To ensure that a representative growth stock like Bayer performs well, I believe that we need to implement measures to lower interest rates. Shall I ask that question? In fact, back in the days of the Gold Rush, the companies that made significant profits were those selling picks, jeans, and tents. The AI era has shown us something similar, with Nvidia, for instance, gathering a lot of money. Along with that, companies involved in AI semiconductors, like HBM, have created a significant trickle-down effect. However, if we look at this from a stock price perspective, there have been various debates recently about the winter of the AI semiconductor sector, and there are concerns about a potential bubble. It could also be a seasonal issue. What do you think about this, sir? It’s quite complex. I believe it was in April of last year that there was a SupEx committee within the SK Group.

The mental model for SK's growth is similar to that of a Suss. The Specs Committee includes all representatives of the SK Group, including vice presidents. There, I had the opportunity to meet with 13 representatives from the SK Group, give a lecture, and have a Q&A session and discussion like today. They asked the same question. One person, whose identity I won't disclose, asked me how I feel about semiconductors. I mentioned that Saltlux is a customer, right? They are also a partner. But I responded, I don't know. I could sense their disappointment, as they seemed frustrated that I didn't have an answer, especially since AI needs semiconductors. So, I said, I think the HBM market is going to grow significantly. I also mentioned that this would likely lead to an increase in Hynix's stock price, which is substantial. So I said...

I think it's correct, but what I'm afraid of is whether we can really see growth in the existing market with only the memory that goes into GPUs. The question is whether the demand for DRAM can continuously rise as much. To be honest, I'm not sure about that right now. The reason is that while the demand for GPUs is increasing, and there is certainly a demand for special memories like HBM that goes into GPUs, I'm not sure if the generalized computing demand for standardized memory will continue to increase alongside it. For example, if we were to eliminate search engines like Google and Naver or websites like Amazon, it’s possible that the overall demand for computing-related semiconductors, including memory and other types of semiconductors, could actually decrease. Ironically, that could happen. So, that’s what I’m concerned about.

It's very difficult to make a judgment easily, and I'm not an expert on AI demand either. However, I would like to discuss a few things today regarding semiconductors. So, how is the semiconductor market doing? When this graph comes up, everyone tends to get confused. Please don't take it too seriously; I'll introduce it in a more relaxed way. 

The graph is indeed complicated. The dots represent the size of the brain of artificial intelligence, or the neural networks. And if you look here, GPT-3 is located around here. Each point corresponds to different AI models, up until 2021. I haven't been able to find the graph beyond that year. On the vertical axis, it says training compute and petaflops.

You don't have to understand this, but here you can see that one space increases by ten times. Each of those large spaces increases by ten times. This is called a logarithmic scale. To put it simply, if I explain it like this, it's about how quickly GPUs and CPUs can perform calculations. The vertical axis represents petaflops, which is a thousand times more than teraflops, and teraflops is a thousand times more than gigaflops. So, the units we usually use aren't really applicable here; in other words, it's incredibly fast. The problem is that this looks linear, but the vertical axis increases by ten times. So, in reality, the speed of semiconductors is increasing tenfold every one or two years. The speed required for semiconductors is increasing exponentially every one or two years. But what's amazing is...

What this means is that what we have here is our semiconductor's Moore's Law. Moore's Law states that every two years, the integration density of semiconductors doubles, they become faster, and their prices decrease. Next, what we have above is the prediction of the development speed of semiconductors needed for traditional artificial intelligence, which is more important. The red line represents the prediction of the development speed for super-advanced artificial intelligence. This is where we have all the generative AIs you know, like ChatGPT, Bard, and many others. However, ChatGPT 3, which has already been out for over two years, is positioned above that line. In terms of speed, it is much faster than the semiconductor development speed. The speed of traditional artificial intelligence development is not just double; it is actually 100 times faster. From the perspective of generative artificial intelligence, the speed at which semiconductors are catching up is completely...

It seems like a vague idea, but it’s coming at us at 100. Next, looking at the graph here, this shows the recent speed of development in super-advanced artificial intelligence, which is twice the speed of existing artificial intelligence development. This means that the speed of AI development is 10,000 times faster than the existing semiconductor speed. However, this is just looking at the overall speed. If we convert it based on a two-year period, super-advanced AI has increased by 750 times in two years. This is not due to revolutionary technology in semiconductors but rather a costly trick. Now, semiconductor companies are involved, but Samsung is struggling to produce this, making it a very challenging trick. Strictly speaking, this is about stacking existing HBM RAM. The reason it is technically difficult is that when stacking the fine processes, holes need to be drilled from top to bottom to manage the data.

We need to be able to give and take, so we have to be angry when we give and take. This is a technology that hasn't been achieved yet. Of course, the best places that do this realistically are Hynix, Samsung Electronics, and Microchip, but even those are struggling to keep up, and the gap is widening. This means that it won't last for five or ten years. So my question is this: Can semiconductors continue to develop as rapidly as artificial intelligence? There is a graph that shows this is impossible. It's very technical, but it involves FETs, which you may know about. It's difficult, but you can think of it simply as a transistor. What is a transistor? It’s when you mix impurities into a silicon crystal and apply a very small voltage, allowing you to control a large current. It's like a switch.

But the switch is made incredibly small, just a few nanometers, and this is what connects an enormous number of transistors, which is what a semiconductor chip is. Inside the chip, there are hundreds of millions of transistors, right? There are tens of billions of them. So, in order to achieve better performance, we need to keep making them smaller, right? This way, they consume less electricity and generate less heat. Therefore, everyone places great importance on the manufacturing process. Is it a 5-nanometer process or a 3-nanometer process? This manufacturing process refers to the size of the channel, which we call the channel size in transistors. It's the size of a single transistor. This is actually the core of Moore's Law, and while we are trying to keep up with it, RAM has been progressing twice as fast as Moore's Law up until now. But the problem is whether we can keep making them smaller. There are limitations, and the reason for those limitations is...

If it gets any smaller, it becomes impossible for humans to control. In other words, a quantum effect occurs. This quantum effect makes it so that we don't know where things are going, and management becomes unmanageable. The difficult part of quantum mechanics is the principle of uncertainty. In the very small micro world, we can only understand things in terms of probabilities; we cannot control atoms, especially electrons. When that state occurs, they operate randomly. So when will that happen? Looking at the graph, if it happens quickly, it will be around 2030, and if it happens later, it will be around 2034 or 2035. It's approaching; we have about 5 or 6 years left. If we look at it in the long term, we have about 7 to 8 years remaining. By that time, we won't be able to advance semiconductor processes any further. We might need to bring in aliens, right? Otherwise, we need to find a completely different method.

The current method of integrating semiconductors by placing them on silicon crystals is progressing at a pace that is only a fraction of the speed at which AI is advancing. Even NVIDIA's GPUs and the NPUs, which refer to Neural Processing Units, are changing the entire architecture. The issue is that semiconductors focused on integration, like DRAM, will continue to operate in the same way for the next ten years. Therefore, it is necessary to find ways to overcome this limitation. Initially, my response to your question may have seemed complicated, but from my perspective, AI semiconductors can be broadly categorized into two types. First, there are learning semiconductors, which are produced exceptionally well by companies like DiA globally. Second, there are inference semiconductors, such as LPUs and NPUs, which are currently being developed by Samsung Electronics.

Alternatively, a well-known AI semiconductor company in the country is focused on creating products that are entirely based on inference. Once the learning phase is completed, they no longer study but instead aim to create semiconductors that facilitate thinking. This includes two main types: inference semiconductors and memory semiconductors. So, there are three major tasks at hand. Among these, if we consider which area is likely to grow the most, we need to address the memory semiconductor, specifically HBM memory semiconductors. 

Why is it called HBM? HBM stands for High Bandwidth Memory. The reason high bandwidth is essential is that data needs to be transferred in large quantities at once, which requires a wide bandwidth. It's about how much data can be read from and written to memory in a unit of time, similar to how a highway can handle a dump truck. In terms of memory, we have HBM and HBM2, and below that, there is PCI-based GDDR6X.

We are currently discussing the advancement of bandwidth, but GDDR alone is insufficient. If you look at the vertical axis, each increment represents a factor of ten. So, in fact, we are increasing by a factor of ten. HBM3 would be around this point. The GDDR we typically use, like the DDR RAM for PCs or servers, is stacked in layers—like twelve layers—which makes it difficult to achieve the same performance. Yet, it only reaches this level of speed, which is three to four times faster. This is only a small representation on the chart because the vertical axis is logarithmic; one increment equals ten times. Therefore, moving from HBM2 to HBM3 represents a tenfold increase, which is roughly one increment. It's important to note that this is not solely about advancing memory semiconductor technology or development techniques; the emergence of HBM2, HBM3, and HBM4 is quite significant.

We have high expectations for innovation, and while we cheer and invest, the gap is widening compared to the growth rate of artificial intelligence technology itself. Last week and the week before, Jensen Huang mentioned something important. Currently, we have TPU from Google here, and the H100 here. The advancements are happening at a rapid pace. The AI chips like GPU and MPU are improving, but memory technology is lagging behind. The memory is not progressing as fast, and this gap is expected to widen even further. Jensen Huang said a couple of days ago that the issue lies with memory semiconductors. We can produce faster, but memory cannot keep up. Nevertheless, the only option available, despite being costly and inefficient, is HBM at the moment.

It's not that Samsung, Hynix, or Micron can't do it; rather, it's that the current technology we have is the best we can achieve. There is a significant gap between what we have now and the potential. So, when should that breakthrough happen? As I mentioned earlier, the current HBM (High Bandwidth Memory) is made by stacking layers using traditional semiconductor processes. With that, we can increase performance by ten times or even twenty times, but look at the difference we're talking about here. It's not just ten times; it's a difference of a hundred thousand times or even a million times. Right now, the speed is such that we need another method to emerge within the next six to seven years, aside from HBM. From the GPU's perspective, it's like walking on an unpaved road in a tractor while I'm enjoying ice cream at a highway rest stop after switching to a Ferrari on the highway. Meanwhile, that friend is still struggling along on that unpaved road with a tractor.

Yes, it was ten times, a hundred times, a thousand times. So now, if we see a hundred, there will be a ten thousand times difference. Therefore, this is what Jensen is saying: HBM is very slow, but there is nothing else to use. So right now, the choice is HBM. This is what Jensen Wang is saying. But then, what needs to happen? I believe someone will emerge. That someone could be Samsung or Hynix. As you saw earlier, by 2035, there needs to be a company that can achieve a breakthrough within the next ten years. So who could that be? Looking globally, the most likely candidates are Samsung and Hynix. And there’s a second possibility; the truly scary ones are from China. China is claiming they will work on HBM, even HBM 2, but as I mentioned, this is not about now, it's about ten years from now. So if I connect what I just mentioned with the present, it looks like this.

From the perspective of AI, we are now at the peak of the hype cycle, and the bubble has started to deflate. So, the first question people ask me is, Is the bubble deflating? What does it feel like on the ground? Yes, the bubble is indeed deflating. But is that a problem? No, it's not a problem. Those who can survive the deflation of the bubble will be the ones who make money. 

So, how much longer do we have? It’s about three to five years left. After that, the AI era will fully commence. When this AI era begins, there will be software companies like AI oil companies, but as I mentioned earlier, during the gold rush, who were the ones selling jeans and picks? One of them would be semiconductor companies. So, what problems do semiconductor companies need to solve during this time? As I mentioned,

From a memory perspective, HBM cannot handle it. Currently, HBM1 is expensive, but the speed still cannot keep up. However, as humanity has envisioned, it will emerge. In five years, or at most seven years, there will be a completely new method. The stacking method has its limitations; it will still be stacked, but among them, some interesting ideas continue to emerge. Of course, one of the candidates that could create a significant breakthrough is quantum computing. Quantum computers operate in a completely different way from traditional semiconductors. However, I am personally quite skeptical about whether quantum computers will have the economic viability for us to use in a practical sense within the next five to ten years. Nevertheless, in the medium to long term, it will be completely...

There are other ways to approach this, which is why it is a significant solution, especially regarding quantum computers. However, if we consider the speed of commercialization and the process of market growth, there is still a need for memory to keep up with GPUs, MPUs, and LPUs. The possibility of needing alternatives to HBM 3, 4, and 5 is very high. This is because, due to quantum effects, we can no longer make components smaller, nor can we infinitely increase the performance. Therefore, raising the performance to the third or fourth level is already challenging. Achieving the tenth level would be miraculous, and reaching the twentieth or thirtieth level would be extremely difficult, as the performance would drop significantly. This is because we need to delve deeply into the process, which is quite complex. As a result, there is extensive research on technologies that can help with this. Now, going back, the Vice Chairman of SK Hynix mentioned this.

They said it's not true. Now, a new world will open up in semiconductors, so they asked me to make sure to talk a lot about how well Hynix is doing. The key is, if we're looking for a company that can solve that problem, the possibilities are high with Samsung and Hynix. Especially now, Hynix is overwhelmingly ahead in HBM technology, and according to a recent article, they will start mass production of HBM3 next year. They also mentioned that it's not easy to reach 20 or 30 layers. Therefore, a completely different approach is needed. The problem lies with the Chinese companies. Looking at what the Chinese are doing, they can't export high-priced equipment, so they are using low-cost equipment to do all sorts of things. Our guest, Jeon Byeong-seo, the director of the China Economic and Financial Research Institute, describes it as the kitchen knife technique. In other words, even without a sashimi knife, they can slice fish with a kitchen knife.

It feels like we're doing something similar, right? It's the same expression. Our country was like that. When Hyundai and Samsung learned from Toshiba and Toyota, they saw that Korean people were excellent. We really came with axes and said we wanted to try it out, and they were doing their thing with knives. We were there with axes, saying we would cut hair with them. Just imagine how scared they must have been. So, there is a process in China. Currently, both domestic semiconductor experts and AI experts like me are most concerned with the fact that, from both AI and semiconductor perspectives, China and the United States are significant players. Therefore, in every aspect of AI and semiconductors, it seems like we are trying to contain China. However, this is just buying time; I don't think we can completely stop them. So, how this will change is really interesting, especially regarding South Korea's Hynix.

From Samsung Electronics' perspective, it is a significant challenge, and we must create innovation. In fact, if we look back about 20 years, who could have imagined that we would catch up with the steel and chemical industries? Who would have thought we would catch up with LCDs? Now, OLEDs are taking over, and the only thing left is semiconductors. However, in terms of artificial intelligence and robotics technology, we are already lagging behind China. This may sound uncomfortable to some, as they might think, Why are we talking about China? or Are we being pro-China? But this is a realistic concern. Yes, it is worrying. For the representative growth stock, pharmaceutical companies like Bayer need to see interest rate cuts to thrive, and I believe that. However, I also think that there is still potential for the pharmaceutical industry to take another leap.

"HBM 어휴 답답해" 젠슨황이 가장 걱정하는 반도체는 이것...'10년 뒤의 중국'이 가장 무섭습니다 / 이경일 대표 (2부)

"지식과 자산의 복리효과" 그 길을 복길이 여러분과 함께 걷습니다.
( 업로드 : 오전 11시45분, 오후 4시, 오후 7시 )
제휴 문의 : ystreet2021@gmail.com

인터넷 언론사 제호 : 와이스트릿 / 등록번호 : 서울 아53714 / 등록일자 : 2021.05.24
주소 : 서울시 영등포구 국제금융로8길 27-9 여의도 동북빌딩 910호
대표번호 : 02-6949-1701 / 팩스 : 02-6949-1702
이메일 : ystreet2021@gmail.com
발행인 : 이대호
Copyright © 와이스트릿  YSTREET ALL RIGHTS RESERVED. 와이스트릿의 모든 콘텐츠는 저작권법의 보호를 받으며, 무단 전재·복사·배포 등을 금지합니다.


"HBM 어휴 답답해" 젠슨황이 가장 걱정하는 반도체는 이것...'10년 뒤의 중국'이 가장 무섭습니다 / 이경일 대표 (2부)

Scripts