Recently, AMD and Nvidia received news that the US government is demanding that they cut off the supply of high-end GPUs to the China region.

 

On the evening of August 31st, Jiweiwang suddenly announced that AMD and Nvidia China have successively received notices from headquarters to cut off the supply of top-level computing chips for artificial intelligence and data centers to customers in China.

 

Regarding AMD:

 

Suspend shipments of all MI100 and MI200 in China;

Statistics on the shipped quantity of Ml100 in China;

Collect a list of customers and shipping details for MI200 in China.

 

On the Nvidia side:

 

Suspend the shipment of A100 and H100 to all customers and agents in China, and other GPU cards will not be affected;

 

The A100 inventory of each server OEM can continue to be delivered to their respective industry customers, and Nvidia China has not yet issued any letters to OEMs.

 

At present, Nvidia H100 has just been released at this year's GTC conference and has not yet been shipped. But the A100, which has been shipped for 2 years, breaks 16 performance records and leads the way in the latest MLPerf.

 

It can be seen that the power outage of only these two high-end GPUs, Nvidia, directly hit the training of intelligent computing and AI large models.

 

As of the time of publication, AMD and Nvidia have not provided any comments.

 

The United States is increasing its efforts again!

 

Unfortunately, this news does not seem to be unfounded.

 

According to a document submitted by Nvidia to the US Securities and Exchange Commission, the US government issued a notice to Nvidia on August 26th, proposing new export control licenses for the existing A100 and upcoming H100, which will take effect immediately.

 

 

In addition, DGX or any other system using A100 or H100 and A100X, as well as any future system that can achieve or exceed A100 peak performance and I/O performance, are within the scope of the new license requirements.

 

According to the document, the regions targeted by the ban are Chinese Mainland, Hong Kong and Russia. However, Nvidia does not have paying customers in Russia.

 

Nvidia stated that this move may affect the development of H100 and the ability to support existing A100 customers, and may further require the company to transfer certain businesses from China.

 

2022CTC大會(huì)H100發(fā)布

 

Currently, Nvidia is in contact with the US government in an attempt to seek exemptions for the company's internal development and support activities.

 

A spokesperson for Nvidia also stated, "We are discussing countermeasures with Chinese customers to meet their planned or future demand for alternative products. If customers do not require sufficient alternative products, we will seek exemptions or licenses for them, but we cannot guarantee that the US government will approve or respond in a timely manner."

 

Affected by the ban, Nvidia's stock price fell nearly 6.6% after opening.

 

 

On August 24th, Nvidia had expected sales for the third quarter to be approximately $5.9 billion. However, with the ban in place, Nvidia's potential sales in China for the third quarter could potentially lose $400 million if customers do not want to purchase replacement products and the US government does not timely or refuses to issue licenses to important customers.

 

In addition, according to a spokesperson for AMD, the US government has also notified them to stop exporting top-notch artificial intelligence chips to China.

 

AMD stated that the new licensing requirements will prevent MI250 chips from being shipped to China, but MI100 chips should not be affected.

 

AMD believes that the new regulations will not have a substantial impact on its business. But the stock price still fell after the opening.

 

 

對國內(nèi)影響有多大？

 

In recent years, the United States has not only continuously strengthened export restrictions on chips against China, but also attempted to bring manufacturing back home.

 

Biden has officially signed the Chip and Science Act, the US Department of Commerce has cut off equipment for advanced process chip manufacturing below 14nm from China, implemented export controls on EDA software tools, and the US is facilitating alliances with other countries.

 

The current export ban on chips by the United States on Nvidia and AMD will not only hinder Nvidia's $400 million business in China, but also affect research in the domestic AI field.

 

 

However, currently, the impact on the consumer electronics industry is not significant. In addition to these two chips, chips such as Qualcomm, MediaTek, and Samsung are also available.

 

Intelligent computing

 

However, most domestic servers cannot do without both of these chips. Obviously, this supply cut by the United States is aimed at computing power, with the aim of hindering China from leading the world in the field of artificial intelligence.

 

Building a powerful intelligent computing center requires high-end GPUs. Without intelligent computing, many fields such as smart transportation, smart cities, and industrial connectivity will be affected.

 

Therefore, chip supply interruption will directly affect the development of the domestic cloud computing industry and artificial intelligence industry.

 

According to the latest global supercomputing TOP500 list in June, the top ten supercomputers mostly use AMD, Nvidia, and Intel processors or technologies.

 

Among them, the Shenwei the Taihu Lake Light supercomputer in China uses the Shenwei 26010 multi-core processor that we independently developed. The Tianhe-2 supercomputer uses the Xeon Phi 31S1P coprocessor based on Intel's integrated multi-core architecture.

 

Therefore, the discontinuation of AMD and Nvidia has not had an impact on the two most influential supercomputers in China.

 

 

However, for domestic enterprises that provide servers, they are not so lucky. The servers are commonly used by enterprises, such as data centers and cloud computing.

 

At present, domestic companies such as Alibaba are starting to develop their own cloud native processors, which is worthy of recognition.

 

元宇宙

 

You should know that GPUs are the foundation of the core computing resources in the metaverse. In the future, the size of the metaverse market may exceed 470 billion US dollars.

 

There are a large number of programs in the metaverse that require computation, and the virtual content, blockchain networks, and artificial intelligence technologies that make up the metaverse cannot do without the support of computing power.

 

Without strong computing power, the metaverse is like a castle in the air.

 

Therefore, a more realistic modeling and interaction in the metaverse requires stronger computing power as a prerequisite, which further illustrates the importance of our self-developed GPU.

 

 

 

AI Big Model

 

In addition, without chips from companies such as NVIDIA and AMD, domestic tasks in image, speech recognition, and other machine learning will also be seriously affected.

 

The parameter scale of the GPT-3 that caught fire in 2020 reached 175 billion, and its training dataset size also exceeded 500GB.

 

Training such a large AI model will inevitably consume a lot of computing power. GPT-3 alone consumed 10000 GPUs and took 30 days to complete.

 

So, how will cutting off high-end GPUs affect the training of domestic AI models?

 

Taking NVIDIA A100 as an example, its deep learning performance can reach 3.5 times that of V100 in actual testing in 2021. In the latest AI chip scoring competition results, A100 broke 16 performance records.

 

Compared to its predecessor A100, which had only 54 billion transistors, Nvidia installed 80 billion transistors in the H100 and adopted a customized TSMC 4nm process.

 

In terms of computing power, the performance of H100's FP16, TF32, and FP64 is three times that of A100, with 2000 TFLOPS, 1000 TFLOPS, and 60 TFLOPS, respectively. In addition, H100 has added support for FP8, with a computing power of up to 4000 TFLOPS, which is six times faster than A100.

 

In AI training, if H100 is used to train GPT-3, the speed can be increased by 6.3 times. If combined with new accuracy, chip interconnection technology, and software, it can be increased to 9 times.

 

 

In addition, with the support of the new Hopper architecture, training of large models can be completed within days or even hours.

 

Among them, the Transformer Engine can package and process FP8 data at twice the speed of FP16, so each layer of the model that can be processed with FP8 can increase the speed by twice.

 

 

If the United States really loses supply of high-end GPUs, is there a feasible alternative to domestic GPUs?

 

Are domestic GPUs ready?

 

In a more detailed summary, Xinzhixun stated that overall, domestically produced GPUs are still in an early stage of development.

 

In the field of graphics cards, only Jingjiawei has achieved some results, but there is still a significant gap compared to Nvidia and AMD.

 

For the high-performance computing field affected by this power outage, the article analyzes that only Tiantian Zhixin, Biren Technology, and Xindong Technology have launched corresponding products.

 

According to official data released by Bi Ren Technology, the BR100 has surpassed NVIDIA's A100 in terms of AI computing power. However, as the product has just been launched, it still needs to be tested by the market.

 

The United States has been continuously suppressing domestic demand, and domestic GPU manufacturers are also accelerating their own research and development in order to achieve self-sufficiency in critical moments.

 

Now, this self-developed path still has a long way to go.

 

 

 

Reference materials:

https://mp.weixin.qq.com/s/V7X2gRVZchHVg9WTrI_P-g

https://laoyaoba.com/n/831077

https://www.reuters.com/technology/nvidia-says-us-has-imposed-new-license-requirement-future-exports-china-2022-08-31/

https://www.reuters.com/technology/amd-says-us-told-it-stop-shipping-top-ai-chip-china-2022-08-31/

Translated from:https://mp.weixin.qq.com/s/FzGMX_s0-KHS1mk7lvMMDw