It was reported on April 13 that in the early morning of this morning, NVIDIA GTC, an annual event affecting artificial intelligence and high-performance computing technology, arrived as scheduled. This is the second time that the NVIDIA GTC conference has been held online since last year.

Last year, NVIDIA released A100 GPU and a series of servers, clusters and supercomputing, which shocked the AI field. Now, NVIDIA CEO huangrenxun returns with a series of new software and hardware products.
Half a year after announcing the $40billion acquisition of arm, NVIDIA released three new processors based on arm, including the world’s first CPU NVIDIA grace designed specifically for TB accelerated computing, the new bluefield-3 DPU, and the industry’s first 1000tops computing autonomous vehicle SOC.
After NVIDIA announced three new processors including server CPU, so far, the US stock NVIDIA has risen by 5.62%, while Intel has fallen by more than 4% and AMD has fallen by more than 5%.
1、NVIDIA Grace
At the press conference, Huang Renxun announced his first blockbuster new product, NVIDIA grace, a CPU specially designed for large-scale artificial intelligence and high-performance computing applications.
It is understood that the vast majority of data centers will continue to use the existing CPU, and grace will mainly be used in the segment of the computing field, which is expected to be available in 2023. Because it is difficult to completely put the super large-scale model into the GPU memory, if it is stored in the system memory, the access speed will be greatly limited. This CPU is mainly designed to solve this bottleneck.

NVIDIA named this CPU after grace hopper, a pioneer computer scientist who invented the world’s first compiler and is known as the “first lady of computer software engineering”. There are three specific innovations:
(1) Built in next generation arm neorse kernel, each CPU can_ int_ Base benchmark runs more than 300 instances per unit time;
(2) The fourth generation NVIDIA nvlink is adopted, and the connection speed from CPU to GPU exceeds 900gb/s, reaching 14 times the bandwidth of the current server; CPU to CPU speed exceeds 600gb/s.
(3) It has the highest memory bandwidth. The new memory lpddr5x technology is adopted. The bandwidth is twice that of lpddr4, and the energy efficiency is increased by 10 times. It can provide more computing power.
Next year, two powerful AI supercomputers will be launched, both of which will adopt NVIDIA grace. It is said that it is closely combined with NVIDIA GPU, and its performance will be 10 times higher than that of the most advanced NVIDIA DGX system (running on X86 CPU).
Among the two AI supercomputers, the Swiss National Computing Center (CSCs) is building a system with computing power up to 20exaflops, and Los Alamos National Laboratory of the United States will also equip its researchers with new AI supercomputers.
2、Bluefield-3 DPU
Huangrenxun believes that the data processing unit (DPU), which is responsible for transmitting and processing data in the data center, is forming “three pillars of future computing” together with CPU and GPU.
NVIDIA’s new bluefield-3 DPU contains 22billion transistors. It uses 16 arm A78 CPU cores and 18m IOPs elastic block storage. Its encryption speed is 4 times that of the previous generation and is fully downward compatible with bluefield-2.

The new generation bluefield-3 DPU is expected to release samples in the first quarter of 2022. The computing power of the fourth generation Bluefield DPU reaches 1000tops and the network speed reaches 800gbps.
Bluefield-2 can replace the workload equivalent to 30 CPU cores, while bluefield-3 achieves a 10x increase in accelerated computing performance and can replace 300 CPU cores to protect, unload and accelerate network traffic at a rate of 400gbps. The processor is also the first DPU that supports the fifth generation PCIe bus and provides data center time synchronization acceleration.
Bluefield-3 provides developers with a complete and open software platform through NVIDIA DOCA (Architecture integrating data center on chip) software development package to develop network, storage, security, management and other applications for software definition and hardware acceleration on Bluefield DPU. DOCA was released and available for download today.
NVIDIA drive Atlan is a new generation of AI autonomous vehicle processor. Its computing power will reach 1000tops, about four times that of the previous generation Orin processor, surpassing the intelligent computing power of most existing driverless taxis.

This is the first time that the drive platform integrates DPU. It brings a data center level network for autonomous vehicle through the arm core, and is committed to applying it to 2025 models.
The SOC adopts the architecture of the next generation GPU, a new ARM CPU core, a new deep learning and computer vision accelerator, and has built-in Bluefield DPU for advanced network, storage and security services. The network speed can reach 400gbps.
Huangrenxun praised that “Atlan integrates all the technologies of NVIDIA in the fields of AI, automobile, robot, security and Bluefield secure data center, which can be called a technological miracle.”
“We release exciting new products every year. Three types of chips have leapt year by year, one architecture.” Huangrenxun said that the data center roadmap includes three types of chips: CPU, GPU and DPU, of which grace and Bluefield are essential and key components. Each chip architecture goes through a two-year polishing cycle (there may be changes in the cycle), one year focusing on the X86 platform and the other on the arm platform.
In addition, NVIDIA also announced the partnership with Amazon AWS, ampere computing, MediaTek, Marvell and other arm based CPU platforms.
In terms of software, Megatron, a large-scale language model training and reasoning Q & A engine, Jarvis, a real-time conversational AI platform, Morpheus, Omniverse Enterprise Edition, cuquantum, a GPU accelerated quantum circuit simulation framework, and a series of new developments were announced for the first time.

For the comprehensive report of electronic enthusiasts, please refer to NVIDIA, Orient Securities, etc. for reprint, please indicate the source above.

Leave a Reply

Your email address will not be published.