Today I want to discuss a very niche topic—graphics cards and artificial intelligence.
As you may have noticed, the 2023 World Artificial Intelligence Conference (WAIC) is being held in Shanghai these days. During this conference, Elon Musk opened the event, and AMD’s Lisa Su, along with numerous Nobel Prize and Turing Award winners, frequently mentioned a term—computational power, which fundamentally relates to graphics cards.
As a crucial hardware foundation, it can be said that graphics cards are pivotal to the prospects and future of artificial intelligence.
You might find it strange that a device primarily used by gamers has evolved to become a key player in artificial intelligence.
Nowadays, major manufacturers looking to train AI models first stock up on graphics cards. Recently, there was news that the Wall Street Journal reported that the U.S. is expected to ask American chip companies to stop selling AI chips to Chinese customers as early as this month. Since GPUs are essentially a type of chip, this means that starting from the second half of this year, we might face a graphics card crisis.
You may not realize that graphics cards are indeed vital for the future of technological development. Some even say that the decisive factor on future battlefields is closely related to graphics cards, which is somewhat magical.
For readability, this article will not differentiate between “graphics card,” “GPU,” and “GPGPU”; all will be referred to as “graphics card.”
Initially, there were no graphics cards, and of course, there were no video games. After the first video game was launched at MIT in 1961, nobody expected that people would enjoy gaming so much. The early gaming enthusiasts were precisely the early computer pioneers.
For instance, everyone knows the Unix operating system, which was originally developed by the great Richard for playing a space battle game, later becoming the operating system for the U.S. military, universities, and research institutions.
And then there’s Elon Musk, who created a small game through programming at the age of 12, and even now, he is still a die-hard fan of “Elden Ring” and “Cyberpunk 2077.”
However, the earliest games did not run on graphics cards; at first, there were no graphics cards, and the graphics were quite simple, requiring no such hardware.
It wasn’t until the 1990s that as games grew larger and graphics became more detailed and began to transition to 3D, the demands on CPUs increased. However, CPUs are not adept at graphics processing. Many realized that there should be dedicated hardware to handle this, improving efficiency and experience, and that people would likely buy it.
Driven by this notion, dozens of companies quickly emerged, and the graphics card industry entered a phase of fierce competition.
During this process, three young men decided to make a big impact at a remote fast food restaurant in San Jose, leading to the establishment of a small company that was so under the radar at the time that nobody noticed it, but it later became famous and is now valued at over a trillion dollars—NVIDIA.
NVIDIA did not emerge as a leader immediately; it struggled to survive, working with the super enterprises of the time to provide better graphics processing chips.
Its sole aim was to develop products that gamers would be willing to pay a premium for, enhancing user experience. Thus, there were no complex business battles; it was purely about continuously hiring top talent, developing more powerful graphics cards, and providing better gaming experiences.
From today’s perspective, NVIDIA’s success echoes Ren Zhengfei’s words, “doing what the market needs them to do,” meaning continuously researching products that gamers are willing to spend money on.
As we know, while male consumption power may not be great, in terms of graphics cards and gaming, men have indeed supported a gigantic sector.
If NVIDIA has any decisive strategy, it would be Jensen Huang’s deep partnership with Microsoft in the late 1990s after a period of confusion.
Do you remember when we were kids, there seemed to be a lot of dedicated gaming consoles like the VTech and the Atari arcade games? However, starting from university, gaming mainly shifted to computers, primarily on Windows operating systems.
Playing games on Apple computers was simply unthinkable, and as for Linux, it was mainly used for office development and server deployment, meaning most people rarely saw Linux, yet it’s almost impossible to avoid it daily; when you open an app on your phone, the servers that those apps request data from are mostly running on Linux.
Without a deep partnership with Microsoft, NVIDIA could not have reached its current trillion-dollar scale because Microsoft’s user base is too large. This immense user base provides enough purchasing power for NVIDIA to sell enough products, earn money, and invest in the next round of research and development, solidifying its position.
Conversely, without NVIDIA, Microsoft’s operating system would not have such a wide audience, as Microsoft’s efficiency in office tasks cannot compare to that of a MacBook. Consequently, many of us tech enthusiasts typically have a basic setup: using a Mac for work, cloud computing for coding, and a desktop equipped with a Microsoft operating system and NVIDIA graphics card for gaming. For many, Microsoft’s operating system is primarily for gaming.
In other words, Intel, Microsoft, and NVIDIA have mutually achieved a grand slam.
The typical process is that when a new graphics card is developed, people buy it, install it, and play games for two years, then game developers create new games that current graphics cards cannot support.
Gamers become restless, urging Huang to release higher-performing graphics cards. Generally, higher performance also means higher prices, but gamers are willing to spend, allowing Huang to borrow money from investors for R&D, knowing that once developed, people will buy it, enabling him to recoup the investment.
It is precisely because gamers are always willing to spend money that NVIDIA has more funds to develop stronger graphics cards, forming a positive feedback loop that has lasted over two decades, burning through hundreds of billions in funding, eliminating countless competitors, and achieving the current scale, showing no signs of stopping. Today, the graphics card landscape primarily consists of NVIDIA and AMD, and from a market scale perspective, NVIDIA essentially stands alone.
Game graphics have evolved over decades, from this:
to this:
So the question arises, how did gaming graphics cards become related to artificial intelligence?
Let me tell you a story that our teacher mentioned years ago when explaining the difference between graphics cards and CPUs.
In 1979, during the Iranian protests, demonstrators stormed the U.S. embassy and took 66 embassy staff hostage. When checking the numbers, they found a few were missing, but all the personnel data had been shredded by the Americans, turning into fine strips of paper.
The Iranians, undeterred, employed thousands of elementary school students to compare the strips one by one, ultimately restoring the original data, creating a new chapter in brute-force decryption. This event was later adapted into the film “Argo.”
At this point, it becomes clear what the difference between a CPU and a graphics card is. A CPU is like a top university student who can assemble and debug a very complex imported device in a lab, but does it have an advantage when it comes to calculating a sack of arithmetic operations?
Clearly not; in this case, a group of elementary school students who have just mastered basic arithmetic would outperform the CPU, just as the Iranians did with their puzzle-solving, relying on brute force.
Since the advent of artificial intelligence, today’s graphics cards seem omnipotent, but the reality is that their basic principle is simply a bunch of elementary students capable of performing large-scale parallel computations.
Let me explain a bit; you will immediately understand what graphics cards are calculating. The essence of playing games is that the graphics card continuously draws images for you. For instance, the familiar “frame rate”—100 frames mean the graphics card draws a hundred pictures in one second, one after another, making the video appear very smooth.
However, this “drawing” does not literally mean painting on a canvas but involves a series of complex yet straightforward graphical computations. These graphical computations mainly consist of matrix operations. You might be confused by the term “matrix operations”. Let me show you a random diagram:
This is “linear algebra,” typically taught in the first year of engineering and science majors. For those who have studied it, this is merely elementary arithmetic; it may look complex but is actually just a series of addition and multiplication, very tedious. However, this type of operation is very common; for instance, the GPS we use daily essentially involves such matrix operations.
You must have heard about how a bunch of university students from Harbin Institute of Technology helped with the development of China’s first atomic bomb, right? A professor at my university participated in those projects, and essentially, it was just manual calculations of calculus—transforming differential equations into matrices, followed by tedious addition and multiplication operations, where the abacus’s advantages became apparent.
In other words, the primary advantage of graphics cards is that they are packed with a lot of elementary students who excel at multiplication and addition. After decades of evolution, with gamers voting with their wallets, graphics cards have continuously advanced, and “parallel brute force” has become increasingly impressive. However, initially, people did not understand that they had other uses beyond gaming until someone started using them for mining.
Mining, which involves extracting Bitcoin or Ethereum and other cryptocurrencies, fundamentally consists of calculating a very complex and tedious encryption algorithm, and graphics cards are naturally suited for such brute-force operations, leading NVIDIA to ride this wave and see its stock price soar during the pandemic.
However, even NVIDIA’s management did not anticipate that their graphics cards could be used for other purposes.
Around 2011, NVIDIA, already a dominant player in the gaming industry, suddenly learned of something they had never expected.
It turned out that MIT was conducting some of the earliest artificial intelligence training, and researchers wanted to enable computers to quickly recognize images of cats.
How to recognize them? They could only process pixel by pixel, involving a massive amount of repetitive and tedious calculations. Initially, MIT researchers used CPUs to handle this, but the results were poor. As mentioned earlier, a CPU is like a brilliant graduate student who can perform very complex tasks but can only do one thing at a time. When faced with a sea of arithmetic operations, it feels like “having strength but no way to use it,” resulting in extremely slow calculations.
Eventually, they began using GPUs, or the graphics cards we know, which allowed for a parallel computing approach, akin to the Iranians directing a group of elementary students to solve a puzzle, known as “parallel computing,” which is naturally suited for “simple and large-scale” operations.
Thus, graphical display, artificial intelligence, and decryption algorithms are fundamentally all about “brute force operations.” Now you understand why graphics cards can be used in aerospace, weather forecasting, energy exploration, etc., because these fields all involve super-large-scale matrix operations. As long as it involves super-large-scale brute-force calculations, it’s the domain of graphics cards.
This is why many people say that future battlefields will also belong to graphics cards, especially in air combat—how fast can human reactions be? How fast can those intelligent weapons equipped with graphics cards react, quickly gathering information and computing to make evasive maneuvers or trigger actions? Just thinking about it is chilling.
Therefore, NVIDIA also has an important business in “data centers,” where they create large server rooms to provide computational power to others for a fee.
It is precisely because of the widespread application of NVIDIA graphics cards in gaming, data centers, and artificial intelligence that the market outlook is promising, making NVIDIA a member of the trillion-dollar club today.
Yes, although gaming still retains a somewhat frivolous impression in the public’s mind, the application and evolution of artificial intelligence are closely related to gaming. In fact, the first product that overturned human understanding of artificial intelligence was born through a “Go battle.” Nowadays, AI graphics cards are an important foundation for training artificial intelligence.
At this year’s WAIC, research institutions also released reports on gaming and artificial intelligence, analyzing their symbiotic development relationship from three aspects: theoretical research, graphics card hardware iteration, and application scenarios. Xiamen University, in collaboration with several universities such as Communication University of China, Central Academy of Fine Arts, Beijing Institute of Technology, and Shanghai Jiao Tong University, is preparing to establish a “Joint Research Center for Game Artificial Intelligence.” That’s great.
Some may ask, what does this have to do with us? Can China quickly catch up?
That’s a narrow perspective. China’s greatest advantage is its vast consumer market, and market demand often gives rise to limitless possibilities. If cultivated well, under the global wave of artificial intelligence, we can certainly occupy an important position. This might be why Musk passionately expressed at the WAIC opening ceremony, “I believe China will possess very strong artificial intelligence capabilities in the future.”
Conclusion:
The rise of artificial intelligence demonstrates that:
The consumer market is the breeding ground for hard science. The essence of the consumer market is the human longing for a better life. After all, everyone works hard to earn money, save to buy things they like, and use smarter tools, primarily because these things bring joy or solve problems. This longing for a better life is the most primitive driving force behind technological progress.
Moreover, we can see that often greatness is not planned but grown. The folks at PS didn’t know what their product would become, and ASML, which makes photolithography machines, couldn’t replicate TSMC’s processes. Jensen Huang never expected that his gaming cards would one day become the cornerstone of artificial intelligence.
Of course, while discussing games and graphics cards, the reality extends far beyond gaming; many fields work this way. You cannot predict what it will become in the future; no one can say which fruit will drop from which branch and create a new industry.