NVIDIA PVA SDK 2.10 helps Jetson developers accelerate sensor processing using the PVA with a new compiler, host emulator, and AI agent skills to build, test, and debug PVA code.
Get started 👇 https://x.com/i/article/2115552972788916224
Accelerate Multimodal Sensing for Physical AI with NVIDIA PVA SDK 2.10
The NVIDIA Programmable Vision Accelerator (PVA) is a dedicated, power-efficient engine on Jetson for multimodal sensor processing for Physical AI. PVA SDK 2.10, now generally available to all NVIDIA Jetson developers, brings PVA development into the agentic era, putting PVA expertise within reach so every developer can program and iterate at agentic speed.
At its foundation is a completely rebuilt, NVIDIA-engineered compiler based on the LLVM framework, delivering modern, high-performance code generation for PVA. It is paired with the newly built PVA emulator, which enables testing PVA-native code on a host machine. Together, they let developers compile, run, and verify code with or without target hardware.
PVA Solutions is now available on GitHub, offering optimized applications for agents to build on. PVA Agent Skills tie it all together, guiding agents through each step of PVA development.
Available on supported NVIDIA Jetson Orin and Thor modules, PVA enables power-efficient acceleration of multimodal sensor processing. PVA SDK 2.10 and its supporting resources help developers use this dedicated engine while preserving GPU capacity for demanding AI workloads in physical AI applications.
More AI Through Efficient Sensor Processing
Physical AI systems bring sensing, perception, and reasoning to the edge. These systems must continuously process data from cameras, radar, LiDAR, and other sensors while running increasingly capable AI reasoning models—all within tight power, thermal, and latency budgets.
PVA accelerates image processing, traditional computer vision, radar and LiDAR processing, and other signal-processing workloads, freeing GPU capacity for demanding AI tasks.
What’s New in PVA SDK 2.10
- Compile C and C++ code with NVVPU, a new NVIDIA-built compiler based on LLVM.
- Develop and test without target hardware using the native host emulator.
- Debug through supported Debug Adapter Protocol clients, including Visual Studio Code, and profile application performance.
- Build on open-source PVA Solutions and use PVA Agent Skills to generate, build, test, and debug custom operators.
The release also adds a native host emulator, support for debug adapter protocol clients including Visual Studio Code, and profiling utilities. Developers can test PVA code without target hardware, debug applications in the emulator or on a Jetson system, and use NVIDIA Nsight™ Systems to identify performance bottlenecks.
PVA Solutions 0.6.0
PVA Solutions 0.6.0 is now available on GitHub under the Apache License 2.0. The release provides source code, samples, and optimized PVA operators for image processing, computer vision, radar, LiDAR, and AI pipelines. New capabilities include a YOLOv5 object-detection pipeline that runs preprocessing and postprocessing on PVA alongside GPU inference, new radar-processing operators, and additional image-processing and linear-algebra operators. Detailed documentation explains the optimization techniques used in each reference implementation, helping developers ramp up quickly, customize the code for their applications, and accelerate delivery of high-performance PVA applications.
PVA Agent Skills
PVA Agent Skills help developers use AI coding agents to install the PVA SDK, generate, build and test custom PVA operators. Drawing on PVA documentation and programming references, the skills guide agents through operator implementation, build and test workflows. Developers can use these skills to create new operators or debug and fix issues in existing implementations, reducing the specialized PVA knowledge needed to get to production operators.
New to AI-assisted development on Jetson? Follow the Jetson AI Lab tutorial to prepare your Jetson for a coding-agent workflow.
Where Is PVA Used Today?
Rebotnix makes smart city cameras with agentic reasoning capabilities for faster city-level decision-making. The company uses PVA alongside other NVIDIA Jetson hardware accelerators to process dual thermal-camera streams while keeping GPU resources available for AI. This approach supports Rebotnix moving from Jetson Orin NX 16GB to the 8GB configuration while maintaining performance and reducing hardware cost. Its BLADE platform brings vision and thermal AI to autonomous systems.
RidgeRun helps customers evaluate their current workload, identify bottlenecks, plan offloading to the PVA, and implement PVA applications for NVIDIA Jetson. RidgeRun helped a customer implement low-power image processing by offloading workloads to PVA. This approach keeps GPU resources available for AI inference while supporting the energy-efficiency needs of battery-powered systems, where lower processing power can enable longer operation.
Join NVIDIA partners RidgeRun and Connect Tech for the livestream, How to Maximize Memory Efficiency and Reduce Costs on VIDIA Jetson. Explore model quantization, efficient inference runtimes, and NVIDIA Jetson Agent Skills to fit larger AI models into less memory.
How to Get Started with PVA SDK 2.10
- Download: Visit the PVA Developer page to access PVA SDK 2.10.
- Learn: Follow the documentation and tutorials to build your first application.
- Build: Explore PVA Solutions on GitHub for optimized operators, sample pipelines, and optimization guidance.
- Develop with AI: Use PVA Skills from the NVIDIA Skills repository to generate, test, and debug operators with your AI coding assistant.