The Power and Meaning of the NVIDIA BlueField DPU for AI Factories
As organizations accelerate innovation, developing AI-powered and life-changing products like self-driving cars or large language models (LLMs), efficient infrastructure is critical for scaling operations and staying competitive. Historically, data centers have focused on using central processing units (CPUs) for general-purpose computing and graphics processing units (GPUs) for intensive parallel processing tasks central to AI and machine learning. As AI models grow in scale and complexity, data centers have become the new unit of computing, pushing the boundaries of traditional cloud networks. To enable the transformation toward data center-scale computing, the data processing unit (DPU) has emerged as a third pillar in computing.
The rise of AI factories
Earlier in our AI factory series, F5 defined an AI factory as a massive storage, networking, and computing investment serving high-volume, high-performance training and inference requirements. Like traditional manufacturing plants, AI factories leverage pretrained AI models to transform raw data into intelligence.
What is a Data Processing Unit (DPU)?
A DPU is a programmable processor designed to handle vast data movement and processing via hardware acceleration at a network’s line rate. In late 2024, we announced BIG-IP Next for Kubernetes deployed on NVIDIA BlueField-3 DPUs. NVIDIA BlueField is an accelerated computing platform for data center infrastructure, purpose-built to power NVIDIA AI factories. While the CPU is responsible for general-purpose computing for computational applications and the GPU excels in accelerated computing tasks such as AI-related large-scale vector and matrix computations and graphics rendering, the NVIDIA BlueField DPU is often incorporated into a PCIe (peripheral component interconnect express) network interface card (NIC), which is responsible for network connectivity for the host or chassis of an AI cluster. In other words, the NIC now essentially has become a powerful processor, optimized for processing data as it moves in and out of the server. The BlueField DPU can also function as an inter-cluster networking device when multiple hosts or chassis are in a single AI cluster.
Unlocked power
By handling software-defined networking, storage management, and security services, BlueField DPUs reduce the computational burden on CPUs, allowing them to focus on the tasks at which they excel. This offloading capability is crucial for AI factories, where vast amounts of data must be processed and transferred rapidly to meet the demands of complex AI models and real-time inference tasks.
BlueField DPUs contribute significantly to energy efficiency and scalability within AI factories. As AI factories require massive computational resources, the efficient management of power and cooling becomes paramount. DPUs, with their specialized acceleration engines and high-performance network interfaces, ensure data is processed and transported with minimal latency and power consumption. This efficiency not only reduces operational costs but also enables AI factories to scale effectively. With BlueField DPUs, AI factories and large-scale infrastructure can achieve a balanced, high-performance, and high-efficiency infrastructure that supports the continuous innovation and deployment of AI technologies.
Where are BlueField DPUs deployed in AI factories?
When looking at F5’s AI Reference Architecture, DPUs are commonly deployed within the functional areas of RAG corpus management, fine-tuning, training, and inference services as well as the storage clusters supporting these functions. Additionally, DPUs are found in numerous applications where high-performance data throughput and power efficiency are required, including examples like DPUs supporting 5G radio access network (RAN) deployments.
The F5 AI Reference Architecture highlighting areas where DPUs are commonly deployed.
Offloading and accelerating application delivery and security to the DPU
The new requirement for efficient traffic management and robust security for AI factories represents an important shift focusing on data flow and fortified infrastructure to prevent security threats. F5 BIG-IP Next for Kubernetes deployed on NVIDIA BlueField-3 DPUs enables low-latency, high-throughput connectivity by offloading and accelerating data movement from the CPU to the DPUs. Additionally, it integrates comprehensive security features, such as firewalls, DDoS mitigation, WAF, API protection, and intrusion prevention, directly on the programmable NVIDIA BlueField-3 DPU. This allows you to create an architecture that isolates AI models and apps from threats, ensuring data integrity and sovereignty.
BIG-IP Next for Kubernetes supports multi-tenancy, enabling the hosting of multiple users and AI workloads on a single infrastructure, enabling network isolation. Managing a large-scale AI infrastructure becomes straightforward with BIG-IP Next for Kubernetes, as it provides a central point for managing networking, traffic management, security, and multi-tenant environments. This simplifies operations and reduces operational expenditure by offering detailed traffic data for network visibility and performance optimization. The integration between BIG-IP Next for Kubernetes and the NVIDIA BlueField-3 DPU helps ensure AI factories operate at their full potential while reducing tool sprawl and operational complexity.
Powered by F5
For enterprises investing in AI, ensuring their infrastructure is optimized and secure is non-negotiable. F5 BIG-IP Next for Kubernetes deployed on NVIDIA BlueField-3 DPUs is a strategic investment to deliver high performance, scalability, and security, maximizing the return on large-scale AI infrastructure. For organizations deploying GPUs and DPUs to support AI factory investments, contact F5 to learn how BIG-IP Next for Kubernetes can enhance your AI workloads.
F5’s focus on AI doesn’t stop here—explore how F5 secures and delivers AI apps everywhere.
Interested in learning more about AI factories? Explore the other AI factory series blog posts F5 has published to date: