Shulou(Shulou.com)11/24 Report--

Chips are crucial to artificial intelligence. Su Zifeng, CEO of AMD, saw an once-in-a-lifetime opportunity.

Compile: Poetry |

Author: David Meyer |

After decades of software domination, the "silicon" part of Silicon Valley is back on the stage. It turns out that it takes a lot of hard-core hardware to make the miracle of generative artificial intelligence a reality, and chipmaker Nvidia seized this opportunity with its powerful graphics processor to become the market leader.

Demand for Nvidia's AI-friendly processors was so strong that investors valued the company's stock market at more than $1,000bn in May, roughly equivalent to Saudi Arabia's GDP in 2022. In an artificial intelligence-driven economy, it is true that chips may become as important as oil, but in the fast-changing technology industry, even leaders still have to be constantly prepared to rely on proven reserves to remain dominant for a long time.

For Jensen Huang, chief executive of Nvidia in a leather jacket, the most serious threat comes from a chipmaker across town. The chipmaker has the unique advantage of both graphics processing capabilities and the corporate image formed by competing with giants over the years. Under the leadership of Lisa Su, AMD aims to occupy a significant share of the artificial intelligence chip market, and as the artificial intelligence revolution begins, even replacing Nvidia as the industry leader.

"I think this is an opportunity for us to write a new chapter in the AMD growth story," Su Zifeng said in an interview with Fortune magazine in mid-September. "globally, there are very few companies that have (intellectual property) and customer base, and frankly, we have the opportunity to really shape the way artificial intelligence is applied globally. I think we have such an opportunity."

Su Zifeng has good reason to talk about the opportunities of AMD in the artificial intelligence chip market, which she predicts will be worth $150 billion by 2027. AMD is perennial runner-up in the PC and server microprocessor market (led by Intel), but it has become the world's second-largest GPU maker because of its acquisition in 2006 of ATI, a Canadian chipmaker specialising in video game accelerators (the chip is ideal for training artificial intelligence models such as OpenAI's GPT-4 and Google's Bard).

Nvidia is estimated to have more than 90 per cent of the artificial intelligence training market, but the company is struggling to meet demand for its most powerful artificial intelligence chip, the H100, and Su Zifeng is preparing to launch a direct attack on MI300, a competitor to the H100, this quarter.

Brian Colello, head of technology at Morningstar, said: "it is certain that in 2024 there will be a scenario where Nvidia's GPU is sold out, customers can only use AMD's GPU, and AMD can win business based on availability alone."

"for Nvidia's valuation (worth trillions of dollars), there are two questions that need to be addressed: how big the market is and whether Nvidia can absolutely dominate." Collow said. If Nvidia can capture 95 per cent of the artificial intelligence market, far away from the second AMD, that is one thing. But he points out that if it is a score of three or seven, "that would be very good for AMD."

As for how to divide market share, it boils down to competition in terms of performance, flexibility and availability (in the context of supply chain uncertainty), all of which are time-tested areas for Su Zifeng.

Distant relative and close competitor when Huang Renxun, a sometimes domineering Ferrari driver, runs the company he founded, Su Zifeng-thoughtful, outspoken and humane-has entered her 10th year at the helm of an established company, which is widely believed to have saved it.

Su Zifeng emigrated to the United States from Taiwan at the age of three. She and Huang Renxun are both from Tainan and are actually distant relatives. Su Zifeng later became a highly respected electrical engineer at IBM, in charge of the emerging products department. After a stint as Chief Technology Officer of Freescale Semiconductor (Freescale Semiconductor), Su Zifeng joined AMD in 2012 as Senior Vice President and became President and CEO two years later.

Although Su Zifeng's initial strategy to save AMD included expanding its product line from the PC market to areas such as games and high-performance computing, the company still needed a lot of effort to compete with Intel. This may be inevitable, but analysts say it has created Nvidia's dominant position today.

Nvidia has invested heavily over the past decade to make it easier for developers to take advantage of the parallel processing power of its GPU when building data-processing applications through the CUDA interface, but AMD's competition with Intel in CPU means "it doesn't invest as much money in artificial intelligence software in the data center GPU." Said Alan Priestly, an analyst at Gartner.

"the city wall of Nvidia is its software ecosystem." Priestly pointed out. Compared with CUDA, AMD's ROCm software stack is famous for its many vulnerabilities and difficult to upgrade.

Gregory Diamos, co-founder of Lamini, an artificial intelligence start-up and a former CUDA architect at Nvidia, says he thinks AMD is closing the gap. "AMD has put in hundreds of engineers to support the general artificial intelligence program," Demos said. "

But even Su Zifeng admits that there is still a lot of work to be done. "I will not hesitate to announce that our hardware is excellent and our software is constantly improving." "for some AI applications written in the past, it really takes some work to migrate them to AMD," she said. However, Su Zifeng believes that ROCm has made "very good optimization" of the newer AI workloads.

The next stage of artificial intelligence the natural development of artificial intelligence may bring great opportunities to AMD.

Artificial intelligence companies prefer GPU because it can perform multiple tasks at once. The computing power required to render rich, fast-moving graphics and images in video games can be easily used to train large language models, such as OpenAI's GPT-4, to process large amounts of raw data in a relatively short time.

But many analysts believe that the larger share of the artificial intelligence market is not the training of large language models, but the application of large language models: as artificial intelligence becomes a part of daily life, systems need to be built to answer billions of queries that are expected to appear. This is the so-called "reasoning" (because it involves artificial intelligence models using their training corpus to infer the new data presented), and it is an open question whether GPU is still the preferred chip for reasoning.

So-called super-large companies such as Meta and Google are already trying to develop internal artificial intelligence chips such as TPU, and these big companies will eventually need professional chips that provide efficient artificial intelligence services. Many also believe that CPU will play an important role in the reasoning market, a shift that will provide opportunities for AMD to play to its traditional strengths.

The upcoming MI300 series of data center chips combines CPU and GPU. "We actually think we are going to be an industry leader in reasoning solutions because of our architectural decisions," says Mr Su. "

Morningstar's Collow also points out that the market is changing, and that doesn't include efforts by Intel, AMD's arch-rival, to challenge Nvidia to launch new artificial intelligence processors Gaudi2 (for training) and Greco (for reasoning). "for all these companies, there is naturally enough motivation not to be beholden to Nvidia and want more competitors to write software, transmit models and take all necessary measures to ensure the health of the ecosystem," he said. "the ecosystem includes Nvidia, AMD and possibly Intel, as well as their own internal chips being developed."

Nvidia said it was "highly focused" on reasoning. The company's GPU reasoning performance has improved eightfold in the past year, a spokesman told Fortune magazine, Nvidia is "investment reasoning roadmap."

The company also said it realized that it would not always be the only option, and that customers naturally wanted to compare goods. A spokesman for Nvidia said: "A competitive ecosystem is positive in the field of artificial intelligence because it can accelerate technological development faster and more effectively, and we certainly encourage and welcome competition."

Su Zifeng's goal is not only to build a data center with artificial intelligence through the cloud, but also to make artificial intelligence run directly on personal devices and other networked gadgets. "the key to success is that we do account for a large part of the use of artificial intelligence computing," she told Fortune. "

Of course, Nvidia is also targeting these segments and is now even trying to get into the high-performance CPU business through a new "super chip" called Grace, bundling it with the H100 GPU. "Nvidia still wants to dominate the reasoning market, and they are likely to do so," Collow said. for any investor bullish on Nvidia, they may assume that all types of artificial intelligence processes will run through Nvidia's chips and / or networks. "

But even if Nvidia remains in the lead, Collo believes that AMD's strong position in second place is "very enviable and will bring considerable business." Su Zifeng is confident that her company will take full advantage of the explosive growth of artificial intelligence.

Su Zifeng said: "it is clear that with the popularity of generative artificial intelligence in 2022, there is room for this [industry] to grow at an incredible rate. We expect a compound annual growth rate of 50% over the next five years, and when you talk about the scale of tens of billions of dollars, few markets can do that."

This article comes from the official account of Wechat: new Research (ID:chuxinyanjiu), author: Poetry

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.