Nvidia Rolls Out Slew of New AI Hardware for Data Centers

Monday, on the first day of Nvidia’s GTC Spring 2021 event, all eyes were on the announcement of the first Nvidia CPU, codenamed “Grace,” and rightfully so. Grace is a direct shot at Intel’s x86 processor architecture’s market share, which has dominated the data center for decades. Performance advantage with AI applications is one of the few standout characteristics of the latest Intel Ice Lake server chips, but, if Nvidia is to be believed, its upcoming CPU would easily blow past Ice Lake’s performance with AI models.

Read our article about the Nvidia Grace CPU here. Meanwhile, here's what other new AI hardware and infrastructure software products for data centers Nvidia announced at this week's summit, held virtually.

Nvidia BlueField DPU, a Data Center Infrastructure Accelerator

One of Nvidia’s newer concepts in AI hardware for data centers is the BlueField DPU (data processing unit) for data centers, first revealed at the last GTC, in October 2020. On Monday the company unveiled BlueField-3, a DPU it said was designed specifically for “AI and accelerated computing.”

Like Nvidia GPUs, its DPUs are accelerators, meaning they are meant to offload compute-heavy tasks from a system’s CPU, leaving the latter with more capacity to tackle other workloads. DPUs are powered by Arm chips.

Related:Nvidia Is Designing an Arm Data Center CPU for Beyond-x86 AI Models

Nvidia DPUs, based on the BlueField SmartNICs by Mellanox (acquired by Nvidia in 2019), take on things like software-defined networking, storage management, and security workloads. They’re also eventually expected to offload server virtualization, via a partnership with VMware as part of VMware’s Project Monterey.

Nvidia BlueField-3 DPU

Expected to start sampling early next year, BlueField-3 is the world’s first 400GbE DPU, Nvidia said. It is 10 times faster than its predecessor, BlueField-2, which was announced last year and became widely available this week.

Also available starting this week is DOCA, the company’s software platform for building infrastructure applications that take advantage of BlueField DPUs.

Nvidia DGX SuperPOD, Now a Private Cloud for AI Computing

BlueField-2 DPUs underpin a new multitenancy capability in the next-generation Nvidia DGX SuperPOD AI hardware announced Monday. A SuperPOD is essentially a turnkey AI supercomputer, consisting of 20 or more Nvidia DGX servers.

Through Nvidia Base Command, another new piece of software announced at GTC, the DGX SuperPOD can act as a private cloud, offering AI computing infrastructure that can be shared by multiple users in isolation from each other, the company said.

Related:Nvidia Says It's Achieved ‘Near-Bare-Metal’ AI Performance on vSphere

“The world’s first cloud-native supercomputer” is how Charlie Boyle, VP and general manager of DGX systems at Nvidia, described it during a GTC-preview video conference with reporters last week.

Nvidia DGX SuperPOD (background) and DGX Station (front)

Traditionally, a researcher that needs to use a supercomputer at an institution needs to go through a formal application process to secure time on a system and have their code vetted, which can be a drawn-out process.

With Base Command, users can access a slice of DGX SuperPOD capacity on-demand, much like customers consume public cloud infrastructure, Boyle explained.

“We can now provide secure multitenant access to any of those nodes in the supercomputer,” he said.

A30 and A10, AI Hardware for the Masses

Finally, Nvidia introduced two new data center GPUs aimed at enterprise customers that have no use for its top-shelf A100 GPUs, the bread-and-butter AI hardware.

The new Nvidia A30 GPU is for “mainstream AI and data analytics,” while the Nvidia A10 GPU is for “AI-enabled graphics, virtual workstations, and mixed compute and graphics workloads,” the company said.

Rendering of an Nvidia-certified EGX server

Nvidia partnered with major enterprise server vendors, who now offer “Nvidia-certified” servers with the new mainstream GPUs. The vendors are Atos, Dell Technologies, Gigabyte, H3C, Inspur, Lenovo, QTC, and Supermicro. More vendors may be added to the list later, according to the announcement.

Keeping with the theme of making AI infrastructure more accessible to traditional enterprises, these servers are certified to run Nvidia AI Enterprise software suite, built exclusively for VMware vSphere 7.