​​Test System Hardware Acceleration with GPUs and NI LabVIEW​

This white paper explores the strategic use of GPUs as hardware accelerators to handle the increasing demands of computationally intensive test applications. As industries rely more on high-throughput processing to enhance testing efficiency, scalable, high-performance computing becomes essential. GPUs, with their massive parallelism and thousands of cores, offer a significant advantage over traditional CPUs, making them ideal for tasks like image and signal processing, deep learning inference, and matrix computations.

By integrating GPUs with NI LabVIEW and utilizing compatible hardware and toolkits, GPU acceleration can be effectively adopted and seamlessly incorporated into existing test systems. This approach facilitates an intuitive framework for engineers and system architects to maximize hardware acceleration in their workflows. Selecting the right hardware platform provides benefits such as increased processing speed, reduced latency, improved energy efficiency, and scalable performance for future applications.

GPUs excel in tasks that involve neural networks, image processing, and simulations because of their unrivaled parallel processing capabilities. However, the benefits of GPU parallelism become apparent only at large scales because offloading tasks to GPUs involves costly data transfers, synchronization for data consistency, and kernel launch delays. GPUs are not well-suited for small data sets characterized by low data volume and high variance in computational operations.

FPGAs provide a customizable architecture ideal for deterministic systems such as embedded systems and real-time signal processing. They offer optimized processing environments and low power solutions, crucial for AI models with power constraints. However, programming FPGAs for optimal performance involves significant overhead.

General purpose processors typically have a smaller number of computational cores, and each core has its own control unit associated with it, which allows for flexibility and control. CPUs have stronger emphasis on control units when compared to GPUs, which can enable single core context switching.

The NI modular platforms use CPUs for flexibility and complex control logic and FPGAs for determinism and precise control. And now, adding GPUs brings a third pillar of throughput-oriented acceleration. 

LabVIEW offers powerful native capabilities and a broad range of built-in functionalities for diverse applications; this is further enhanced by third-party toolkits that allow easy integration of GPU acceleration. These toolkits are intuitive and resemble native LabVIEW, reducing the learning curve for developers. They add specialized functions for GPU use, helping users tap into GPU power without having to develop all the components from scratch. This section will discuss three significant third-party toolkits in the NI ecosystem that facilitate GPU acceleration for various applications, focusing on their features and effective architectures to meet high performance standards.

The Ngene CuLab Toolkit

Modern engineering and scientific applications require significant computational power, particularly in areas such as RF signal processing, real-time monitoring and analytics, scientific computing and modeling, and industrial machine vision. While GPUs are ideal for these tasks due to their parallelism, LabVIEW has lacked native GPU acceleration support, posing a challenge for engineers unfamiliar with CUDA or C++.

The Ngene CuLab toolkit addresses this gap by offering a high-performance GPU acceleration solution directly integrated with LabVIEW. It features a LabVIEW-like API, enabling engineers to accelerate computations without complex coding. CuLab simplifies integration, offers comprehensive GPU-accelerated functions, and exhibits superior performance, averaging 6X faster than Python frameworks like CuPy.

CuLab offers several key benefits. It seamlessly integrates with a LabVIEW-like API and provides a more than 100X improvement¹ for data processing tasks, when compared to base LabVIEW on a CPU. The platform includes more than 150 GPU-accelerated functions to efficiently handle multidimensional arrays and all numeric data types.

Figure 1. This program showcases the similarity between native LabVIEW and CuLab implementations.

The functionality of CuLab is extensive and mimics native LabVIEW. The comprehensive list of functions includes vector and matrix operations, linear algebra, signal processing, computer vision, math functions, and logical operations.

The Graiphic SOTA Toolkit

The SOTA ecosystem for LabVIEW by Graiphic introduces enriched GPU acceleration capabilities as an alternative to deep learning frameworks like PyTorch and TensorFlow. It transforms LabVIEW into a universal graph editor and orchestrator with expanded ONNX support for training and computation across diverse hardware. Offering both high-level and low-level abstractions, it caters to engineers needing straightforward deployment and researchers seeking detailed control, addressing a wide spectrum of deep learning needs.

Figure 2. The accelerator palette aims for full coverage of standard LabVIEW operations. It effectively manages dynamic orchestration and provides excellent flexibility to support different execution providers. A public reference maps node coverage across multiple backends.

SOTA rivals existing frameworks with a comprehensive feature set, generalized by its Accelerator toolkit which broadens the LabVIEW function palette for orchestrating computational graphs via ONNX. 

The key components of the SOTA ecosystem include the Deep Learning toolkit for training and inference, the Computer Vision toolkit for image processing, the GenAI toolkit for transformer-based models, and the previously mentioned Accelerator Toolkit. This environment’s toolkits excel in computer vision, deep learning, GenAI, and reinforcement learning, with seamless ONNX integration optimizing performance across hardware platforms. It democratizes AI development through visual data flow programming, allowing users without programming skills to create complex workflows.

SOTA optimizes data labeling, AI model training, tuning, and deployment across heterogeneous hardware.

Natan Biesmans LabVIEW G²CPU High-Performance Compute Toolkit

G²CPU offers a high-performance computing toolkit for LabVIEW, allowing platform-agnostic code execution across CPUs, NVIDIA CUDA GPUs, OpenCL devices, and the entire NI ecosystem. Built on the open-source ArrayFire library, it simplifies backend switching for long-term stability.

Key Benefits

• Unified backend abstraction—Write code once and deploy across CPUs, CUDA, and OpenCL without modification, enhancing flexibility. This toolkit can be utilized across the NI ecosystem which includes devices such as Windows and Linux devices, LabVIEW Real Time Module, PXI, and industrial computers.
• Seamless integration—Functions mirror native LabVIEW operations, easing incorporation with minimal learning curve. Debugging is supported with probes that can inspect data across different computing backends directly within the LabVIEW environment.
• High code flexibility—Implement even the most complex algorithms on GPUs while staying in the native LabVIEW IDE.
• High performance—Achieves up to 200X speed improvements² for intensive tasks, with GPU I/O speeds at 7 GB/s on NI PXI and more than 22 GB/s on modern systems. It supports sub-1 ms loops in the NI LabVIEW Real-Time Module for time-sensitive applications. 

Figure 3. This program showcases a Linear Fit algorithm completely written in G²CPU. Notice the similarity to native LabVIEW.

The software toolkit offers a comprehensive suite of high-powered functions for various tasks including mathematics, signal processing, machine vision, UI acceleration, and many more—all of which can be used across any type of GPU and operating system. It features high-speed data transfer capabilities, allowing for seamless integration with devices such as the NI PXI Vector Signal Transceiver (VST), NI FPGA-enabled hardware and NI frame grabbers, along with many others. Made possible using efficient zero-copy data transfers between devices and GPUs, G²CPU is a highly flexible and performant toolkit that can be extended with custom GPU kernels and third-party toolkits, allowing for seamless integration with toolkits like NVENC, LibTorch, OpenCV, and many more. G²CPU is an ideal solution for developers requiring exceptional performance with untethered development in any domain.

This section covers connecting to a GPU server as well as PXI Express to MXI Express configurations.

Network Connection to GPU Server

Connecting to a GPU server can be efficiently achieved through network interface cards (NICs) and smart NICs, which offer advanced capabilities for optimized network performance. Smart NICs integrate CPU and GPU components to support demanding computational and parallel processing tasks. They intelligently offload network processing to reduce overhead, enhancing dataflow efficiency to the GPU server. Smart NICs can improve data handling across networks, supporting distributed computing applications with up to 25 GB/s bandwidth for more effective use of networked GPU resources. Other NI hardware, including NI CompactRIO and NI CompactDAQ systems, can use Ethernet or USB network connections to interface with and use GPU server resources as well. 

Figure 4. A PXI chassis is networked to an external GPU server for high-speed, accelerated computing via NIC or smart NIC.

PXI Express to MXI Express

PXIe-to-MXIe configurations are beneficial for setups needing external GPU capabilities, especially in high-performance computing. By connecting PXIe chassis with MXIe interfaces, users can access powerful external GPUs, enhancing computational capacity beyond local hardware limits. This setup is ideal for constrained spaces and scenarios requiring multiple GPUs across systems for tasks like large-scale data processing, AI development, or simulations. The Gen 3 x16 MXIe interface supports data transfer rates of 16 GB/s, reducing bottlenecks and allowing demanding applications to efficiently utilize external GPU or GPU server resources.

Figure 5. A PXIe chassis connects to a GPU server via MXIe expansion, enabling high-bandwidth data transfer for accelerated processing and advanced computation.

RADX Technologies Inc. offers commercial off-the-shelf (COTS) PXI Express GPU modules designed for the NI ecosystem yet distinct from NI offerings. Based on NVIDIA Turing, Ampere, and Ada GPUs, these modules integrate seamlessly into PXI Express chassis like the NI PXIe-1092 or NI PXIe-1095, enhancing compute power for test and measurement environments. RADX PXIe GPUs support numerous applications, such as data acquisition and machine learning, using tools like LabVIEW and Python, without needing separate GPU servers.

The RADX Technologies in-chassis GPUs offer 1.7 to 30.3 FP32 TFLOPS with 4 GB to 24 GB memory. However, these setups can face data bottlenecks caused by the lack of peer-to-peer communication between GPUs, except when using Linux. Techniques such as zero copy are required to mitigate these limitations and fully leverage GPU capabilities and alleviate those bottlenecks. RADX GPUs are qualified for NI systems and are often paired with G²CPU for GPU-accelerated LabVIEW signal processing.

Figure 6. RADX PXIe-compatible GPU modules and offerings

In conclusion, GPUs are increasingly adopted by test engineers for faster testing because of their exceptional performance and programmability. Their massive parallelism and scalable compute power make them ideal for data-heavy applications like signal processing, image processing, simulations, and model-based control. However, due to their data demands, GPUs remain a niche addition for suitable workflows.

The integration of GPU acceleration into the NI ecosystem, using PXI Express-supported partner GPUs and LabVIEW toolkits, democratizes GPU use for test engineers by providing a flexible, high-performance complement to existing workflows. This integration within familiar environments lowers the entry barrier, enabling faster prototyping, efficient data throughput, and better system scalability.

Understanding how to effectively use GPUs offers test and measurement organizations a competitive advantage in designing systems that are future-proof. Recognizing GPU performance gains, leveraging appropriate tools, and identifying computationally intensive tasks are key steps.