Either he or that man, the familiar leather jacket.
14 On the evening of May 6th, Huang Renxun held an online conference of NVIDIA GTC 2020 in the kitchen. Due to the impact of the COVID-19 epidemic, the planned on-site activities in NVIDIA were forced to be cancelled, and the press release originally scheduled to be released through the media on March 24th also disappeared. After a long wait, Huang Renxun finally met everyone in front of the oven.
This year's GTC has not taken the usual path since the warm-up. Huang Renxun exposed himself the day before the conference, and took out a brand-new Ampere architecture GPU NVIDIA A 100 from the oven.
Surprisingly, although offline activities could not be held, Invista was too lazy to broadcast live online, and directly played the video recorded by Huang Renxun in his kitchen to complete the release of new products. Sure enough, if you have "hard goods" in your hand, you don't care about form.
NVIDIA's first Ampere architecture GPU can be regarded as "the strongest in history". Based on 7nm process, it has 54 billion transistors with an area of 826mm? Compared with Volta architecture, the performance is improved by 20 times, which can be used for both training and reasoning.
NVIDIA A 100 has the third-generation Tensor Core core of TF32, which can improve the AI performance under FP32 precision by 20 times, reaching 19.5 trillion times per second without changing any code.
Multi-instance GPU-MG can divide a single A 100 GPU into seven independent GPUs, and provide different computing power according to different tasks, so as to maximize the utilization rate and return on investment.
The new efficiency technology of NVIDIA A 100 takes advantage of the inherent sparsity of AI mathematics, and the performance doubles after optimization.
NVIDIA summarized the characteristics of NVIDIA A 100 as the following five points:
Huang Renxun said: "The breakthrough design of Ampere architecture has provided the biggest performance leap for NVIDIA's eighth generation GPU so far. It integrates AI training and reasoning, and its performance is up to 20 times higher than that of the previous generation. For the first time in history, the horizontal expansion and vertical expansion load can be accelerated on one platform. A 100 will improve throughput and reduce the cost of data centers. "
NVIDIA A 100 is the first GPU based on NVIDIA Ampere architecture, which provides the biggest performance improvement among NVIDIA VIII GPU. It can also be used for data analysis, scientific calculation and cloud graphics, and has been fully put into production and delivered to customers around the world.
18 The world's leading service providers and system builders are integrating NVIDIA A 100 into their services and products, including Alibaba Cloud, AWS, Baidu Cloud, Cisco, Dell Technologies, Google Cloud, HPE, Microsoft Azure and Oracle Bone Inscriptions.
Huang Renxun also introduced the third generation AI system DGX-A 100 AI based on NVIDIA A100. DGX-A 100 AI is the world's first server with a single-node AI computing power of 5 PFLOPS. Each DGX A 100 can be divided into as many as 56 independent running instances, and 8 NVIDIA A100 GPUs are assembled, and each GPU supports 12 NVLink interconnection buses.
It is understood that compared with other high-end CPU servers, the AI computing performance of DGXA 100 is 150 times higher, that of memory bandwidth is 40 times higher and that of IO bandwidth is 40 times higher.
Huang Renxun said: "AI has been applied to many fields, such as cloud computing, automobile, retail and medical care, and the AI algorithm has become more and more complex and diverse. The computing power requirement of ResNet model has increased by 3000 times from 20 16 to now, and we need a better solution. "
The price of such a powerful DGX-A 100 AI is naturally not cheap. The price tag is199,000 USD, which is about RMB 14 1 10,000 Yuan.
In addition, Huang Renxun also mentioned the new generation DGXSuper POD cluster in NVIDIA, which consists of 140 DGXA 100 systems, and the AI computing power reaches 700 Petaflops, which is equivalent to the performance of thousands of servers.
It is understood that the first batch of DGXSuper POD will be deployed in the Argonne National Laboratory of the US Department of Energy for research on the COVID-19 epidemic.
In addition to the above two blockbuster products, Huang Renxun also announced the launch of NVIDIA Merlin, an end-to-end framework for building the next generation recommendation system, which is rapidly becoming the engine of a more personalized Internet. Merlin reduced the time required to create a 100 TB data set recommendation system from four days to 20 minutes.
NVIDIA has also launched many products related to the AI field, including Mellanox ConnectX-6 Lx SmartNIC, EGX Edge AI platform and a series of software updates and extensions.
1. Ethernet smart network card Mellanox ConnectX-6 Lx SmartNIC
ConnectX-6 Lx is the industry's first security smart network card optimized for 25Gb/s, which can provide two 25Gb/s ports or one 50Gb/s port.
2.EGX edge AI platform
EGX Edge AI platform is the first Edge AI product based on NVIDIA Ampere architecture, which can receive data up to 200Gbps and send it directly to GPU memory for AI or 5G signal processing.
3.spark 3.0
Invista also announced its support for NVIDIA GPU acceleration on Spark 3.0. Spark 3.0 based on RAPIDS breaks the performance benchmark of data extraction, transformation and loading. It helps Adobe Intelligent Services reduce the computing cost by 90%.
4. NVIDIA Jarvis
Huang Renxun introduced NVIDIA Jarvis in detail at the press conference, which is a brand-new end-to-end platform, which can give full play to the powerful functions of NVIDIA AI platform and create real-time multi-modal conversational AI.
5. Fog interactive artificial intelligence
In the live demonstration, an AI system named Misty showed the interactive process of understanding and answering a series of complex questions about the weather in real time.
In the aspect of autonomous driving, Invista also embeds the Ampere architecture into the new NVIDIA drive platform. It is understood that autonomous driving companies such as Ma Xiao Zhixing and Faraday Future have announced the adoption of NVIDIA's DRIVE AGX computing platform.
NVIDIA's NVIDIA Isaac software-defined robot platform will also be used in the BMW Group factory. NVIDIA's robot global ecosystem covers distribution, retail, autonomous mobile robots, agriculture, service industry, logistics, manufacturing and health care.
NVIDIA's three-year conference was full of sincerity, and the first ampere architecture was a big surprise. The performance of NVIDIA A 100 GPU is improved by 20 times, which is a leap in performance.
Although the conference was not broadcast live, it was still full. One DGX-A 100 AI is better than a thousand, which also confirms Huang Renxun's classic saying "The more you buy, the more you earn". NVIDIA's AI solutions have covered all walks of life, and a strong AI ecosystem is taking shape.
Ni Guangnan, an academician of China Academy of Engineering, once said: "The threshold of chip design is extremely high, and only a few enterprises can bear the research and development costs of mid-to high-end chips, which also restricts the innovation in the chip field. 」
NVIDIA's Ampere architecture and a series of AI platforms based on it demonstrated the strength of an AI chip giant at this GTC, and once again set a performance benchmark.
According to Gartner's forecast data, the global artificial intelligence chip market will soar in the next five years, from $4.27 billion in 20 18 to $34.3 billion, an increase of more than seven times, which shows that the AI chip market has great room for growth.
Although there is still a gap between China and western developed countries in the research and development of AI chips, in the past two years, AI chip start-ups in China have received hundreds of millions of dollars. Companies such as Huawei have also developed impressive chip designs.
However, chip development is extremely complicated. There is a shortage of talents in China, and a few China semiconductor companies ranked in the top 65,438+05 in global sales are missing, which shows that China needs to make significant progress to compete with the United States in the semiconductor field.