Nvidia GTC21 launches could further 'democratize' AI for enterprises

The Grace CPU could unleash much faster performance for AI models operating at supercomputing scale. (Nvidia)

Nvidia today announced a wide-ranging series of new products, enhancements and features at the company GTC21 event that broaden the ability of enterprises across a variety of industries to leverage artificial intelligence for more of their applications.

The moves, outlined in the GTC keynote speech by Nvidia CEO and co-founder Jensen Huang, also solidify the company’s ambitions to be a leader in enabling AI, and further transcend its legacy identity as a provider of graphic processing units, according to analysts who watched the news unfold.

Practical for enterprise

“Nvidia has been focused on trying to democratize AI so that it is not just something that the most sophisticated organizations can do,” said Zeus Kerravala, founder and principal analyst at ZK Research, in an interview with Fierce Electronics. “Nvidia is a GPU company, but it’s also a systems company and it is showing here that it can bring the systems and all the tools so that any company in any industry can build AI into their organization.”

Jack Gold, founder and principal analyst of J. Gold Associates, added, “Nvidia is demonstrating that it’s a leader in AI and machine learning. It’s traditional GPU business is not going away, but AI is where gazilions of dollars are to be made in the next few years.”

Among today’s announcements, Nvidia unveiled:

  • The Grace central processing unit (named for computing pioneer Grace Hopper), an Arm-based CPU for AI and High Performance Computing (HPC). It’s designed to address processing bottlenecks posed by rapidly growing AI models that already include “hundreds of millions of parameters, soon to be trillions of parameters in a few years,” said Paresh Kharya, senior director of product management and marketing at Nvidia. Grace, upon its expected availability in 2023, will be capable of delivering performance 10 times faster than Nvidia’s own current DGX-based data center systems that run on x86 CPUs.

Nvidia made clear that existing CPUs will serve the needs of most data centers, and that Grace is destined to be a niche product serving the largest-scale supercomputing needs. To that end, the company announced the first two parties intending to build Grace-powered supercomputers are the Swiss National Supercomputing Centre (CSCS) and the U.S. Department of Energy’s Los Alamos National Laboratory. 

Kevin Krewell, principal analyst at Tirias Research, said in an email to Fierce Electronics, "By connecting Grace to the GPU with NVLink, Nvidia can open up the bandwidth between the CPU and GPU. Grace can be custom-tuned by Nvidia to feed the memory and I/O requirements of the GPU."

  • A next-generation version of the company’s DGX SuperPod, to be augmented with BlueField 2 data processing units to create a cloud-native, multi-tenant supercomputer that allows more enterprises to leverage AI in a more secure way by allowing them to isolate, offload and accelerate AI workloads. The new DGX SuperPod also includes Nvidia Base Command, a solution which coordinates AI training and operation on SuperPod infrastructure for multiple teams working across many distributed locations, a growing reality of working with AI as adoption of the technology broadens to different corners of large enterprises. In addition, Nvidia announced a subscription plan for its A100 system, the work station model that in groups of 20 or more becomes a DGX SuperPod. The A100 subscription will make it easier for smaller scale enterprises and businesses to tap into supercomputing power for AI projects. Nvidia also separately announced it would work with German chemical research company Schrodinger to develop a SuperPod-based solution to accelerate drug discovery.
  • Broader availability for the Jarvis Interactive Conversational AI Framework, which provides developers with pre-trained deep learning models and software tools for creating conversational AI services targeted to the needs of specific industries or companies. The framework had been available in an early access mode since last May, but more enterprises will now be able to access conversational AI solutions built on the framework.
  • Collaborations with AstraZeneca and the University of Florida on AI research projects around drug discovery, as well as an announcement that CareStream Health, ActicSurgical and other firms are working with Nvidia's Clara AGX embedded AI platform for medical devices. The Clara AGX developer kit, which advances the development of software-defined medical instruments, also is now generally available.
  • A new classification of Nvidia-Certified Systems for AI Enterprise. This will allow  customers using the Nvidia AI Enterprise software suite and the VMware vSphere virtualization platform to now run virtualized AI applications on industry-standard servers, opening use of the AI software to many more customers.
  • The Morpheus application framework, a cloud-native suite of accelerated AI capabilities for cybersecurity firms allowing them to use machine learning to identify, capture and take action on threats and anomalies, such as leaks of unencrypted sensitive data, phishing attacks and malware, as they unfold. 
  • Omniverse Enterprise, a design collaboration and simulation platform for enterprises that allows 3D design teams working across multiple software suites and in geographically-dispersed locations to collaborate in real time in a shared virtual space on complex projects. The platform is aimed at speeding up processes that otherwise might require many settings, exchanges of massive computer files and the integration of separate iterations by different teams. The platform launched in an open beta three months ago and has since been downloaded by almost 17,000 users Initial evaluators of the platform included BMW Group, which used it to build a digital twin of entire factory, as well as Foster + Partners and WPP. 
  • The BlueField 3 DPU, the next generation of the company’s Bluefield DPU family, and backward compatible to BlueField 2. Scheduled to be available for sampling next year, it’s the company’s first DPU purpose-built for AI and accelerated computing. One BlueField-3 DPU delivers the equivalent data center services of up to 300 CPU cores, freeing up valuable CPU cycles to run business-critical applications, the company said.
  • A series of collaborations that combine Nvidia GPUs and software with Arm-based CPUs for applications in cloud, high-performance computing, edge computing and personal computing. These collaborations include combining the AWS Graviton2-based Amazon EC2 instances with NVIDIA GPUs. Nvidia also has a new HPC to support AI applications in those environments.
  • The Alps Supercomputer. Nvidia is working with the Swiss National Supercomputing Centre and Hewlett Packard Enterprise to develop what the company called the world’s most powerful AI-capable supercomputer.

AI is hard

With efforts focused on the high-level supercomputing niche, as well as small-scale enterprises and every size of enterprise in many different industries in between, Nvidia seems focused on evolving AI from a high-level concept to something more malleable and of more practical use to enterprises. “AI is hard. People really want to use it, but sometimes lack the resources and skills to do that,” Gold said. “AI is getting to the point where it needs to become easier to work with, and that’s what Nvidia is trying to do.”

Shane Rau, research vice president, computing semiconductors at IDC, said Nvidia is showing that “AI is a component in every part of the stack that it offers--silicon, software, systems and applications. You can get it all from them, or buy it in a more modular way, and layer in your own secret sauce if you have that capability.” Also, particularly with the addition of the Grace CPU, Rau noted that Nvidia now has a stake in GPUs, DPUs and CPUs, the three main chip architectures.

These moves could put more pressure on other companies in the semiconductor space, such as Intel, AMD, NXP and Qualcomm, analysts said. "Could other semiconductor providers and other kinds of companies do what Nvidia is doing here? Yes," Rau said. "Intel has bought a lot of hardware, software and silicon assets that it is building into its own stack. NXP s taking a slightly different approach, but all of these companies are expanding their businesses."

Kerravala added that while Intel has many of the necessary silicon and technology tools, it has not demonstrated that it is bringing these to bear in the market in the form of applications tailored to different industries and AI use cases. "Nvidia is saying that wherever your application is it has an AI solution for you. In healthcare, it's Clara, in the auto industry it's Drive for conversational AI needs its Jarvis, for 3D design it's Omniverse."

Nvidia isn't done making moves yet. It's still in the process of trying to acquire Arm, a company with silicon for GPUs, DPUs and CPUs. Rau said that with this week's announcements and that acquisition--if it navigates regulatory hurdles and closes--Nvidia is well along the path of effectively transforming itself from a GPU firm into a formidable player in the data center ecosystem with the power to influence the data center technology roadmap.

RELATED: Nvidia will buy Arm from Softbank for $40 billion