Five predictions from a senior chip architect about the industry

Source: Content reprinted by Semiconductor Industry Observer (ID: icbank) From Zhihu , author: Xia Jingjing, thank you.

(๑＞ ＜)☆

Each company has a different perspective, and each has its own wonderful angle. What makes me uncomfortable is that many of these predictions are conclusive judgments, without any logical reasoning process, and they don’t even care what was predicted the previous year (did last year’s prediction come true). But what makes me even more uncomfortable is that all the trends are optimistic! No company thinks there will be any crisis in the future (such as a volcanic eruption?) What problems need to be faced?

…Trend is a neutral word isn’t it?

Well, I plan to publish an irresponsible 2022 forecast for the next five years in the chip field! Haha, I will come back every year to see if I can add to the previous one.

First, let me quote a passage from the WeChat public account Zhengshitang: The difference between Sima Qian's "Records of the Grand Historian" and Sima Guang's "Zizhi Tongjian" stems from the difference between a writer and a politician. Sima Qian, who lacked political experience, made up a story, but it was still a story. In the cognition of the scholar Sima Qian, the struggle for the world was the people's hearts, while the politician Sima Guang knew very well that from the division of the Jin Dynasty by the Three Families to the replacement of the Qi Dynasty by the Tian Family, from Wang Mang to Cao Cao, from Sima Yi to Yang Jian, and then to a series of palace coups in the Tang Dynasty, the struggle for power relied on money and the distribution of interests.

What does this mean? In fact, the main thread that dominates the evolution of world civilization and the change of dynasties is productivity and its economic laws. The so-called various stories are several branches of the whole tree.

So what is the productivity of the chip industry and its economic laws? The answer is well known to everyone: Moore's Law. In fact, it was proposed as an economic law, and then the logic of the development of productivity and production relations was written. I wrote a post " Refreshing the cognition of Moore's Law " .

https://zhuanlan.zhihu.com/p/104453844 .

When this logic is not stable, or some variables are no longer stable, the story unfolds.

If we look back, when Moore's Law first became ill, Intel was affected and weakened, Apple and ARM rose, and the logic that real men have their own factories lost its persuasiveness. What is the essential change? It is that the allocation of scientific and technological resources is changing from self-owned to crowdfunding. The only difference from normal crowdfunding is that there is a big rogue Apple. The crowdfunding method it advocates is actually "I will be the first to eat the soup, and after playing almost enough, the younger brothers will happily take over." Think about it, it is not just the process to TSMC, the CPU to ARM (terminal crowdfunding), and even the APP to Android. Unless there are some things that cannot be helped, Apple is like this. My view is that this is an important factor in Apple's rise. But this story seems to be a bit difficult to play in the next few years.

In fact, Moore's Law has not stagnated from a technical point of view. 2nm and 1nm are both feasible. However, the price of each transistor has changed since 28nm, and the economic curve has changed!

Prediction 1: TSMC will be the only one to rise and force Intel to switch to fabless. Taiwan will further become the global wafer manufacturing center.

Last month, Intel CEO paid a special visit to TSMC and posted a video on YouTube. Two days ago, TW also broke the news: digitimes.com.tw/tech/d

This has at least four meanings:

1. The cost of process evolution, which was originally unaffordable for TSMC (at the level of US$30 billion per year, how much should the wafers be sold for to recoup the investment?), has now changed from being a single-party effort of APPLE to crowdfunding by Apple + Intel, which is a major boost for TSMC to become even stronger.

2. The fact that Intel and Apple each built their own factories shows that their process customization demands are very different, and the combined production capacity requirements are greater than previous processes.

3. Intel's dedicated fab capacity is 2W wafer per month (capacity reservation). This capacity cannot be consumed by simple GPUs. It requires a large number of CPUs to enter the market. This will reduce Intel's motivation and cost to build factories in the United States, and will make Intel Fabless. Similarly, Samsung can no longer find people to crowdfund, and will also slow down at a visible speed.

4. Advanced process production capacity is further concentrated in TW. Haha, I don’t know what Americans are thinking. It may be a democratic decision.

Prediction 2: High-profit oligopolies such as Apple and Intel will decline due to rising transistor costs, and many secondary manufacturers with better energy consumption allocation, such as M/O/V and AMD, will rise.

Not to mention Intel, Huawei's restrictions in the past two years have given Apple room to expand its market again, but even so, the cost burden brought by the new process will gradually affect its product profits in the following years. Although Apple's profits are high, you know, the stock market is always an advance payment for the future. Unless it can find a new track, Apple will not be able to continue to lead in new technology (of course, Intel's crowdfunding is also good for Apple). In fact, if you pay attention to the recent news, some signs are emerging. Apple has recently increased its salaries in large numbers. Capitalists are not good people.

And this Mike, I had roast chicken with him in Texas, he drove a red two-seater small sports car which looked like it would burst... He was a fellow of ARM during the ARM A72 era, no small figure in the field of technology, a duck knows first when the river water warms in spring.

The profits of Intel and Apple are so high, just like a giant whale greedily absorbing huge amounts of food to maintain its size and continued growth. This energy loss can fully support countless more energy-efficient sharks to eat their fill. As long as there is meat to eat, the sharks' tough stomachs have a great tolerance for cost increases (there is also a huge price gap between Android and Apple phones).

Expectation: when a whale dies, all things come to life.

Prediction 3: ARM will also gradually decline. In extreme cases, Apple will abandon ARM and independently evolve ISA

Many people in China like to mythify ARM, as if they can kill you with just one move. If you look carefully, ARM's lack of growth and losses have been going on for more than a year or two.

Let me ask you: Where do you think the competitiveness of ARMv9 lies? Why do you buy ARMv9 CPU products? Where will the competitiveness of ARMv10 be in the future? ARM's ISA has always been driven by Apple, and it has been looking for innovative paths (collecting requirements) in the industry's general crowdfunding. However, the space and difficulty of this innovation are increasing year by year, and it is almost so big that Apple might as well give up ARM and customize it based on its own closed system. For Apple, there is no fruit to be picked in the industry. For ARM... Being acquired by NV may be the best destination, but there is also the risk of being separated from Apple... ARM is not so glorious. Compared with TSMC, it does not have the help of Intel.

Prediction 4: NVIDIA is a graphics card company, but the support point of its stock price is AI. If the AI ​​algorithm undergoes some kind of innovation as I expect, Nvidia will be in crisis.

NVIDIA's 800 billion valuation is basically more than half of the estimate for AI. It seems that the people at NV are a little crazy about AI. Brother, you were originally a graphics card.

Of course, there will be a wave of AI algorithms gradually being applied to various industries in recent years, but with the development of AI big models in the next two years (perhaps it will take another two years), everyone will understand that computing power is not everything in AI. To achieve true intelligence, matrix calculation is not the key. Related article: "Talking about Heterogeneity from GPU (13)" https://zhuanlan.zhihu.com/p/457978103 )

I will talk about my understanding of large models or intelligent agents in the next post after the above one.

I think that the current model size is already far larger than the number of neurons in the human brain, but the idiotic situation has not been alleviated at all. Neither logical induction nor structured knowledge is reflected. It is just a feeling that the dictionary has become larger. If we look back at the human brain, we will find that there are two differences at this time. One is that the human brain is spatial computing, and the other is that the number of external connections of neurons is 1K. Future AI training may change from training weights to training both weights and connections, and thus generate truly deep knowledge. If we really go in this direction, GPGPU will not be able to handle it.

Well, I wonder if I, who comes from the Internet field, will rise again. I am secretly happy.

Prediction 5: The distributed working mode caused by the epidemic will replace the original centralized working mode of programmers, and thus trigger a new development mode of software engineering. This new mode will make the interface of each programmer more simplified, and there will be a revival of functional programming. This will also trigger new innovations in the hardware architecture from CPU to data center.

I haven’t thought clearly about my weaknesses. I’ll talk about them next year.

～(￣▽￣～)~～(￣▽￣～)~～(￣▽￣～)~

What stocks should I buy?