Nvidia unveils DPU – a data center system concentrated on a plug-in card

Those who are versed in technical concepts are probably familiar with the processor being abbreviated as CPU, central processor unit, and graphics cards abbreviated GPU, graphics processor unit. In modern times, general calculations on the graphics card, GPGPU or general purpose graphics processor unit, have also taken place in technology fans’ vocabulary. During the GTC 2020 graphics fair, Nvidia introduces another abbreviation to keep track of.

With DPU, or Data Processing Unit, Nvidia combines a number of different components used in data centers on one and the same circuit. A DPU processor uses the DOCA architecture, which specifies both the structure of the circuits and the developer model for how the various components on them are to be used. The DPU processors themselves are part of the Bluefield product family, which will make its debut with Bluefield-2 in 2021. The purpose of the products is to relieve processors in data centers from a series of tasks that are performed more efficiently with a DPU unit.

The data center has become the new unit of computing. 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.

In terms of the set of components, a Bluefield-2 integrates eight high-performance ARM cores of the Cortex-A72 type, which makes it clear why Nvidia recently acquired the company behind ARM’s processor designs and instruction sets. The ARM cores are complemented by an “accelerator motor” that applies long instructions (VLIW), a type of instruction used in both graphics cards (AMD’s VLIW4 / VLIW5) and processors (Intel’s EPIC in the Itanium processors).

According to Nvidia, a Bluefield-2 can replace the work that would require 125 processor cores in a typical data center setup. The processor itself connects to both DDR4 memory and the rest of the Bluefield-2 card over PCI Express 4.0, and the card itself communicates with the host system over NVME and a physical PCI Express 4.0 connection. According to the company, a Bluefield-2-DPU supports as much as 50 million IOPS over NVME and PCI Express 4.0.

As the cards are to be used in data centers, they are also equipped with network connections, also here with technology from an acquired company in the form of Mellanox. The network technology takes shape in the physical connections, which consist of dual 100 Gbps Ethernet connections with Mellanox Infiniband technology. In addition to raw performance, Mellanox technologies are also supported for, among other things, elastic block storage. In terms of streaming video, this can be done with bandwidth of 100 Gbps.

In addition to Bluefield-2, Nvidia also unveils Bluefield-2X, where the additions include third-generation Tensor cores used in the Ampere architecture. In addition to calculating pure AI, these should also be used to accelerate calculations related to security, networking and storage. Examples of how AI calculations can be used for networks include analysis of abnormal network traffic and suspected intrusions into virtual guest systems.

Nvidia’s product plan states that Bluefield-2 will thus reach the market in 2021, followed by Bluefield-3 a year later. The step between the two becomes solid as the calculation capacity of 70 SPECINT and 0.7 TOPS takes a massive jump to 350 SPECINT and 1.5 TOPS with Bluefield-3. The next model Bluefield-4 reaches the market in 2023 and offers slightly reckless 1000 SPECINT and 400 TOPS.

Read more about Nvidia and data center: