Nvidia founder and CEO Jen-Hsun Huang held the annual GTC 2021 conference in his kitchen in the early hours of April 13. At the conference, Huang outlined Nvidia’s vision for the future of the computing industry, focusing on chips, software, services, edge computing, data centers and the cloud.

There were a lot of surprises, especially Nvidia Grace, the world’s first CPU designed for terabyte acceleration based on the ARM architecture, the new Bluefield-3 DPU, and the industry’s first 1000TOPS power self-driving car, the SOC-ATLAN.

Xiaobian also sorted out the following highlights for you —

Nvidia’s first CPU comes out

Huang’s first big announcement was the Nvidia Grace, Nvidia’s first CPU designed for large-scale artificial intelligence and high-performance computing applications.

It was named after U.S. Rear Admiral Grace Hopper, a computer programming pioneer and one of the first female programmers. She created the first modern compiler, the A-0 system, and the first high-level commercial computer programming language, “COBOL”. She was inspired to use the computer term “Debug” after removing moths from computers, and she was dubbed the “Mother of Debug.”

Unlike the CPUs we often use in commercial products such as computers and mobile phones, Grace is positioned as a highly specialized processor for large, data-intensive HPC (data center) and AI applications.

Grace’s innovation can be summarized as the following three points:

  • Built in the next generation ARM Neoverse kernel, each CPU can run more than 300 instances per unit of time in the SPECRate2017_INT_BASE benchmark;
  • Adopt the fourth generation NVIDIA NVLink, the connection speed from CPU to GPU is more than 900GB/s, equivalent to 14 times the bandwidth speed of the current server; The speed from CPU to CPU is over 600Gb /s.
  • With the highest memory bandwidth, the new memory LPDDR5X technology, bandwidth is 2 times that of LPDDR4, energy efficiency increased 10 times, can provide more computing power.

By combining GPUs and DPUs, Grace gives Nvidia a third foundational computing technology and the ability to reengineer data centers to advance the technology, which has the potential to transform Intel’s grip on more than 90 percent of the market for server processors and perhaps even displace Intel’s “No. 1” position.

And Nvidia believes that a custom-designed CPU-GPU platform is the only way it can implement the next generation of super-scale AI and begin to approach computer-based “general intelligence” levels, meaning that GRACE can solve AI problems that are orders of magnitude larger than today’s existing ones.

Researchers, including those at the U.S. Department of Energy’s Los Alamos National Laboratory and the Swiss National Center for Supercomputing, have agreed to build supercomputers using the GRACE chip.



The ALPS supercomputer at the Swiss National Supercomputing Center

Nvidia is embracing the ARM ecosystem

Apart from Grace’s announcement, the most impressive part of the event was Nvidia’s full embrace of the ARM ecosystem. “For good reason, because it’s super energy efficient and its open licensing model has inspired innovators around the world,” explains Huang.

At the event, Huang announced partnerships with several key ARM partners, including AWS, Ampere Computing, Marvel, and MediaTek

  • Nvidia Grace has a next-generation ARM Neoverse kernel built into it, providing a huge performance leap for systems to train large AI models. Specifically, Grace-based systems are tightly integrated with Nvidia GPUs and will perform up to 10 times better than current state-of-the-art Nvidia DGX systems, which run on x86 CPUs.
  • An Amazon EC2 instance based on the AWS Graviton2 will be deployed in the cloud in combination with an NVIDIA GPU. This new combination will deliver the benefits of lower costs, richer game streaming experiences, optimized Android games and AI reasoning in the cloud, and higher AI reasoning performance at lower costs. “The new partnership announced today is the first step in our commitment to expanding the ARM ecosystem beyond mobile and embedded systems,” Huang said.
  • To better support scientific and AI application development, Nvidia has launched a new HPC developer suite for the high-performance computing space. Nvidia’s new HPC Developer Suite provides a high-performance, energy-efficient platform for supercomputers that combines one Ampere Altra CPU (containing 80 ARM Neoverse cores), Dual NVIDIA A100 GPUs (each offering 312TFLOPS FP16 deep learning performance), two NVIDIA Bluefield-2 DPUs for accelerated networking, storage, and security. The developer suite, which includes a suite of Nvidia compilers, libraries, and tools for creating and migrating HPC and AI applications to GPU-accelerated ARM computing systems, will be available in the third quarter of 2021 and has been first deployed by a number of top research institutions.
  • Today Nvidia also announced improved edge video analytics and security features, as well as a new class of ARM-based PCs powered by Nvidia’s RTX GPU.
  • In edge computing, Nvidia is expanding its partnership with Marvell to combine ARM-based Octeon DPUs with GPUs to accelerate AI workloads for network optimization and security.
  • In the PC space, Nvidia has partnered with MediaTek, one of the world’s largest ARM-based SoC providers, to build a reference platform for Chromium, Linux and Nvidia SDK that uses ARM cores and Nvidia graphics cards. Bring the performance of GPU and advanced AI, ray tracing graphics and other technologies into ARM PC platform.
  • In addition, Nvidia is working with other partners such as Fujitsu and Sipearl to expand the ARM ecosystem.

What else is interesting about this conference?

1, 3 BlueField – DPU

The Bluefield-3 DPU will have 20 billion transistors, 16 ARM A78 CPU cores, and 18 megabytes of IOPS elastic block storage. Not only can be backward compatible with the previous generation of products, but also has four times the performance. Jen-xun Huang announced that the NVIDIA Bluefield-3 DPU will provide further acceleration of the infrastructure needed to build very large scale data centers, workstations and supercomputers.

A prototype is expected to be released in the first quarter of 2022. Its next-generation Bluefield-4 DPU will contain 640 transistors, with computing power up to 1,000 tops and network speeds up to 800Gbps.

2, NVIDIA Drive Atlan

The NVIDIA Drive Atlan is capable of 1,000 tops per SoC, a nearly four-fold increase over the previous generation of Orin SoCs (254TOPS), which is better than the total power of most L4 autonomous vehicles today.

The Atlan SoC, with its ampere architecture GPU core, ARM-based Grace CPU core, deep learning and computer vision accelerator units, and Bluefield DPU core, will be demo ready for developers in 2023 and shipped in large numbers in 2025.

In terms of the absolute amount of computing power, we can also see that the computing power upgrade of the Nvidia Drive Atlan is the largest among the four generations of autonomous driving chips.

Mr Huang makes no secret of his love for Nvidia Drive Atlan, saying: “Atlan combines all of Nvidia’s technologies in AI, automotive, robotics, security, and Bluefield’s secure data centres and is a technological marvel.”

conclusion

Today, Nvidia is launching a series of products that give you the most machine learning computing power in almost any industry and area. In short, Huang says, “you can almost say that Nvidia will help you build your life’s work.”

At the time of Huang’s Keynote, the company’s stock briefly crossed the $600 mark. Nvidia now has a market capitalization of $377.1 billion, compared with Intel’s $266.3 billion.

This conference fully embraces the ARM ecosystem, which also shows that industry and market interest in ARM-based solutions is increasing day by day. It remains to be seen if the introduction of Grace will lead developers into a new era of ARM.

Refer to the link: https://blogs.nvidia.com/blog…