vertical: Clara and Holoscan for healthcare, Metropolis and DeepStream for vision, Isaac for robotics, DRIVE for autonomous vehicles, Aerial for telecom RAN, Morpheus for cybersecurity, and cuOpt for optimisation.
- Omniverse is the simulation and digital-twin platform, and it is the standard answer for any scenario about a physically accurate virtual replica or synthetic data generation.
- NIM packages an optimised model as a standard container with an industry-standard API endpoint, while an NVIDIA Blueprint is a reference workflow that wires several NIMs into a complete application pattern.
Last updated: August 2026
3.4 The NVIDIA Solution Portfolio: Domain Platforms & Blueprints
Blueprint anchor: Domain 1 objective 1.6 — "Explain the purpose and use case of various NVIDIA solutions." Sections 3.1–3.3 covered the horizontal stack that every AI workload uses (CUDA-X, AI Enterprise, NeMo, TensorRT, NIM). This section covers the rest of the catalogue: the domain frameworks and delivery vehicles that the exam names in scenario stems. You need a one-sentence purpose for each, not implementation depth.
1. The Portfolio Has Four Layers
Candidates who try to memorise the catalogue as a flat list of product names lose. Organise it as a stack and each new product slots into a layer you already understand.
| Layer | What It Provides | Examples |
|---|---|---|
| 1. Silicon & systems | The accelerated hardware itself | Data center GPUs (A100/H100/H200/B200), Grace CPU and Grace-Blackwell superchips, DGX and HGX systems, IGX for industrial/medical edge, Jetson for embedded edge, ConnectX NICs, BlueField DPUs, Quantum InfiniBand and Spectrum-X Ethernet switching |
| 2. Platform software | Runtime, libraries, orchestration, and management that are workload-agnostic | CUDA and CUDA-X, NVIDIA AI Enterprise, Base Command Manager, Run:ai, GPU Operator, Triton, DCGM, DOCA |
| 3. Domain frameworks | Vertical-specific models, pipelines, and SDKs built on layers 1–2 | NeMo, Riva, Metropolis/DeepStream, Clara, Holoscan, BioNeMo, Parabricks, Isaac, DRIVE, Omniverse, Aerial, Morpheus, Merlin, cuOpt, Earth-2, PhysicsNeMo |
| 4. Delivery vehicles | How the above is packaged, distributed, and consumed | NGC catalogue, NIM microservices, NVIDIA Blueprints, DGX Cloud, Fleet Command for edge fleets |
A scenario question is really asking: which layer-3 framework matches this vertical, and which layer-4 vehicle matches this consumption model?
2. Domain Frameworks — One Line Each
| Platform | Domain | Purpose in One Sentence |
|---|---|---|
| NeMo | Generative AI / LLMs | End-to-end framework for curating data and building, customising, aligning, and evaluating large language and multimodal models |
| Riva | Speech & conversational AI | GPU-accelerated speech recognition, translation, and text-to-speech for real-time voice applications |
| Metropolis / DeepStream | Vision AI | Application framework and SDK for building multi-stream video analytics pipelines (decode → infer → track → publish) |
| Clara | Healthcare | Healthcare-focused application frameworks and pretrained models spanning imaging and clinical workflows |
| Holoscan | Medical devices & real-time sensors | Low-latency streaming platform for AI on live sensor data inside medical and scientific instruments |
| Parabricks | Genomics | GPU-accelerated secondary analysis for genomic sequencing pipelines |
| BioNeMo | Drug discovery | Framework and models for biomolecular AI such as protein structure and molecular property prediction |
| Isaac | Robotics | Simulation, perception, and manipulation stack for developing and training robots |
| DRIVE | Autonomous vehicles | In-vehicle compute platform and software stack for AV perception, planning, and in-cabin AI |
| Omniverse | Simulation & digital twins | Platform for building physically accurate 3D virtual worlds, industrial digital twins, and synthetic training data |
| Aerial | Telecommunications | Software-defined, GPU-accelerated 5G/6G radio access network signal processing |
| Morpheus | Cybersecurity | Streaming AI framework for detecting anomalies, threats, and sensitive-data leakage in network and log telemetry |
| Merlin | Recommenders | End-to-end framework for training and deploying large-scale recommendation systems |
| cuOpt | Operations research | GPU-accelerated solver for routing, scheduling, and combinatorial optimisation problems |
| Earth-2 | Climate & weather | Platform for high-resolution AI-driven weather and climate simulation |
| PhysicsNeMo (formerly Modulus) | Scientific computing | Framework for physics-informed machine-learning surrogates of engineering simulations |
Naming trap: DeepStream is the SDK; Metropolis is the broader vision-AI platform and partner ecosystem that DeepStream belongs to. Both are acceptable answers to a video-analytics stem, but if an option names a non-vision product it is wrong regardless of wording.
3. Delivery Vehicles: NGC, NIM, Blueprints, Fleet Command
The layer-4 products describe how you get and run everything above. These appear on the exam in operations-flavoured questions.
- NGC (NVIDIA GPU Cloud) catalogue. The registry of GPU-optimised containers, pretrained models, Helm charts, and SDKs. Containers are versioned by framework and release tag (for example a monthly PyTorch image), which is how a cluster keeps driver, CUDA, and framework versions consistent across hundreds of nodes.
- NIM (NVIDIA Inference Microservices). A prebuilt container that bundles a specific optimised model with its runtime and exposes a standard, OpenAI-compatible API endpoint. The value proposition is that deploying a model becomes a
docker runplus a Kubernetes service rather than a bespoke serving project. - NVIDIA Blueprints. Reference workflows that compose several NIMs and supporting services into a complete application pattern — a RAG assistant, a video-search-and-summarisation pipeline, a digital human. A Blueprint is a starting architecture, not a single model.
- Fleet Command. Managed service for deploying and updating AI applications across distributed edge locations (stores, factories, hospitals) from a single control plane, which is the edge counterpart to what Base Command Manager does for a data center cluster.
- DGX Cloud. NVIDIA-designed, NVIDIA-supported AI training and inference capacity delivered as a service and hosted with leading cloud providers, bundled with the NVIDIA AI Enterprise software stack. Section 6.4 treats the on-prem-versus-cloud trade-off in detail.
4. Choosing Between Them in an Exam Scenario
Work the stem in this order:
- Is the question about a vertical? Pick the layer-3 framework by industry — hospital device → Holoscan, factory twin → Omniverse, bank fraud stream → Morpheus, delivery routing → cuOpt.
- Is the question about consuming a model? Pick the layer-4 vehicle — one model behind a standard endpoint → NIM; a complete multi-service reference pattern → Blueprint; pulling a versioned framework container → NGC.
- Is the question about running or managing infrastructure? Drop to layer 2 — cluster provisioning and monitoring → Base Command Manager, GPU scheduling and quotas → Run:ai, driver and plugin lifecycle on Kubernetes → GPU Operator, serving many models → Triton, GPU health → DCGM.
- Is the question about a support or licensing guarantee? The answer is NVIDIA AI Enterprise, which is the supported, security-patched distribution with enterprise SLAs — not any individual framework.
Loading diagram...
Test Your Knowledge
A team wants to deploy a specific optimised large language model as a standard, self-hosted API endpoint without building a custom serving stack. Which NVIDIA delivery vehicle is designed for exactly this?
A
B
C
D
Test Your Knowledge
A telecommunications operator wants to run 5G radio access network signal processing on GPU-accelerated servers instead of fixed-function baseband hardware. Which NVIDIA platform addresses this?
A
B
C
D
Test Your Knowledge
How does an NVIDIA Blueprint differ from a NIM microservice?
A
B
C
D