After showing off his muscles with the world’s fastest AI supercomputing, Huang Renxun uses AI to seize the gate of the Metaverse

AI is still the best story of “Nuclear Bomb Factory”.

“AI in the next era”, founder Huang Renxun said this word many times in a keynote speech of 1 hour and 40 minutes at the NVIDIA GTC conference on March 22, Beijing time.

In the black virtual scene, Huang Renxun methodically introduced a series of hardware, software, AI and robot application frameworks for AI computing, and introduced NVIDIA’s achievements in autonomous driving, virtual world, medical and other fields with the help of AI in the past period of time. .

At the GTC2021 in the fall of last November, Huang Renxun had made a high-profile announcement of “entering the Metaverse”. In contrast, this GTC2022 focused on more down-to-earth issues.

Since its birth, “Metaverse” has gone from being popular in the industry to being synonymous with “unrealistic”. Metaverse players who have not left the field after calming down have to think about a serious question: what should we start from to reach such a distant future.

“AI” is the life gate of the Metaverse captured by NVIDIA.

For the Metaverse, image processing and generation capabilities are facing an improvement of tens of millions, and AI is just capable of more complex and finer image processing. Whether it is in replication simulation or innovative construction, AI is the indispensable foundation.

The more basic and crucial thing behind “AI” is “computing power”.

After more than ten years of development, more and more data have been collected, and more and more large-scale algorithm models have been born, followed by a sharp rise in the data and parameters to be processed.

Some professionals believe that in order to realize the Metaverse scene depicted in “Avalanche”, at least a 1,000-fold increase in computing power is required, and industry giants such as Apple, Tesla, and Meta are gradually turning to chip self-development and customization.

The industry calls for a more efficient computing hardware foundation. In the face of the “barbarians” who suddenly came to the door, NVIDIA chose to take the initiative.

Whether it is releasing the H100 GPU and Grace CPU based on the new Hopper architecture, or showing its progress in AI software this time, NVIDIA has revealed its layout and ambitions for the next generation of AI trends.

01 Computing power: the top priority


In the keynote, the first announcement was the H100, the first GPU based on the new Hopper architecture.

NVIDIA H100 uses TSMC 4N (TSMC 4nm) process, integrates 80 billion transistors, significantly improves the speed of AI, HPC, memory bandwidth, interconnection and communication, and can achieve nearly 5TB/s of external interconnection bandwidth.

“20 H100 GPUs can handle the global Internet traffic!” Huang Renxun announced at the meeting.

An order of magnitude performance leap, the H100 is one of the largest graphics processors Nvidia has ever produced. Its FP8 computing power is 4PetaFLOPS, FP16 is 2PetaFLOPS, TF32 computing power is 1PetaFLOPS, and FP64 and FP32 computing power is 60TeraFLOPS.

After showing off his muscles with the world's fastest AI supercomputing, Huang Renxun uses AI to seize the gate of the Metaverse


The large-scale training performance of the H100 is 9 times that of the “predecessor” A100, and the throughput of large-scale language model inference is 30 times that of the A100.

At the same time, Hopper has also built a proprietary engine specifically for Transformer, which reduces training that would have taken weeks to days. With the same model training accuracy, the performance is improved by 6 times.

In addition, H100 is the world’s first accelerator with confidential computing capabilities, both AI models and customer data will be protected.

Grace CPU Super Chip

In addition to the H100, the Grace CPU, which Huang Renxun called “the ideal CPU for the global AI infrastructure”, is also not inferior.

Grace CPU is NVIDIA’s first dedicated CPU for AI infrastructure and high-performance computing. Based on the latest data center architecture Arm v9, it consists of two CPU chips, with 144-core CPU, power consumption of 500W, and performance improved by two to two three times.

After showing off his muscles with the world's fastest AI supercomputing, Huang Renxun uses AI to seize the gate of the Metaverse


Two CPUs are connected through NVLink, which can realize the interconnection between chips, and has the characteristics of high speed and low delay. Grace CPU and Hopper can also be customized through NVLink.

NVLink technology will be widely used in NVIDIA’s chips in the future, including CPU, GPU, DPU and SoC. With this technology, NVIDIA users will be able to use NVIDIA’s platform to achieve semi-custom construction of chips.

EoS The world’s fastest AI supercomputer

The computing power is not enough, and the number is to make up.

Through Huang Renxun’s explanation, we can know that 8 H100s and 4 NVLinks can be combined into a DGX H100. This giant GPU has 640 billion transistors and an AI computing power of 32 petaFLOPS; 32 DGX H100s can form a 256-piece GPU. DGX POD; 18 DGX PODs and a total of 4608 GPUs are built together, which is the EoS supercomputing announced by NVIDIA this time.

After showing off his muscles with the world's fastest AI supercomputing, Huang Renxun uses AI to seize the gate of the Metaverse


In the end, the computing power that EoS can achieve is 275 petaFLOPS according to the traditional supercomputing standard, which will be 1.4 times that of the previous A100-based largest supercomputing Summit in the United States; from the perspective of AI computing, EoS output 18.4 Exaflops, which will be the first in the world today. Four times more than Fuyue.

By then, EoS will be the fastest AI supercomputer in the world.

02 Software: steadily updated

In terms of software systems, Nvidia is still steadily updating.

This time NVIDIA released more than 60 updates to a series of libraries, tools and technologies for CUDA-X, and introduced its progress in climate prediction, Riva, a conversational AI service, and Merlin framework for recommendation systems.

After showing off his muscles with the world's fastest AI supercomputing, Huang Renxun uses AI to seize the gate of the Metaverse

Earth-2 | NVIDIA

At GTC2021 last year, NVIDIA released Earth-2, the first AI digital twin supercomputer. In the past few months, NVIDIA developed a weather forecast AI model FourCastNet based on this.

The model was jointly developed by NVIDIA and researchers from universities and research institutions such as Caltech and Berkeley Lab. By training up to 10TB of Earth system data, the accuracy of predicting the probability of precipitation is higher than previous models.

Subsequently, Huang Renxun introduced Riva, NVIDIA’s conversational AI service.

Version 2.0 of Riva supports recognition of 7 languages, converts neural text to gender-voiced speech, and can be custom tuned by users through its TAO transfer learning toolkit.

Maxine is a toolkit of 30 AI models that optimize the audiovisual effects of video communications in real time.

After showing off his muscles with the world's fastest AI supercomputing, Huang Renxun uses AI to seize the gate of the Metaverse

Maxine | NVIDIA

When a remote video conference takes place, Maxine helps the speaker maintain eye contact with the rest of the meeting, even when you’re reading a manuscript or browsing other web pages. If the participants contain different nationalities and use different languages, Maxine can switch to another language in real time through the AI ​​model.

The Merlin framework is aimed at recommender systems.

It enables enterprises to rapidly build, deploy and scale advanced AI recommender systems. Huang Renxun used WeChat as an example in the live broadcast. After using Merlin, WeChat’s short video recommendation delay was shortened to a quarter of the original, and the throughput was increased by 10 times. Migrating from CPU to GPU, Tencent’s cost in this business was reduced by two. one part.

03 Metaverse and a new wave of AI

While improving computing power and making up for the shortcomings of the CPU, NVIDIA has not forgotten the “sea of ​​stars” in the Metaverse that it ultimately pursues.

Huang Renxun’s avatar, Toy Jensen, once again came on stage to have a dialogue with the deity, and it is worth noting that this time Toy Jensen was able to make eye contact and dialogue with Huang Renxun in complete real-time.

Faced with tricky questions such as “what is synthetic biology” and “how do you make it”, Toy Jensen gave smooth answers.

Behind Toy Jensen is Nvidia’s Omniverse Avatar framework, which enables businesses to quickly build similar avatars that look, move, and sound like they are.

The real-time dialogue is the technical support provided by the Riva mentioned above and the large language model Megatron 530B NLP, so that the avatar can understand the question and reply in real time.

After showing off his muscles with the world's fastest AI supercomputing, Huang Renxun uses AI to seize the gate of the Metaverse

Toy Jensen in conversation with Jensen Huang | NVIDIA

Building avatars and interacting in real time is undoubtedly the norm in the future Metaverse world, and in just a few minutes of display, Nvidia told us that this does not seem impossible.

In addition, in Huang Renxun’s view, new chips, software and simulation functions will set off a “new wave of AI”. The first wave of AI learning is perception and reasoning, while the next wave of AI development will be robotics.

At present, NVIDIA has gradually built NVIDIA Avatar for virtual images, DRIVE for autonomous driving, and Metropolis for manipulation and control systems around the four pillars of real data generation, AI model training, robot stack and Omniverse digital twin. , Isaac for autonomous infrastructure, and end-to-end full-stack robotics platforms like Holoscan for medical devices.

At the end of the keynote speech, Huang Renxun spent about 8 minutes leading the audience to sort out the newly released technologies, products and platforms from scratch, and summed up 5 trends affecting the development of the industry: million-X million times computing speed leaps, Transformers that dramatically accelerate AI, become data centers for AI factories, exponentially growing demand for robotic systems, and digital twins for the next AI era.

And the improvement of “computing power” will still be the basis of all breakthroughs.

“We’re going to accelerate the entire stack at data center scale over the next decade, making another million-X million times the performance leap. I can’t wait to see what the next million times performance leap will bring.”

Posted by:CoinYuppie,Reprinted with attribution to:
Coinyuppie is an open information publishing platform, all information provided is not related to the views and positions of coinyuppie, and does not constitute any investment and financial advice. Users are expected to carefully screen and prevent risks.

Like (0)
Donate Buy me a coffee Buy me a coffee
Previous 2022-03-23 10:28
Next 2022-03-23 22:36

Related articles