In the Warring States era of AI chips, giants "encircled and suppressed" Nvidia: Microsoft, Intel, Huawei, who can share the pie

NVIDIA, which has enjoyed the dividends of AI, finally has a sense of crisis. Shortly after the launch of its new generation of GPU H200, on November 16, Microsoft immediately launched its first self-developed AI chip, plus Intel, Google, AMD, these opponents who also covet the cake of AI chips, I have to say that Nvidia has been forced to the "bright top", and the battle of giants is about to break out.

In the domestic market, due to the restrictions on the A800 and H800 due to the escalation of the chip ban, Huawei's Ascend has ushered in widespread application and has also become a potential threat to NVIDIA. In the eyes of industry insiders, the AI chip market will show a trend of diversification in the future, and it will no longer be a dominant pattern, especially under the test of geopolitics, Chinese AI chip companies also have the opportunity to go overseas and occupy a place in the global AI wave.

New product competition "on the job"

On November 16, Microsoft CEO Satya Nadella released two self-developed chips at the Ignite2023 developer conference - Microsoft Azure Maia 100, a cloud AI chip, and Microsoft Azure Cobalt 100, a server CPU.

Among them, Maia 100 is the first artificial intelligence (AI) chip designed by Microsoft for large language model training and inference in the Microsoft cloud, using TSMC's 5nm process, with 105 billion transistors, optimized for AI and generative AI, and supporting Microsoft's first implementation of sub-8-bit data types (MX data types).

On November 16, local time, Microsoft's stock price rose 1.76% to close at $376.17 per share, with a market capitalization of $2.8 trillion.

Previously, according to media reports, Microsoft is very likely to release a chip codenamed "Athena" at this conference, and this chip has been provided to a small group of Microsoft and OpenAI employees, who are testing the technology, but the final product released is not called Athena, but Azure Maia 100.

Just three days before Microsoft announced the Azure Maia 100, Nvidia also just announced a GPU called the NVIDIA HGX H200. According to reports, H200 is the first GPU to use HBM3e, its faster running, larger memory capacity will further accelerate generative AI and large language models, while advancing scientific computing for HPC workloads, compared with the previous generation of A100, its capacity has almost doubled, and the bandwidth has also increased by 2.4 times, so that the inference speed of large models is nearly double that of H100.

Microsoft's Maia100 will be available in 2024, and Nvidia's H200 will be available through global system manufacturers and cloud service providers starting in the second quarter of 2024.

It is worth mentioning that OpenAI, a major customer of Nvidia, has taken the lead in trialing the Maia 100 chip, which is being tested on GPT-3.5 Turbo. In order to solve the shortage of artificial intelligence chips and save chip costs, OpenAI has previously been revealed to be considering manufacturing its own AI chips, and has even considered potential acquisition targets.

In addition to these, giants such as Intel, Google, and AMD are also "sharpening their knives". Google first announced the details of its own AI supercomputing in April this year, and launched a new AI chip Cloud TPU v5e in August; in June this year, AMD directly released a GPU called MI300X, which is directly benchmarked against NVIDIA's H100, the company claims that MI300X memory is higher than NVIDIA H100 in both density and capacity; Intel has released a high-performance deep learning AI training processor Gaudi 2 in May 2022, and plans to upgrade Gaudi in the near future HBM capacity of the AI chip, and said that the third-generation Gaudi AI chip will be launched next year.

The market landscape is waiting to change

There are good reasons for giants to compete for AI chips – the dividends are attractive.

In this year's artificial intelligence boom set off by generative AI, compared with those artificial intelligence companies, Nvidia, which provides AI chips for various large models, has received the most tangible benefits, in addition to soaring in the capital market, becoming the world's largest semiconductor chip company by market capitalization, and the surge in GPU sales has also helped its data center business set a record revenue in the last two fiscal quarters.

In the capital market, Nvidia has made a lot of money. The company's stock price opened at the beginning of the year at $148.51 per share, rose as high as $502.66 per share, and closed at $494.8 per share on November 16, local time, with a market capitalization of $1.22 trillion.

In terms of performance, in the second fiscal quarter of fiscal year 2024 ending July 30, 2023, Nvidia achieved revenue of $13.51 billion, an increase of 88% quarter-on-quarter and 101% year-on-year, of which the data center division, which benefited the most, reached a record revenue of $10.32 billion in the second fiscal quarter, an increase of 141% quarter-on-quarter and a year-on-year increase of 171%.

In addition, several other data of Nvidia also ushered in significant growth. In the second fiscal quarter, its net profit was US$6.188 billion, up 203% sequentially and 843% year-on-year, non-GAAP net profit was US$6.74 billion, up 148% sequentially and 422% year-on-year, and gross margin was as high as 70.1%, compared with 43.5% in the same period last year, up 26.6% year-on-year.

Such dividends will continue. According to third-party institutions, the AI chip market size will reach $53.4 billion in 2023, an increase of 20.9% over 2022, an increase of 25.6% in 2024 to $67.1 billion, and by 2027, AI chip revenue is expected to be more than double the market size in 2023, reaching $119.4 billion.

Yuan Bo, a senior communications engineer and strategic planning expert, told the China Times that AI is the core technology involved in the fourth industrial revolution, and in the next 5-10 years, the global demand for AI chips will be very large, so from the perspective of market demand, it is natural that the world's mainstream chip companies have devoted themselves to the research and development of AI chips.

So, after the competition of giants, can the AI chip market still be dominated by Nvidia?

"Global AI chips used to be dominant, but from the perspective of the United States' AI industry strategy against China, the global security supply chain structure has been destroyed, and it is no longer safe to use a company's chips to build the country's future core industrial capabilities. Therefore, the AI chip market will show a trend of diversification in the future, which is also an opportunity for China. For China's local AI chip companies, whether it is a national strategy or market prospects, in the case of unstable external supply chains, more Chinese companies will be driven to use localized chips and localized AI training frameworks, and Chinese AI chip companies will also have the opportunity to go out of China and occupy a place in the global AI wave. Yuan Bo said.

Huawei is on the rise

When it comes to China's local AI chips, Huawei's Ascend has been the hottest recently.

Due to the recent escalation of the U.S. chip ban, the A800 and H800 provided by NVIDIA for the Chinese market have also been restricted from sale, and Huawei's Ascend concept has also ushered in a wave of capital heat.

iFLYTEK previously said that the company has launched a special research project with Huawei Ascend at the beginning of 2023 to jointly build a new foundation for general artificial intelligence in mainland China, so that the domestic large model architecture is based on the software and hardware of independent innovation, and the current Huawei Ascend 910B capability has basically been able to benchmark against NVIDIA A100.

However, Yuan Bo believes that the applicable scenarios of Huawei's Ascend chip and Nvidia GPU are different, and in the field of AI machine learning training, its ability has approached or even partially surpassed the level of NVIDIA, but image graphics processing is weak, so it can only make up for some of NVIDIA's scenarios, and China is still very limited in AI. "Nvidia is a general-purpose GPU, and Huawei Ascend is a dedicated artificial intelligence chip, which can only replace NVIDIA in some fields, and at the same time, the existing AI algorithms required for business scenarios need to be re-adapted, and the workload is still relatively large. From this point of view, whether it is AI chips or supporting AI training frameworks, the gap between domestic chips and NVIDIA is still relatively large. ”

Of course, Nvidia will not sit idly by and remain indifferent to the Chinese market, which contributes one-fifth of its revenue. Recently, it was reported that Nvidia has developed the latest improved series of chips for the Chinese market - HGX H20, L20 Pure and L2 PCle, which are improved from H100. The reporter of this newspaper asked Nvidia for confirmation on this, and the other party said that it could not respond to this.

"Even if Nvidia launches an improved model for the Chinese market again, it does not mean that it is safe to use Nvidia's new model of GPU, because the previous improved model is still restricted by the United States, and this risk is still very large. Therefore, in some artificial intelligence fields, especially in some scenarios involving the national economy and people's livelihood and large models of key industries, priority will definitely be given to the use of domestic substitution. Therefore, it is also a once-in-a-lifetime opportunity for Chinese AI chip companies, including Huawei, to promote the maturity of domestic AI chips and AI training frameworks. Yuan Bo said.

Editor-in-charge: Huang Xingli Editor-in-chief: Han Feng