NVIDIA Holoscan SDK
Holoscan is NVIDIA's open source SDK for real-time AI processing of sensor streams, such as surgical video, ultrasound, cameras and radio signals, at the edge or in the cloud. It runs on IGX, Jetson, DGX Spark and x86 systems with NVIDIA GPUs and is licensed under Apache-2.0.12
Also known as Holoscan, Holoscan SDK
At a glance
- What is it?
- Holoscan is a sensor processing platform built around an open source SDK. An application is a graph of operators, small units of work that receive data, process it on the GPU and pass it on. The SDK supplies the runtime that schedules these operators, passes data between them, and can split one application across several machines. NVIDIA describes it as domain-agnostic, although many reference applications come from medical devices. Version 4.6 is the latest in the documentation. Holoscan Sensor Bridge, a separate open project, connects sensors to the GPU over Ethernet through an FPGA interface.134
- What does it do?
- Holoscan lets a team build a pipeline that takes raw data from a sensor, runs signal processing and AI models on it, and shows or forwards the result within a tight time budget. Built-in operators cover input and output, model inference, processing and visualization (Holoviz). The runtime offers several schedulers, distributed applications across nodes, and data flow tracking to measure where latency comes from. HoloHub adds community reference applications, operators, tutorials and benchmarks, from endoscopy tool tracking to software-defined radio.35
- Who needs it?
- Engineers building devices or instruments where data must be processed as it arrives, with predictable delay: surgical and imaging systems, ultrasound and endoscopy, industrial inspection, scientific instruments, software-defined radio receivers, and robots. It suits C++ or Python developers deploying on NVIDIA edge hardware such as IGX or Jetson, or on x86 workstations.25
- What does it need?6
- A supported platform: Jetson AGX Thor (JetPack 7.0), IGX Orin (IGX Software 1.1.1), Jetson AGX Orin or Orin Nano (JetPack 6.2.1), DGX Spark, GH200 (headless) or an x86_64 workstation
- On x86: Ubuntu 22.04 or 24.04 (RHEL 9 through the container only) with an Ampere or newer GPU recommended
- NVIDIA driver 535 or newer for discrete GPUs
- CUDA 13 or CUDA 12 depending on platform; Python 3.10 to 3.13 for the Python wheels
- A Quadro or NVIDIA RTX series GPU if you need GPUDirect RDMA
- Torch or ONNX Runtime installed separately if your models need them (Debian and wheel installs)
- What it is not
- Holoscan is software, not a medical device; using it does not make a product cleared for clinical use, and regulatory work stays with the device maker. It is not IGX hardware either, although IGX is a common target. It is not a general video analytics toolkit for many city or retail cameras; that is DeepStream's role. And it is not limited to healthcare: NVIDIA positions it as domain-agnostic.7
Availability and licensing. The Holoscan SDK is open source under Apache-2.0 and can be built from source on GitHub. It is distributed as an NGC container, Debian packages, Python wheels on PyPI and Conda packages, with no price stated on the developer page. Version 4.6.0 was published on GitHub on 3 September 2026, following 4.5.0 on 31 July 2026. Holoscan Sensor Bridge is a separate Apache-2.0 repository, and HoloHub is a separate open source collection of operators and applications.138
The problem it solves
Medical devices, instruments and robots produce continuous streams of high-bandwidth data. Moving that data from a sensor into a GPU, running AI on it and displaying the result usually takes several copies through system memory and a lot of custom code, which adds delay and makes timing hard to predict.
Each team also tends to rebuild the same plumbing: capture drivers, scheduling, inference wrappers and visualization.
Holoscan provides that plumbing as a reusable runtime and operator library, with GPUDirect-based sensor input and tools to measure latency, so teams can focus on their own algorithms.
How it works
- Capture. Sensor data arrives through capture cards, cameras or Holoscan Sensor Bridge, which streams sensor data over Ethernet; reference applications use GPUDirect to write it straight into GPU memory.
- Build the graph. Developers connect operators through input and output ports into fragments; an application is one or more fragments.
- Schedule. A scheduler decides when each operator runs; NVIDIA recommends the event-based scheduler for parallel pipelines. Conditions control whether an operator is ready.
- Infer and process. The inference operator runs models with TensorRT, and optionally Torch or ONNX Runtime, alongside custom CUDA or Python processing.
- Visualize or forward. Holoviz renders results, or data is sent onward to other systems.
- Distribute and measure. Fragments can run on different machines, and data flow tracking reports latency between operators.
When an application runs on the GXF executor, each operator corresponds to a GXF component (a codelet); a GPU-resident executor is also available.14
Diagram as a list
Applications & solutions
- Sensors: endoscopes, ultrasound, cameras, radiosProduce raw streaming dataConnects to Capture cards, GPUDirect and Holoscan Sensor Bridge
- Holoviz and output operatorsDisplay results or forward them
Models & frameworks
- Inference operator (TensorRT, Torch, ONNX Runtime)Runs AI models on the streamConnects to Holoviz and output operators
Inference & runtime software
- Operator graph (fragments, ports, conditions)The application's processing stepsConnects to Inference operator (TensorRT, Torch, ONNX Runtime), Holoviz and output operators
Operations & orchestration
- Schedulers and executors (GXF-based)Decide when operators runConnects to Operator graph (fragments, ports, conditions)
- Data flow trackingMeasures latency between operatorsConnects to Operator graph (fragments, ports, conditions)
Accelerated computing
- IGX, Jetson Thor and Orin, DGX Spark, x86 GPUsRuns the application at the edge or in a labConnects to Operator graph (fragments, ports, conditions)
Networking, power & facilities
- Capture cards, GPUDirect and Holoscan Sensor BridgeMove sensor data into GPU memoryConnects to Operator graph (fragments, ports, conditions)
Capabilities
Operator graph runtime48
Applications are graphs of operators connected through ports, organized in fragments, with conditions, resources and several schedulers.
Why it matters: Gives a structured way to build low-latency streaming pipelines instead of custom threading code.
Limits: The MultiThreadScheduler is now legacy; new apps should use the event-based scheduler, and API changes between 4.x releases can require code updates.
Distributed applications4
Fragments of one application can run on different physical nodes, with the runtime handling communication between them.
Why it matters: Lets heavy processing run on a separate GPU node while capture stays near the sensor.
Limits: Network latency between nodes adds to the end-to-end time budget.
AI inference and visualization operators68
Built-in operators for IO, ML inference, processing and visualization; the inference operator picks optimized TensorRT paths automatically for eligible pipelines, and Holoviz renders output.
Why it matters: Covers the common steps of a sensor AI pipeline without writing them from scratch.
Limits: Torch and ONNX Runtime must be installed separately for Debian and wheel installs; the container lacks the ONNX Runtime backend.
Sensor input with GPUDirect and Sensor Bridge110
Reference applications use GPUDirect for direct transfer to GPU memory, and Holoscan Sensor Bridge offers an FPGA interface with a standard API for bringing sensor data over Ethernet.
Why it matters: Cuts copies between sensor and GPU, which is often the main source of delay.
Limits: GPUDirect RDMA needs Quadro or NVIDIA RTX series GPUs; Sensor Bridge requires compatible FPGA hardware.
Latency measurement3
Data flow tracking profiles an application and analyzes data flow between operators in its graph.
Why it matters: Makes it possible to show where time is spent, which matters for devices with timing requirements.
Limits: Measurements describe the tested setup only; results change with hardware and load.
HoloHub reference applications5
A public hub of reference applications, operators, modules, tutorials and benchmarks for Holoscan, from surgical video to software-defined radio and ROS 2 robots.
Why it matters: Gives working starting points close to many real projects.
Limits: Community examples vary in maturity and need review before production use.
Practical use cases
A surgical video system needs AI overlays, such as tool tracking, with very little delay.
- Approach
- Capture the endoscope feed with GPUDirect, run a tracking model through the inference operator and overlay results with Holoviz.
- Role of NVIDIA Holoscan SDK
- Provides capture, scheduling, inference and display in one low-latency pipeline.
- Data, infrastructure and skills
- Capture hardware, a validated model, an IGX or RTX workstation, and a regulatory plan if the system will be used clinically.
- Type of benefit
- Lower processing latency
- Caveats
- Clinical use needs device-level validation and regulatory clearance that Holoscan does not provide.
- First step
- Run the endoscopy tool tracking reference application from HoloHub with its sample data.
An industrial inspection or instrument maker must process a high-bandwidth sensor stream in real time.
- Approach
- Bring sensor data in through Holoscan Sensor Bridge or a capture card, run signal processing and AI operators on the GPU, and stream results to the control system.
- Role of NVIDIA Holoscan SDK
- Supplies the data path into the GPU and the streaming runtime.
- Data, infrastructure and skills
- Sensor and FPGA or capture hardware, domain algorithms, and edge hardware such as IGX or Jetson.
- Type of benefit
- Real-time decision support
- Caveats
- Integrating custom sensors takes hardware and FPGA work.
- First step
- Review the Holoscan Sensor Bridge documentation and pick a supported sensor board.
A medical robotics team needs one runtime from simulation to the deployed robot.
- Approach
- Use Isaac for Healthcare, which builds on Isaac Sim, Isaac Lab, Omniverse and Holoscan, and deploy the sensing and control pipeline with Holoscan.
- Role of NVIDIA Holoscan SDK
- Acts as the real-time runtime layer for sensing on the robot.
- Data, infrastructure and skills
- Robot hardware, simulation assets and robotics engineers.
- Type of benefit
- Shorter path from simulation to device
- Caveats
- Simulation results must be confirmed on the physical system.
- First step
- Explore the Isaac for Healthcare workflows listed in the NVIDIA-Medtech GitHub organization.
Who uses it
Moon Surgical · Medical devices (surgical robotics)
Moon Surgical: FDA-cleared ScoPilot camera control running on NVIDIA Holoscan
Moon Surgical's ScoPilot feature lets the laparoscope on its Maestro surgical robot follow a chosen instrument tip, running locally on NVIDIA Holoscan. It received FDA clearance in March 2025; Moon Surgical says Maestro had been used in more than 1,100 patients.
In production
Works with
Optional integration
- NVIDIA JetsonSupported on Jetson AGX Thor, Jetson AGX Orin and Orin Nano.
- NVIDIA DGXDGX Spark is a supported platform.
Complementary tools
- NVIDIA IsaacIsaac for Healthcare is a domain framework built on Isaac Sim, Isaac Lab, Omniverse and Holoscan.
- NVIDIA OmniverseIsaac for Healthcare combines Omniverse simulation with the Holoscan runtime.
Alternative approaches
- NVIDIA DeepStream SDKFor many-camera video analytics DeepStream is the closer fit; Holoscan targets low-latency sensor pipelines.
Optional integration for
- NVIDIA JetsonHoloscan SDK is a supported SDK on JetPack.
Relationship labels follow NVIDIA's documentation. "Alternative approaches" does not mean one is better: each profile says when it fits.
Getting started
Check your platform
Compare your hardware, OS and driver with the Holoscan installation page; pick the CUDA 13 or CUDA 12 build that matches.
Check: Your platform appears in the support table and nvidia-smi shows driver 535 or newer.
Install the SDK
Choose the NGC container (nvcr.io/nvidia/clara-holoscan/holoscan), the Debian package, or the Python wheel (pip install holoscan-cu13 or holoscan-cu12).
Check: Importing holoscan in Python, or building a C++ example, succeeds.
Run an example
Build and run one of the SDK examples, then a HoloHub reference application close to your domain.
Check: The example runs and shows output in Holoviz.
Add your model
Export your model for TensorRT, or install Torch or ONNX Runtime separately, and configure the inference operator.
Check: Your model's results appear in the pipeline at the expected rate.
Measure latency
Enable data flow tracking and record end-to-end and per-operator latency on the target device.
Check: Measured latency fits your time budget under realistic load.
Plan production deployment
For developer kits like IGX Orin, NVIDIA points to a separate OpenEmbedded/Yocto-based stack for production images.
Check: A deployment path and image build process are agreed.
Official resources
Could this technology help you?
Describe your project to the Solution Architect. It starts with NVIDIA Holoscan SDK as context but recommends independently, including when you do not need it.
Sources
Each statement above links to the source it comes from. Labels say who reported it.
- NVIDIA Holoscan SDK (developer page) (opens in a new tab)
- nvidia-holoscan/holoscan-sdk repository (opens in a new tab)
- Holoscan SDK User Guide: overview (opens in a new tab)
- Holoscan SDK User Guide: core concepts (opens in a new tab)
- HoloHub (opens in a new tab)
- Holoscan SDK User Guide: SDK installation (opens in a new tab)
- NVIDIA Healthcare and Life Sciences (opens in a new tab)
- Holoscan SDK releases on GitHub (opens in a new tab)
- London AI Centre and NVIDIA launch MONAI, a new AI framework for healthcare (opens in a new tab)
- Holoscan Sensor Bridge repository (opens in a new tab)
- NVIDIA-Medtech GitHub organization (opens in a new tab)
Thank you. Your correction was sent.
The editors check it against the sources. If you left an email address, they may reply about it.