Skip to content

What's New in Holoscan 4.4: New Docs, Native UYVY Capture, and a Stronger Holohub Ecosystem

Written July 6, 2026

Holoscan SDK 4.4 follows the focused, developer-facing cadence of Holoscan 4.3 with a release that sharpens the day-to-day workflow: a modernized documentation experience, first-class support for UYVY camera input, clearer Holoviz diagnostics, and ecosystem updates that make Holohub easier to extend and deploy against.

You can grab it today as a Docker container (v4.4.0-cuda13, v4.4.0-cuda12-dgpu, v4.4.0-cuda12-igpu), a Python wheel for CUDA 12 (holoscan, pip install holoscan==4.4.0) or CUDA 13 (holoscan-cu13, pip install holoscan-cu13==4.4.0), or as Debian packages (4.4.0.0-1).

Documentation, Rebuilt for How You Work

The Holoscan SDK User Guide has migrated to a Fern-based build system and is now published at docs.nvidia.com/holoscan/sdk-user-guide. The new site adds an AI-powered "Ask a question" assistant, a "View as Markdown" option on every page, and direct integrations with coding assistants — including "Connect to Cursor" — from the "More actions" menu on each page.

Two new guides round out the developer experience:

  • Holoscan CLI Guide — documents the holoscan-cli PyPI package, an all-in-one command-line tool to discover, build, run, test, package, and lint Holoscan projects from source.
  • Optimizing JetPack — guidance on installing the real-time kernel and running with a minimal headless BSP to achieve lower, more predictable latency on Jetson developer kits.

The offline PDF user guide is no longer generated; the online guide is the canonical reference going forward.

Native UYVY Input for CSI and MIPI Cameras

FormatConverterOp and FormatConverterGpuResidentOp now accept "uyvy" as an in_dtype, in addition to the existing "yuyv" mode. The new path handles YUV 4:2:2 packed input with U/Cb-first byte order — the format many CSI and MIPI camera sensors produce — and converts it to RGB888 via NPP's nppiCbYCr422ToRGB.

If you've been byte-reordering UYVY frames before format conversion, you can drop that preprocessing step and feed the sensor output directly into the converter.

format_converter:
  in_dtype: "uyvy"
  out_dtype: "rgb888"

Holoviz Diagnostics Keep Improving

Holoviz now surfaces more actionable errors when Vulkan instance creation fails. Error messages identify the specific layers and extensions that could not be loaded, which makes bring-up on new or unusual hardware configurations considerably faster.

Holoviz also emits a warning when an InputSpec with a higher bit-depth pixel format (for example, 10-bit) is rendered to a lower bit-depth swapchain surface (for example, 8-bit). The warning calls out the format mismatch and the resulting bit-depth quantization — a follow-on to the startup warnings introduced in 4.3, now covering the render path as well.

Holoscan Sensor Bridge 2.6.0 GA

The Holoscan Sensor Bridge module is updated to 2.6.0 GA for source builds and release containers. Sensor Bridge is the FPGA-based ingest path for JESD-compliant sensors on Holoscan platforms, and this release keeps the SDK aligned with the latest GA capabilities for GPU-direct sensor I/O.

Safer PoseTree Usage

All PoseTree interface methods (such as get, set, remove, and has) are now marked [[nodiscard]]. Callers that previously discarded these return values will receive compiler warnings or errors. Update call sites to inspect the returned expected_t value, or explicitly discard it when intentional:

std::ignore = pose_tree.set(frame_id, pose);

What's New on Holohub

The reference catalog at Holohub saw meaningful ecosystem work alongside the SDK release:

  • holoscan-cli migration — ./holohub is now a thin wrapper that bootstraps and delegates to the standalone holoscan-cli package. The same commands still work; the CLI itself is now versioned and installable independently via pip.
  • DAQIRI networking — networking examples and operators (adv_networking_bench, basic_networking_ping, psd_pipeline, network_radar_pipeline, simple_pdw_pipeline, velodyne_lidar_app) now use DAQIRI instead of the retired Advanced Networking Operator (ANO) surfaces.
  • Holoscan Sensor Bridge module — Sensor Bridge is now a curated Holohub module. Any project can fetch and build HSB sources with --build-with=<operator_name>, and the AI Surgical Video Workflow has been updated to use it.
  • Deltacast external module — Deltacast operators and demos are now maintained by DELTACAST.TV as an external Holoscan module, with one-line build support preserved in the Endoscopy Tool Tracking demo.
  • AJA setup guide in Holohub — AJA capture-card setup documentation has moved from the SDK User Guide into Holohub, centralizing hardware bring-up guidance alongside the AJA Source operator. This aligns with the SDK 4.4 breaking change that removes aja_setup.md from the user guide.
  • Holoscan GStreamer module packaging — the GStreamer module now builds Python wheels for multiple Python versions and imports from the holoscan namespace, making it easier to integrate GStreamer encoding and recording into Holoscan pipelines.

The Holohub module template also received a round of improvements — CMake export interfaces, Python package staging, and split C++/Python example pipelines — lowering the bar for partners and community contributors shipping reusable Holoscan modules.

Getting Started

The fastest way to try 4.4 is pip install holoscan==4.4.0 (CUDA 12) or pip install holoscan-cu13==4.4.0 (CUDA 13), or pull the container from NGC. Full details, including platform compatibility and the complete release notes, are in the Holoscan SDK User Guide and the v4.4.0 release notes. Join the conversation on the NVIDIA Developer Forum or Discord, and explore the source on GitHub. As always, contributions to Holohub are welcome — if you've built something on Holoscan, we'd love to see it land in the reference catalog.