Nvidia adds DPUs to GPU lineup for artificial intelligence apps
Nvidia has rolled out a new chip technology that supplements its GPUs, along with a new development kit and architecture to create more capable AI-based applications.
Nvidia doubled down on its commitment to enhance AI application performance with new data processing units, a software development kit and a supporting architecture that optimizes newly developed apps and services to run faster across storage, networking and security systems.
The Nvidia BlueField-2 DPU includes all of the capabilities of the latest Mellanox SmartNICs, combined with Arm. The programmable chip delivers data transfer rates of 200 gigabits per second to speed data center security, networking and storage tasks.
The BlueField-2X is enhanced with the company's Ampere GPU with AI capabilities. It permits real-time security analytics, such as abnormal traffic indicating data breaches, and encrypted traffic analytics.
Jensen Huang
"We see the data center becoming the new unit of computing," said Jensen Huang, Nvidia's CEO, during the company's GPU Technology Conference. "DPUs are an essential element of modern data centers where CPUs, GPUs and DPUs can be combined into a single computing unit."
One BlueField-2 DPU is capable of delivering the same data center services that could consume up to 125 CPU cores, according to Nvidia.
The adoption of [DPUs] will be pretty slow for a while. There's an awful lot of tire kicking that has to happen.
Ezra GottheilPrincipal analyst, Technology Business Research, Inc.
"The adoption of [DPUs] will be pretty slow for a while," said Ezra Gottheil, principal analyst with Technology Business Research, Inc. "There's an awful lot of tire kicking that has to happen," he said.
Gottheil believes acceptance of the new technology will likely occur first among large cloud providers, such as AWS, Google and Microsoft, followed by third-party and corporate application developers focused on the AI and machine learning.
"This platform has to be accepted by developers, but most of the developers are inside the cloud service providers," Gottheil said. "They have the R&D to pour into this and they can move a lot faster than the enterprise [IT] guys can," he said.
Another reason large cloud providers may gravitate toward DPUs first, particularly AWS and Google, is they make customized servers for their own mammoth data centers. This dovetails with Nvidia's strategy with its DPUs and the technology used in the Arm processors, which Nvidia recently purchased.
"[Nvidia] is clearly going into the business of creating specialized microprocessor parts that can be assembled like blocks," Gottheil said. "They are subdividing the processing capabilities among DPUs and GPUs to take the weight off CPUs. This approach makes sense to me."
DPUs mark another step in chip evolution
One consultant sees Nvidia's move to DPUs as similar to what happened with the trend in the 1980s when computing architectures migrated slowly from mainframes to client-server systems.
"The same thing happened thirty years ago when the computing architecture was concentrated around the CPU and then morphed into multiple chips to better accommodate client-server architectures," said Frank Dzubeck, president of Communications Network Architects, Inc. "This is interesting technology but it is more evolutionary than revolutionary."
While AWS, Google and Microsoft have yet to endorse Nvidia's DPUs, server manufacturers Dell, Lenovo and Asus plan to integrate the technology into their servers. Last week, VMware disclosed a partnership with Nvidia as part of Project Monterey, along with Red Hat which said it will offer its hybrid cloud portfolio components on Red Hat Enterprise Linux and Red Hat OpenShift. Check Point Software said it will integrate the chip in its cybersecurity offerings.
Nvidia BlueField roadmap
Nvidia laid out a roadmap for the BlueField line of DPUs starting with the BlueField-2 series, which is currently being sampled. The BlueField-3 chip is now in development but nearing completion, Huang said. BlueField-4 will be delivered sometime in 2023 and will support the company's CUDA parallel platform along with Nvidia AI.
"We are going to bring a ton of technology to networking," Huang said. "In just a couple of years, we'll increase nearly 1,000 times in compute throughput on the DPU," he said.
Meanwhile, the Nvidia data-center-on-a-chip (DOCA) SDK allows developers to create applications that work hand-in-glove with DPU-accelerated data center infrastructure services, similar to the Nvidia CUDA programming model that enables developers to build GPU exploitive applications.
DOCA, which is an open development platform, is integrated into the Nvidia NGC, a software catalog that provides a containerized software.
Nvidia's EGX platform will also be combined with Ampere GPU and BlueField-2 DPU onto a single PCIe card. Huang said during his keynote that using a common platform makes it easier for enterprise users to build a more modern data center.
Also highlighted at the conference was a handful of improvements Nvidia is making to the Arm platform to give users more development choices for both the Arm and Nvidia platforms.
Those improvements include: Nvidia AI to speed AI training; Nvidia Rapids, a suite of software libraries to run data science and analytics on GPUs; and Nvidia HPC SDK, which is made up of compilers, libraries made for high performance computing.
The BlueField -2 DPUs are expected to make an appearance in new systems from server vendors sometime in 2021. The BlueField-2X DPUs, still under development, are also scheduled for release in 2021. DOCA is available now to Nvidia's business partners.
The DOCA SDK is available now to Nvidia's early access partners.
Jensen Huang
