What does Nvidia want to summon after putting together the last piece of the "CPU puzzle" with all three cores?

Core Comments - Understand industry current affairs news at the fastest speed and comment on industry development trends from the most unique perspective.

Jiwei.com reported that NVIDIA, which is known for its GPU, released its first data center CPU based on the Arm architecture, which may surprise many people. At this year's GTC 21, Huang Renxun, the leather jacket master, once again gave his heavyweight keynote speech in his kitchen, introducing NVIDIA's latest progress in AI, automobiles, robots, 5G, real-time graphics, collaboration and data centers.

With the CPU puzzle complete, NVIDIA has entered the era of the "three-core" combination of CPU, GPU and DPU.

Facing the future era of AI, autonomous driving, and smarter 5G, the giants seem to have made similar choices. Last year, Intel, which is known for its CPUs, released its self-developed GPU; AMD is also planning to acquire FPGA based on its CPUs and GPUs.

The era of chip integration continues to deepen, which also means that competition in the chip industry has entered a new stage.

Integrated CPU, DPU and GPU 

The focus of the whole event was the first data center CPU, Grace. So far, NVIDIA has filled the blank of CPU and completed the three basic computing capabilities.

Nvidia's first data center CPU is named after U.S. Navy Rear Admiral and computer programming pioneer Grace Hopper, and its next-generation natural language processing (NLP) model will have more than one trillion parameters for training.

Grace-based systems, tightly coupled with NVIDIA GPUs, will deliver 10 times the performance of today’s most advanced NVIDIA DGX systems (running on x86 CPUs).

However, it is worth noting that Grace is a highly specialized processor that is mainly aimed at large data-intensive HPC and AI applications. Huang Renxun pointed out that the vast majority of data centers will continue to use existing CPUs, while Grace will mainly be used in niche markets in the computing field and is expected to be available in 2023.

In terms of DPU, the BlueField-3 DPU released by NVIDIA achieved a 10-fold performance leap, capable of replacing 300 CPU cores and protecting, offloading and accelerating network traffic at a rate of 4000Gbps.

In terms of GPU, NVIDIA launched the NVIDIA A10 and A16 GPUs. Compared with the previous generation of virtual workstations, A10 provides up to 2.5 times the performance, A16 provides 2 times the user density, and its total cost of ownership is lower. In addition, when combined with NVIDIA RTX Virtual Workstation (vWS) and NVIDIA Virtual PC (vPC) software, these two GPUs will provide more powerful performance, video memory, and speed to accelerate any workflow from graphics, AI to VDI.

At this point, NVIDIA has upgraded its data center product roadmap to a three-chip combination of GPU+CPU+DPU. Huang Renxun said that Grace and BlueField are essential key components. This is an important upgrade to maintain its competitiveness in the field of AI.

New competitive landscape in the era of chip integration

So after gathering the "three cores", which businesses of NVIDIA will be helped to advance rapidly?

First of all, for NVIDIA, the collection of these three computing chips is of milestone significance for promoting the entire data center business. The computing power growth of today's data center business is no longer the responsibility of a single type of computing chip, and Huang Renxun's remarks seem to confirm this.

Huang Renxun believes that to ultimately build a data center server-level platform from a CPU core, GPU, DPU, system software, and all the algorithms above it are needed. By combining GPU and DPU, Grace enables NVIDIA to provide a third basic computing capability.

When Nvidia announced the launch of DPU last year, Huang Renxun said that data centers have become a new type of computing unit. The combination of CPU, GPU and DPU can form a fully programmable single AI computing unit, providing unprecedented security and computing power.

As the largest general-purpose CPU supplier, Intel, which has long dominated the data center field, also stated that the company is transforming from a CPU company to an XPU platform company, and "X" represents many different products.

After completing the puzzle of the data center roadmap, Nvidia's ambitions have gradually grown, and this business has become increasingly important to Nvidia. After announcing the launch of the new CPU, Nvidia's stock price closed up 5.62%.

According to Nvidia's fourth quarter and fiscal year 2021 financial report ending January 31, 2021, the data center's quarterly and full-year revenue both set records. In the fourth quarter, the data center business earned $1.9 billion in revenue, a year-on-year increase of 97%. Full-year revenue set a record of $6.7 billion, an increase of 124%.

The promotion of data center business has also brought NVIDIA a very high revenue growth rate. According to the latest report released by Gartner, among the top ten semiconductor suppliers in the world in 2020, NVIDIA ranked ninth, with the highest revenue growth rate of 45.2%.

Mercury Research analyzed that the addition of this new CPU will enable Nvidia to compete head-on with Intel.

Intel and AMD have also recently released new processors. Just last week, Intel released the third-generation Xeon Scalable Processor (codenamed Ice Lake), which uses Intel's 10nm process and is the industry's only data center CPU with built-in artificial intelligence acceleration, supporting end-to-end data science tools and a wide range of intelligent solution ecosystems. AMD's latest third-generation AMD EPYC processor (codenamed Milan) series based on the Zen 3 architecture uses TSMC's 7nm process.

Secondly, in addition to the high-performance computing market, Nvidia's three-chip combination can also enhance its competitiveness in the autonomous driving car market.

Huang said the field of self-driving cars is "one of the biggest challenges facing machine learning and robotics, and also one of the toughest and most impactful challenges."

NVIDIA is building modular end-to-end solutions for the $10 trillion transportation industry, enabling partners to take full advantage of the components they need. Based on NVIDIA's next-generation GPU architecture, new Arm CPUs, and deep learning and computer vision accelerators, Huang Renxun launched the next-generation NVIDIA DRIVE Atlan for self-driving cars, which has a performance of more than 100 trillion operations per second (TOPS), about 4 times that of the previous generation Orin processor, and exceeds the total computing power of most L5 driverless taxis.

With the global data explosion and the development of the artificial intelligence revolution, computing power growth can no longer meet the demand for data growth. According to IDC statistics, the global demand for computing power doubles every 3.5 months, far exceeding the current growth rate of computing power.

Driven by this, global computing, storage and network infrastructure are also undergoing fundamental changes. Some complex workloads cannot be handled well on general-purpose CPUs, so more computing chips such as GPUs, FPGAs, and DPUs are needed to meet the needs of applications such as AI, 5G, and autonomous driving that cannot be met by a single chip. The era of "X" PU chips has arrived, which also means that competition in the chip industry has entered a new stage.