Accelerating Data Center AI with the NVIDIA Converged Accelerator Developer Kit

Discuss (0)

The modern data center is becoming increasingly difficult to manage. There are billions of possible connection paths between applications and petabytes of log data. Static rules are insufficient to enforce security policies for dynamic microservices, and the sheer magnitude of log data is impossible for any human to analyze.

AI provides the only path to the secure and self-managed data center of the future.

The NVIDIA converged accelerator is the world’s first AI-enhanced DPU. It combines the computational power of GPUs with the network acceleration and security benefits of DPUs, creating a single platform for AI-enhanced data center management. Converged accelerators can apply AI-generated rules to every packet in the data center network, creating new possibilities for real-time security and management.

Image shows NVIDIA's new converged accelerator which combines a Bluefield2 DPU and Ampere GPU.
Figure 1. In standard mode, the BlueField-2 DPU and GPU are connected by a dedicated PCIe Gen4 switch for full bandwidth outside of the host PCIe system.

At NVIDIA GTC, we are introducing two new converged accelerators. The A100X combines an A100 Tensor Core GPU with a NVIDIA BlueField-2 data processing unit on a single module. The A30X combines an A30 Tensor Core GPU with the same BlueField-2 DPU. The converged cards have the unique ability to extend the BlueField-2 capabilities of offloading, isolating, and accelerating the network to include AI inference and training. 

Both accelerators feature an integrated PCIe switch between the DPU and GPU. The integrated switch eliminates contention for host resources, enabling line-rate GPUDirect RDMA performance. The integrated switch also improves security by isolating data movement between the GPU and NIC.

AI enhanced DPU

The converged accelerators support two modes of operation:

  • Standard–The BlueField-2 DPU and the GPU operate separately.
  • BlueField-X–The PCI switch is reconfigured so the GPU is dedicated to the DPU and no longer visible to the host system.

In BlueField-X mode, the GPU is dedicated exclusively to the operating system running on the DPU. BlueField-X mode creates a new class of accelerator never before seen in the industry: a GPU-accelerated DPU.

Image shows that in Bluefield-X mode, the CPU in the host server connects to the Converged Accelerator. The Converged Accelerator's PCIe switch is connected to the CPU and DPU. While the GPU is only connected to the PCIe switch and DPU.
Figure 2. In BlueField-X mode the x86 host only sees the BlueField-2 DPU, allowing the DPU to run AI workloads on the network data.

In BlueField-X mode, the GPU can run AI models on the data flowing through the DPU as a “bump in the wire.” There is no performance overhead and no compromise on security.  The AI model is fully accelerated without consuming host resources. 

BlueField-X unlocks novel use cases for cybersecurity, data center management, and I/O acceleration. For example, the Morpheus Cybersecurity framework uses machine learning to take action on security threats that were previously impossible to identify. Morpheus uses DPUs to harvest telemetry from every server in the data center and send it to GPU-equipped servers for analysis.

With BlueField-X, the AI models can run locally on the converged accelerator in each server. This allows Morpheus to analyze more data, faster, while simultaneously eliminating costly data movement and reducing the attack surface for malicious actors. Malware detection, data exfiltration prevention, and dynamic firewall rule creation are Morpheus use cases enabled by BlueField-X.

The Morpheus example only scratches the surface of what is possible with BlueField-X. Our customers routinely share ideas that we had not yet considered. To enable more creative exploration of AI-enabled networks, we are introducing the NVIDIA Converged Accelerator Developer Kit

With this developer kit, we provide early access to A30X accelerators for select customers and partners building the next generation of accelerated AI network applications. Discover new applications for BlueField-X in edge computing or data center management. Some ideas to help get you started include:

  • Transparent video preprocessing–Bump-in-the-wire video preprocessing (decryption, interlacing, format conversion, etc) to improve IVA throughput and camera density.
  • Small-cell RU solution–RAN signal processing on a converged accelerator to increase subscriber density and throughput on a commodity gNodeB server.
  • Computational storage–Bump-in-the-wire storage encryption, indexing, and hashing to offload costly CPU cycles from a storage host preparing data for long-term storage.
  • Cheating detection–Detect malicious gameplay/cheating in a streaming gaming service

Get started with the NVIDIA Converged Accelerator Developer Kit

The NVIDIA Converged Accelerator Developer Kit contains sample applications that combine CUDA and NVIDIA DOCA, and documentation to help you install, configure your new converged accelerator. Most importantly, we provide access to an A30X and usage support in exchange for feedback. 

To get started, simply register your interest on the NVIDIA Converged Accelerator Developer Kit webpage. If approved, we will contact you once the hardware is ready to ship and you can start building the next generation of accelerated applications.

We hope that you share our excitement for building a new class of real-time AI applications for data center management and edge computing. Let the discovery begin.