Nvidia Wants To Bring New BlueField DPUs To ‘Every Server’

The GPU juggernaut is introducing a new type of processor backed by VMware and Red Hat that offloads hypervisor functionality from the CPU to improve data center economics and introduce new security and management capabilities. ‘We believe that the DPU belongs in every server going forward, regardless of the application workload running there,’ Nvidia’s Manuvir Das says.


Nvidia is expanding its data center ambitions with a new type of processor that will offload “critical” networking, storage and security workloads from the CPU and enable new security and hypervisor capabilities, which is why the chipmaker believes they should go into “every server going forward.”

The Santa Clara, Calif.-based company revealed the BlueField family of data processing units, or DPUs, at its fall GTC 2020 conference Monday, saying that the PCIe-based products feature DOCA, a new “data-center-infrastructure-on-a-chip architecture,” and rely on a mix of Arm processor cores, ConnectX-6 SmartNIC technology from Nvidia’s Mellanox Technologies acquisition and, in the case of a few products, Nvidia’s data center GPU architecture.

[Related: Omni-Path Spin-Out Aims To Help HPC Partners Keep Nvidia In Check]

Sponsored post

While the DPUs have been in development before Nvidia’s plan to acquire British chip designer Arm took shape, they represent another step toward the company’s goal of becoming a “data center-scale company,” as Nvidia CEO Jensen Huang recently said. In the company’s most recent earnings report, data center revenue eclipsed gaming sales for the first time in its history.

“The data center has become the new unit of computing,” Huang said of the newly revealed DPUs in a statement. “DPUs are an essential element of modern and secure accelerated data centers in which CPUs, GPUs and DPUs are able to combine into a single computing unit that’s fully programmable, AI-enabled and can deliver levels of security and compute power not previously possible.”

The first product in the lineup, BlueField-2, is sampling with customers now and will be available in 2021 in servers from Asus, Atos, Dell Technologies, Fujitsu, Gigabyte, H3C, Inspur, Lenovo, Quanta and Supermicro. The second product, BlueField-2X, features an Nvidia Ampere GPU for AI-accelerated security, network and storage tasks and will also be available next year.

Beyond OEM support, the products also have buy-in from VMware, which announced at last week’s VMworld 2020 that as part of an expanded partnership it will support Nvidia DPUs as part of its Project Monterey, a new kind of hybrid cloud architecture that will deliver purpose-built performance for AI, machine learning, high-throughput and data-hungry applications. Linux vendors Red Hat and Canonical have also pledged support as has security vendor Check Point Software Technologies.

With a single BlueField-2, the DPU can deliver the same level of performance for software-defined networking, software-defined security, software-defined storage and infrastructure management that 125 CPU cores can, meaning that CPU cores will have headroom to deliver additional performance for enterprise applications, according to Nvidia.

Manuvir Das, Nvidia’s head of enterprise computing, said DPUs should be thought of as SmartNICs with additional acceleration capabilities that replace standard network interface cards, or NICs, and substantially improve the economics of data centers.

“What we expect to see is significant [total cost of ownership] benefits for the customer through the introduction of the DPU because it will really free up the server to do a lot more application workload[s], which will reduce the TCO profile,” he said. “There are studies, for example, that have been published from public cloud vendors talking about more than 30 percent of the work of the CPU being offloaded, which of course leads to a TCO benefit there.”

Beyond the CPU-offload benefits, Nvidia’s DPUs bring additional security to data centers because the Arm cores embedded within them serve as a second security plane separate from the host CPU, providing a new level of isolation between security services and applications.

“If I’m running the security plane on the host CPU itself, as I do today, and if the host CPU is compromised because one of the applications running there is compromised, then my security subsystem itself is compromised and then all bets are off,” Das said. “Whereas by removing that security subsystem, that functionality, the distributed firewall onto this other security domain, which is the Arm CPU running on the DPU, I now have a level of protection because if an application compromises the host CPU, it does not compromise the security plane and the entire data center.”

The additional benefit is specifically aimed at virtualization and will allow users to manage bare-metal servers with a hypervisor by offloading it to the CPU.

“In this model by shifting the hypervisor functionality to the DPU, that hypervisor is now capable of managing bare-metal servers as well, with applications running bare metal. And so it really brings those two worlds together for the first time, where you can manage both virtualized and bare-metal servers in the same way,” said Das.

The BlueField-2 DPU is a PCIe 4.0 card that comes with eight 64-bit A72 Arm processor cores, two VLIW acceleration engines and the Mellanox ConnectX-6 Dx SmartNIC. It also comes with root of trust, key management, a cryptography engine and two 100-Gbps Ethernet and InfiniBand ports.

With Nvidia’s BlueField-2X, the company has incorporated a GPU from its new Ampere architecture for AI use cases such as network anomaly detection and response, real-time network traffic analytics and real-time analytics of uploaded video content. Das said the GPU is connected to the DPU on the system-on-chip through an internal PCIe 4.0 connection. He added that the GPU is not the same as Nvidia’s A100, though he declined to provide more details, except that it can deliver 60 tera operations per second of AI performance and is not visible to the main part of the server.

Beyond Nvidia’s plans to launch the BlueField-2 and BlueField-2X in 2021, the company revealed a DPU product road map that includes the BlueField-3 and BlueField-3X in 2022 as well as the BlueField-4, which will come out in 2023 and, for the first time in the product line, integrate the GPU and Arm cores in the silicon and provide 600 times better performance than the BlueField-2, according to the company.

“We believe that the DPU belongs in every server going forward, regardless of the application workload running there,” Das said.

Matthew DuBell, a consulting solutions architect for the Business and Analytics Advisors group at World Wide Technology, a top Nvidia partner in St. Louis, told CRN last week that new SmartNIC technologies like Nvidia’s DPUs will become increasingly important as organizations need to better insight into network performance and resolve bottlenecks. That’s something he has encountered, for instance, in deploying multiple Nvidia DGX systems.

“I have a lot of clients buy one or two DGX-1s or [DGX-2s], and they did interesting, unique things with that when they really want to go to scale and hook up five, 10, 20 of them. That’s where the network performance and insight became critical,” he said.