Skip to content

HOLOSCAN REFERENCE DESIGNS

From raw signals
to agentic devices

Turnkey hardware, firmware, and software solutions built with sensor and IC partners. Start with a working Holoscan system, then make it your own—with GPU algorithms, AI models, and agents that turn measurements into decisions.

Apache 2.0 Licensed Open SourceInspect the source. Modify the design. Build your product.

Choose your starting point

Ultrasound probe resting on an NVIDIA DGX Spark

Software-defined ultrasound

Available now

Stream raw channel data from TI acquisition electronics over HSB-based Ethernet into GPU memory. Build your beamformer, train on captured RF data, and steer the next transmit from software.

  • Raw RF acquisition
  • GPU beamforming
  • Active transmit control
Explore the ultrasound design
Endoscopy video with GPU-powered tool-tracking overlays

Software-defined endoscopy

Early access under NDA · General availability in 2027

An FPGA design, Holoscan operators, and a sample pipeline for real-time display. Yuan is demonstrating a PCIe-based FPGA capture card built on the design at GTC Berlin.

  • Camera acquisition
  • GPU imaging + AI
  • Live visualization
Explore the endoscopy design

Endoscopy example image courtesy of CAMMA, IHU Strasbourg, and the University of Strasbourg

Minimal FPGA logic
Maximum software freedom

The FPGA provides the essential glue between sensors and GPUs: interfaces, capture, timing, and transport. Move the application’s signal processing, reconstruction, AI inference, and agent reasoning onto the GPU with CUDA and Holoscan.

  1. Sensor + front end

    Acquire raw measurements
    Drive the next acquisition

  2. FPGA I/O

    Capture, timing, transport
    Minimal sensor-to-GPU logic

  3. GPU + Holoscan

    CUDA processing, AI, agents
    Interpret and decide

Raw signals into the GPU. Decisions back to the sensor.

Raw to insights

Give your algorithms access to the original measurements. Reconstruct an image, extract a signal, or run a model directly on raw data in the same GPU pipeline.

Change software, explore more

Replace a CUDA kernel, tune a model, or develop with coding agents. Test against recorded data, then bring the same algorithm into the live instrument.

Extend every layer

Start with the demonstration pipeline and its source. Adapt sensor integration, firmware, and operators to the device you want to build.

Let intelligence decide
what to sense next

An agentic device can act on what it learns. Holoscan’s low-latency execution and deterministic control capabilities let you connect perception to the next acquisition.

For active-steering sensors such as ultrasound and radar, use GPU results to update transmit settings, focus on a region of interest, or change the scan sequence. The sensor becomes part of the feedback loop.

Explore GPU-in-the-loop ultrasound

One runtime for perception and control

Holoscan brings raw-data processing and control into one application graph. CUDA Graphs reduce dispatch overhead for eligible GPU stages. Pinned CPU cores control thread placement, while CUDA green contexts let you assign GPU resources on supported targets.

Explore the Holoscan runtime

Open designs.
Open data.
Open models.

Reference designs connect our open data and open models initiative to real instruments. Collect the measurements that train new models, then bring those models back into the live acquisition loop.

Build datasets for your own research and products, or contribute captures to our open data initiatives. Choose what you share and help expand the data available to the community.

  1. Collect raw data

    Capture sensor measurements with acquisition settings and context. Keep the information you need to revisit reconstruction, test algorithms, and train new models.

  2. Train and share models

    Use your own datasets or openly available captures to develop and evaluate models. Contribute data and models to help others reproduce results and build on your work.

  3. Bring insights back to the device

    Deploy models in the Holoscan pipeline. Use their results to improve interpretation, guide acquisition, and collect the next set of measurements.

Your next product
starts further ahead

Get hardware, firmware, and software working together from the start. These partner-built reference designs give you a launchpad to accelerate the path from first prototype to market.

Spend your development time on the algorithms, agents, and capabilities that make your product different. Our reference designs are built for Apache 2.0 licensed reuse. Ultrasound is available now; endoscopy is in NDA early access ahead of general availability in 2027.

Choose a reference design Build with the community