Nvidia announced a new line of data processing units (DPUs) and a data center architecture it's calling DOCA, during the GTC 2020 virtual event today.

"The data center has become the new unit of computing," said Nvidia CEO Jensen Huang. "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 and can deliver levels of security and compute power not previously possible."

According to Manuvir Das, head of enterprise computing at Nvidia, the idea behind the DPU is to shift functionality running in the hypervisor from the CPU to the DPU. This, he says, frees up the CPU to run applications and has the added benefit of isolating a compromised host from the security layer in the event of a breach.

The DPUs themselves are built on technology acquired from Mellanox, which Nvidia purchased in 2019 for $6.9 billion. In fact, Mellanox announced BlueField-2 — Nvidia's base-model DPU — back in February before the deal closed. Since then, the BlueField-2 product line has evolved dramatically with Nvidia integrating an Ampere-based GPU into the more powerful BlueField-2X.

Both DPUs are built around Mellanox's ConnectX-6 DX smartNIC, which features integrated Arm processor cores that allow the product to achieve data transfer rates of 200 Gb/s and accelerate data center workloads like security, networking, and storage tasks.

The GPU-enhanced BlueField-2X ups the ante by enabling artificial intelligence (AI) functionality.

"One example of this is anomaly detection whereby we compute the patterns of traffic and behavior that represent 'normal,' so that whenever there is an intrusion, we can identify deviation from that abnormal behavior and proactively block that attack," Das said.

Thanks to these capabilities, Nvidia claims a single BlueField-2 DPU can deliver the same data center services as 125 CPU cores, freeing up those resources for other workloads.

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

Just as important as the hardware is the framework on which the code runs. To this end, Nvidia introduced DOCA, an architecture that does for its DPUs what CUDA did for its GPUs.

DOCA provides developers with an open platform for building hardware-accelerated networking, storage, security, and management applications for the BlueField-2. It will be integrated into Nvidia's GPU Cloud software library, which makes third-party software designed to take advantage of CUDA and, now DOCA, acceleration available in a containerized format for customers to implement in their systems.

According to Nvidia, Asus, Atos, Dell Technologies, Fujitsu, Gigabyte, H3C, Inspur, Lenovo, Quanta, and Supermicro have all announced plans to integrate Nvidia DPUs into their enterprise server offerings. Additionally, Nvidia is working with several software partners including VMware with Project Monterey, Red Hat, Canonical, and Check Point Software to leverage the capabilities offered by DPUs.

The BlueField-2 is in customer trials now, with the new system expected to hit the market in 2021. The BlueField-2X is still under development and expected to reach the market sometime next year.

Meanwhile, the DOCA software development kit is currently available in early access to Nvidia partners.

Nvidia expects to launch the follow up to the BlueField-2, the BlueField-3 and 3x, in 2022, and a fully integrated BlueField-4, which will integrate smartNIC and GPU functionality on a single die, in 2023.