According to Trendforce data, South Korea's SK Hynix and Samsung will control 90% of the world's HBM memory production capacity in 2023.

Therefore, just when the market value of Lao Huang's Nvidia exceeded one trillion, SK Hynix's share price secretly more than doubled last year, and now its market value has exceeded one hundred billion.

01 The history of HBM

Even Samsung, which had chosen the wrong route before, has seen its stock price rise to its highest level in the past two years.

Maybe everyone has no idea how important HBM memory technology is. Let's put it this way, South Korea has listed this technology as a national strategic technology at the beginning of this year, and wants to maintain its leading position by providing tax incentives to SK Hynix and Samsung.

Even when Lao Huang released the latest B200 AI chip, he had already given tens of billions of yuan in deposits to SK Hynix and others, directly covering the output of these factories this year.

You know, SK Hynix just achieved mass production and delivered HBM3e memory at the end of last month. Isn't this much more powerful than the Xiaomi SU7 next door?

Such a refreshing business made Samsung next door so greedy that it hurriedly released its own HBM3e samples in February and immediately sent them to customers all over the world for inspection.

02 So why is HBM memory so popular?

This has to mention the long-term development of memory technology. In the information age, whether it is gaming or work, the running speed of a computer system actually depends on the cooperation of the processor and memory.

Theoretically, if the two speeds are close, they must be the best partners.

But for a long time in the past, the performance of processors has increased exponentially, and the transmission speed between memory and hardware has simply not been able to keep up.

Over the past 20 years, the peak computing power of hardware has increased 90,000 times, but the memory and hardware interconnect bandwidth has only increased 30 times.

To use an analogy that is not very appropriate, the processor is equivalent to the Chinese little master, and the memory is the assistant who serves the dishes on the side. No matter how high the skill of the little master is, if the assistant cannot match the dishes, the food will still not be served quickly.

Therefore, in recent years, memory has become a drag on computer performance. Some people even call this phenomenon a "memory wall."

These problems are quite acceptable for those of us who play games. After all, there must be a lot of people using 1066 to play AAA.

03 The AI Models

What really keeps them busy is actually the explosion of AI in the past two years. Because the foundation of large AI models is inseparable from massive data and computing power, and to support such large data processing and transmission, the "memory wall" must be broken.

In other words, playing games is like ordering four dishes and one soup for the little boss, and casually training the AI next door is equivalent to ordering a full banquet.

Therefore, for major AI manufacturers, the memory wall is like a sophon that locks their own development level, and HBM is the wall breaker.

Different from the "bungalow design" adopted by traditional DDR memory, HBM, the wall breaker, adopts a "building design", which is a dimensionality reduction blow.

Building a building in a chip mainly relies on advanced packaging technology such as Through-Silicon Via (TSV).

To put it simply, different chips are stacked together, a bunch of holes are punched in the middle, and then they are connected with conductive materials such as copper tubes.

In a physical sense, the data transmission distance, floor space, etc. are reduced.

In this way, signal delay can be reduced, low power consumption of the chip can be achieved, bandwidth can be increased, and many other advantages.

As technology levels improve, TSVs can be made smaller and denser, and more and more die layers can be stacked, which can further increase bandwidth, transmission speed and maximum capacity.

As mentioned earlier, Samsung's latest product is the HBM3e with a 12-layer stack capacity of up to 36 GB.

Although it sounds so simple, in fact, HBM technology is difficult in terms of materials, design, packaging, heat dissipation, etc.

Most of these technical difficulties were first overcome by South Korea's SK Hynix.

Therefore, SK Hynix was the first to break through and achieve mass production such as HBM3 and HBM3e.

By the HBM3 generation, a data path of up to 1024 bits has been achieved, the operating rate has reached an astonishing 6.4 Gb/s, and the bandwidth is as high as 819 Gb/s.

The latest HBM3e is actually the fifth generation product of the HBM memory technology family. Its maximum data processing speed has reached 1.18TB (terabytes) per second, which is equivalent to processing more than 200 full high-definition (FHD) movies in 1 second. level movie.

Don't look at the HBM track. There are currently three companies: SK Hynix, Samsung and Micron. However, whether it is in terms of development time, technological breakthroughs or mass production speed, SK Hynix has always been the leader. They alone occupy about half of the market. market share.

Behind him is Samsung, which is also a Korean company. This guy established two HBM teams in one breath this year to catch up with the progress.

Although Micron in the United States has said that it skipped the research and development of HBM3 and directly successfully completed the research of HBM3e.

It has even become one of Nvidia's suppliers now, but Shichao feels that such a big tone has yet to be verified by the market.

Therefore, if you think about it this way, Nvidia's AI arms dealer plan really has to be based on the supply of these HBM manufacturers.

Although everyone is saying that the hard currency in the AI industry is H100 and B200.

Who would have thought that in the end, it would be stuck with South Korea?

View more at EASELINK