Ancient civilizations hunted for spices; in the 20th century we fought wars for oil. In 2023, the world's most valuable commodity will be an envelope-sized computer chip.
The H100, a rectangular black maze of circuitry with a shimmering microchip in the center, is niche hardware even in Silicon Valley. boring standards.
Its $40,000 price tag is exceptional for a computer processor, although even if you have the money, good luck with it. The half million H100s expected to be released this year are already sold out.
It's not like one chip will do you much good. Last week it became known that Saudi Arabia ordered about 3000 H100.
Nearly all of the chips are being snapped up by the world's richest companies: Microsoft, Amazon, Google, and Mark Zuckerberg's Meta are thought to account for the vast majority of demand.
Joe Biden banned their export to China under rules designed to prevent secret technology from falling into the hands of the enemy.
Though more than a trillion microchips—the brains of our electronics—are produced each year, the H100 is at the center of the global race for intelligence.
China's semiconductor manufacturing equipment imports by country
The chip is described by its manufacturer, Silicon Valley giant Nvidia, as being the first in the world to be designed for «generative» artificial intelligence, a type of powerful, «creative» AI that powers technologies such as ChatGPT. that took the world by storm.
This year, the advent of systems like ChatGPT has reignited warnings that AI should be taken as seriously as nuclear weapons and climate change.
Its creators were called to Downing Street and the White House and threatened with new rules.
The strikes that have brought Hollywood to a halt, for example, are partly because actors and writers fear they will be replaced by artificial intelligence copies that require no wages.
How the chip works
Enter a query in ChatGPT, and it will confidently give a competent answer in a split second. However, the hard work is done by thousands of Nvidia chips on a huge, humming server farm hundreds of miles away.
The AI boom has turned the California-based company into one of the $1 trillion-plus elite companies, with founder Jensen Huang becoming the 32nd richest person in the world with a net worth of $39 billion.
Despite the emergence of competition, Jensen Huang's Nvidia dominates to the point that it is often considered «the only game in town». Photo: I-Hwa Cheng. /Bloomberg
The position of the H100 as the key to the AI boom is really a fluke. For most of its 30 years, Nvidia has been making chips known as GPUs used to support 3D graphics in video games.
Traditional computer processors, known as central processing units, can perform complex calculations, but usually only one at a time. In contrast, GPUs can perform hundreds of simple calculations at the same time.
This helped them with tasks like changing color pixels in a video game. What the company didn't count on was that its chips would also be ideal for the fast data processing needed for artificial intelligence.
In 2012, Google researchers used 16,000 processors to train rudimentary AI software that could recognize images of cats. A year later, the researchers achieved the same results with four GPUs.
“Part of this is incredible foresight, and part of it is just blind luck,” says Jay Goldberg, a Silicon Valley chip manufacturing expert.
>“The way AI works is the same mathematical equation, just repeated over and over again. again in large numbers. And GPUs are really good at it.”
Despite luck, people are now buying chips as fast as Nvidia can make them. “GPUs are a lot harder to get than drugs right now,” Elon Musk told a conference in May.
While competing companies make GPUs, one tech investor describes Nvidia as “the only game in town.” And for those who want the best, only the H100 will do.
The current model that runs ChatGPT, for example, was developed using about 25,000 A100, the chip's predecessor.
The new version is an order of magnitude more powerful: in 20 hours it can do what would have taken the A100 a week.
For all the “intelligence” that is being attributed, the best AI systems are really the product of the best training, and it really is a matter of having the best computer.
In the first version of the ChatGPT model, there were about 100 million variables known as parameters; the latter is believed to have trillions.
Therefore, it is not surprising that the search for the best chips has turned into an act of desperation. “There is one thing that can compete with H100 GPUs: Taylor Swift concert tickets,” Aravind Srinivas, an AI entrepreneur, tweeted.
One online commentator calls it «GPU OPEC».
Silicon Valley bosses trade stories about their efforts to get legendary chips: some use their well-connected VC investors to contact, others beg friends at universities to let them connect to their research labs.
Yes, however, this is only one guaranteed way to get stable supplies: hundreds of millions of dollars. This means that there are currently two main camps of customers who can buy chips in the quantities they need: the tech giants and the nation states.
Global race for AI supremacy
In a few years, AI has moved from a promising technology to an essential national priority.
In 2017, China laid out a «next-generation artificial intelligence blueprint», promising to make the country a «major» center of innovation by 2030. According to the US Center for Security and Emerging Technologies, Beijing aims to promote «artificial general intelligence.» , a system capable of rivaling human intelligence.
The plan was met with a volley of US sanctions designed to undermine China's AI efforts.
Last year, the White House banned Nvidia from selling its H100 and A100 chips to China without a license. The company has been quick to release new versions of the chips to bypass controls, and the Biden administration is now expected to respond with a new set of rules.
Saudi Arabia and the United Arab Emirates are investing heavily in chips developing their own «big language models» like ChatGPT.
According to the Financial Times, Saudi software, developed at the oil-rich state's Kaust Research Institute, is mostly created by Chinese researchers whose education prevents them from working in America.
In the US, the Biden administration is spending tens of billions of dollars to encourage microchip companies to set up factories in America. The plan has been called «one of the biggest government interventions since World War II.»
Nvidia may be a US company, but its most powerful chips are made entirely in Taiwan, leaving the entire AI industry dangerously vulnerable to a threat Chinese invasion.
1904: Chip production in the West declines
AI weapons are a clear example of how important new technology is. Autonomous drones that can see the world around them and react accordingly have already changed the battlefield in Ukraine, and artificial intelligence has successfully controlled F-16 fighter jets.
But the risks are wider: generative AI programs like ChatGPT threaten massive cyberattacks and the industrialization of fake images.
Mike Wooldridge of the Alan Turing Institute says widespread misinformation is the biggest immediate risk associated with AI.
In 2016, Russia was accused of interfering in the election of Donald Trump using an army of Internet trolls in Moscow offices. By next year's elections, AI will likely be able to create an endless supply of divisive online comments and fake images that are indistinguishable from the real thing.
However, old-fashioned economic conflict is also a concern.
Sam Altman, CEO of OpenAI, the company behind ChatGPT, predicted that once AI software reaches a level of sophistication where it can improve itself, it will start a «recursive cycle of innovation» in which it will create «phenomenal technologies.» wealth,” not to mention instant scientific breakthroughs.
Sam Altman, CEO of OpenAI, predicts bots will reach critical mass, after which they will learn on their own Credit & Copyright: JOEL SAGET/AFP
Not everyone is so optimistic, but even less radical predictions about how AI will change economy will come true, countries will not want to be left behind.
Britain is left behind
And Britain? Five years ago, British tech investor Ian Hogarth predicted that the development of machine learning systems would lead to what he called “AI nationalism.”
Hogarth suggested that the rise of AI would mean an «accelerated arms race» leading to a more closed economy, preventing foreign takeovers and limiting the exit of talent, adding that «AI policy will become the single most important area of public policy.»
He warned that the UK was misguiding its own AI strategy by allowing the sale of companies like London-based Deepmind, one of the world's leading AI labs, which Google bought for a comparatively paltry £400m in 2014.
If five years ago his ideas seemed outlandish, now they are taken seriously. This year, Hogarth was named chairman of the Rishi Sunak Artificial Intelligence Foundation's task force, set up to encourage cutting-edge security research.
The UK has arguably founded one of the world's leading AI companies, and last year's government review described the field as «high-level economics, security, health and wellness.»
Number of A100 GPU instances worldwide
But when it comes to the processing power needed for today's technology, experts warn that we're dangerously behind. A survey of UK computing capacity published this year found that less than 1,000 high-performance Nvidia GPUs are available to researchers in the UK — just a fraction of what it took to train the latest ChatGPT model.
The most powerful British computer Archer-2 was developed four years ago and did not foresee the AI boom.
“The UK is currently not well-prepared for large-scale GPU-centric computing,” says Mark Parsons, chief executive of the Edinburgh Parallel Computing Centre, the UK's leading supercomputing center where Archer-2 is based. “We just don’t have that capability anywhere.”
An industry source warns that in a few years “the economy will have a computing crisis that will be as devastating as the energy crisis of the past year.”< /p>
The government allocated £. 900m for a new supercomputer and research infrastructure, although they are unlikely to be online for at least a year, and in terms of AI, for life.
Instead, Rishi Sunak focused on other areas, such as putting the UK at the center of a global effort to contain the risk of AI escaping human control.
Supercomputer performance ratings
One option likely to be discussed at the global security summit in November is a global agency in the UK, modeled after the International Atomic Energy Agency, a body created at the start of the Cold War to prevent nuclear winter.
Preparations for the summit are behind schedule but moving fast, with officials debating at what level China will be invited.
New superpowers
However, all this may be secondary. Unlike nuclear weapons, it is not nation states that are at the forefront of AI development, but a handful of giant technology companies.
Google will spend $33 billion on buildings and equipment this year, with spending on infrastructure like Nvidia chips rising by about a third.
Netflix — Derided in an episode of Black Mirror. who envisioned a future in which AI holograms would replace actors is advertising for machine learning jobs that pay up to $900,000, more than four times the prime minister's salary.
If the power of AI is a product of the computing power behind it, national governments pale in comparison to the resources available to private enterprises.
Inflection AI, a startup run by DeepMind co-founder Mustafa Suleiman and funded by Microsoft, is building an artificial intelligence supercomputer powered by 22,000 H100 chips.
Amazon, Microsoft and Google, the so-called «hyperscalers» that run the data centers that power an increasing portion of the world's IT, are lining up orders worth tens of thousands each.
The meta could demand 100,000, according to one analysis.
Goldberg says the «super seven» tech giants, including three Chinese companies Baidu, Alibaba and Tencent, are responsible for bringing Nvidia's top-end chips to 100% demand.
Altman is arguably the most prominent figure in the field AI today, talks about OpenAI in the same language that is commonly used for nation states. He says that in the future its leader may be democratically elected.
At the moment, although they are accountable to shareholders and employees, this has led to questions about which side these companies really are on.
In 2019, Microsoft employees spoke out against the company's work with the US military, writing that «we didn't sign up for weapons development.»
Last year, protests at Google forced the company to stop work on a contract to supply drones to the Pentagon.
Alex Karp, chief executive of Palantir, the AI company used by the CIA to hunt down Osama bin Laden. , recently warned that such hesitation would be comparable to Robert Oppenheimer's refusal to build an atomic bomb, and that no such predicament existed in China.
“Our indecision, apparent or not, regarding the further use of artificial intelligence for military purposes will be punished,” he said.
Perhaps the time for a decision is running out.
Demis Hassabis, founder of DeepMind and Britain's most prominent leader in artificial intelligence, says that artificial general intelligence, the point at which artificial intelligence advances at potentially unstoppable speed, could happen within a decade.
As for the company behind this revolution, this month Nvidia introduced an even more powerful version of the «GH200». The world is vying for the superchip at the heart of the AI race, but it could soon be yesterday's news.
Свежие комментарии