When to Use the PVA?

⟵ ⟶

RidgeRun NVIDIA PVA Development

Table of Contents [Sticky]

• Introduction
  • What is the PVA?
  • When to Use the PVA?
  • Development Workflow
  • Services
• Getting Access to the PVA
  • Vision Programming Interface(VPI)
  • GStreamer VPI Plug-In
  • Custom Applications
• Algorithms
• Contact Us
• Sponsor your Favorite Feature

When to Use the PVA?

The PVA is designed for low-power, fixed-function vision processing using a Vector Processing Unit approach, specially designed for Computer Vision workloads. Integrating the PVA into your system can be highly beneficial in the following scenarios:

• Offloading Preprocessing and Postprocessing Tasks in AI Pipelines

In typical AI inference pipelines, preprocessing and postprocessing stages often involve computationally intensive CV operations such as image cropping, resizing, and image enhancement. By offloading these tasks to the PVA, the load on primary processors like the GPU and CPU is reduced, allowing them to focus on core AI inference tasks. This offloading enhances overall system efficiency and performance.

• Executing Standalone Computer Vision Algorithms

The PVA specializes in handling pure CV. Tasks like feature extraction and image processing algorithms can be executed directly on the PVA. This capability is advantageous in applications requiring real-time processing with low power consumption, such as autonomous vehicles and robotics.

• Performing Mathematical Computations

Due to its vector SIMD VLIW DSP architecture, the PVA is well-suited for mathematical computations. Offloading these computations to the PVA can free up the GPU and CPU for other critical tasks, optimizing the overall computational workload.

It is highly recommended for algorithms that match the following characteristics:

• Streamlined Data Flow: Algorithms with linear, well-defined stages and minimal branching.
• Vector-friendly Operations: Algorithms with high parallelism capacity, like pixel-wise operations.
• Low Memory Footprint: Fits within the PVA’s internal memory and minimizes external memory access.
• Deterministic Execution: Algorithms with minimal to no branching.

PVA Block Diagram (Simplified). Adapted from developer.nvidia.com

Use the PVA when your application demands:

• Power Efficiency: Ideal for edge devices and battery-powered systems.
• Real-Time Processing: Deterministic behavior enables consistent low-latency performance.
• Supported Algorithms: When the workload involves operations like convolution, optical flow, remapping, or image scaling.
• CPU/GPU Offloading: Reduces contention on primary processors by shifting processing to the PVA.
• Vector-friendly Algorithms: Vector-friendly algorithms are best-suited for PVA.

Some example applications include:

• Object detection pipelines (pre-processing stages)
• Image pyramid generation for scale-invariant feature detection
• Optical flow for motion estimation
• Filtering operations in multi-camera systems

For more complex algorithms, it is recommended to use the GPU or the CPU. Some characteristics are:

• Pointer chasing algorithms: algorithms with aggressive and scattered memory look-ups may present a lower performance on PVA, given the size of the vector memory (VMEM).
• Floating-point intensive algorithms: for high precisions, like Float64, it is better to use the GPU.

⟵ ⟶