Rebirth 08: Rise from copycat phones

Chapter 432 The best suppression is dumping
Just as the wearable device department of Zhiyun Group, the investment department, supply chain department, and marketing department began to prepare for the large-scale production of AI glasses, striving to release this new product in the spring of next year.

Other business departments of Zhiyun Group have not stopped their development!
By the end of October, Weizhi Technology, a subsidiary of Zhiyun Group, was responsible for the development of X86 instruction set chips.

A new product launch conference was held earlier than in previous years, and a new generation of CPU mid-range desktop chips based on the 12-nanometer process were released.

This mid-range desktop CPU chip, together with several high-performance mobile chips and high-end desktop chips released in January this year, constitute the X86 instruction set of Zhiyun Group, a complete set of new-generation chips based on the -nanometer process.

This series of 12-nanometer CPU chips has surpassed competitors Intel and AMD's chips in transistor density, while also having better power consumption performance.

From a hardware perspective, Zhiyun's series of CPU chips have already surpassed its competitors...

Unfortunately, there are still some gaps in ecology.

There is no way around this. Intel has accumulated too many ecological advantages in the field of X86 instruction sets. It has many more advanced instruction set technologies and unique technologies, which Zhiyun Group cannot use. Zhiyun Group can only develop similar technologies on its own to supplement them.

This ecological gap cannot be made up in a day or two.

However, the CPU business itself is a supplementary business in Zhiyun Group. Whether it is the purely self-developed SOP instruction set or the X86 instruction set acquired through acquisitions, the purpose of its chip products is very simple, which is to avoid being strangled by competitors.

Therefore, slightly worse performance is acceptable.

What is really important is that not long after the release of the new CPU product, Zhiyun Semiconductor, a subsidiary of Zhiyun Group, released a new generation of server GPU.

Strictly speaking, it is not a brand new product, but an upgraded version of the APO4500 graphics card released at the beginning of the year: APO4600.

This graphics card uses more advanced technology, supports larger video memory and memory bandwidth, and improves the performance support for some algorithms to a certain extent. The overall performance is improved by about 4500% compared to the APO graphics card which also uses a nm process.

The APO4600 graphics card is also Zhiyun Group's last major upgrade and update in the / nanometer process era. There are no subsequent server GPU design plans based on this process node.

The performance of this graphics card is still quite powerful. Even when compared with the AI5500 graphics card used by Zhiyun Group, the only difference is in the video memory capacity, video memory bandwidth and communication speed between GPUs... Of course, this difference alone will lead to a big gap in performance during artificial intelligence training.

Why is the NVIDIA X2 graphics card under Zhiyun Group not suitable for AI training? It has a large number of core processing units and strong theoretical computing power, but the maximum memory capacity is only 16GB, the memory bandwidth is only 32GB, and the interconnection technology between GPUs only supports dual-card interconnection, with a bandwidth of only 32GB... This leads to low memory bandwidth and capacity, which means low training efficiency, and the interconnection technology only supports dual-card interconnection, with a bandwidth of only 32GB, which means that users cannot use this thing to form a GPU server array with thousands of graphics cards.

Zhiyun Group directly eliminated the possibility of using the X series as an artificial intelligence computing power card... Otherwise, if everyone uses X series graphics cards worth 10,000 or 20,000 yuan to train AI, then how can APO graphics cards worth hundreds of thousands of yuan be sold?
Similarly, there is the same gap between the APO series graphics cards and the AI ​​series.

The latest version of the APO4600 graphics card and the AI5500 graphics card used within Zhiyun Group are both based on -nanometer process chips. The number of transistors is actually similar, with more than transistors per transistor, and the number of processing units is also similar.
(AI4500 appeared several times in the previous article, which was a misunderstanding. These two are actually the same thing. Both are chips that use a 5500-nanometer process, and will be unified as AI later).

However, the difference between the two is huge, mainly reflected in the video memory capacity, video memory bandwidth, interconnection bandwidth between GPUs and special ecological support.

The APO4600 graphics card has a memory capacity increased from 24GB to 32GB, a memory bandwidth increased to 800GB, and an interconnection bandwidth increased to 500GB.

The AI5500 graphics card for personal use has a memory capacity of 80GB, a memory bandwidth of 1.6TB, and an interconnection bandwidth of 1TB.

These three key data will eventually be reflected in artificial intelligence training, and there will be a relatively large difference in training efficiency. Moreover, this difference will become increasingly apparent when supporting large-scale artificial intelligence training.

Large-scale artificial intelligence training, such as the ultra-large-scale generative AI training of Yun AI, cannot be handled by one or two servers. It requires a GPU server array consisting of tens of thousands or even more graphics cards for training.

The more GPUs are combined together, the higher the requirements for memory bandwidth and interconnection bandwidth... Otherwise, the computing power is sufficient, but the data is blocked.

The above are huge differences in hardware. In addition to hardware issues, there are also huge differences in software ecology.

In order to adapt to Zhiyun Group's own underlying artificial intelligence algorithms, the AI ​​series graphics cards have many specialized modules since the earliest AI series. Not only are the chips specially developed, but even the R&D support software is specially adapted and developed.

The AI ​​chip is designed for Zhiyun Group's own underlying artificial intelligence algorithms...it does not support a large number of various messy artificial intelligence algorithms out there.

If other companies take it, not only will they not be able to use it, but even if they find a way to use it, the performance will not be much better than the APO graphics card.

The relationship between the APO graphics card and AI graphics card in Zhiyun Group is somewhat similar to the smart terminal SOC under Zhiyun Group. The S series is a specially optimized and designed self-use chip. Many of the core modules in it can only be used by Zhiyun mobile phones and other companies cannot use them at all.

Instead of using the S series chips, other mobile phone companies might as well use the W series chips directly.

After all, the W series chips are general-purpose SOCs designed and developed specifically for the open source version of the Android system. They have a mature development ecosystem and can be used by any mobile phone manufacturer. They are also quite easy to use.

Zhiyun Group likes to do this kind of thing. It uses one for its own use and another one for external use.

On the one hand, it is to maintain the advantages of its own products. On the other hand, it is also because the ecosystem is gradually separated. Many things developed within Zhiyun Group involve many very unique and exclusive technologies and requirements, which can only be specially designed and cannot be turned into general products.

Over time, the current model has been formed.

However, although the APO graphics card is just a reduced-power version of the AI ​​graphics card, it is undeniable that the APO graphics card in Zhiyun Group is still the most advanced server computing power card that can be publicly purchased on the planet.

In terms of hardware performance and ecological support, they are far superior to AMD's similar competing graphics cards...the kind that have no competitors.

After the APO4600 graphics card was officially released, it naturally did not cause any response among ordinary consumers. This thing was too far away from ordinary people like them.

And with a price of over 150,000 yuan per copy, no one would be willing to use this thing to play games!

But for major high-tech companies, it is different. The performance of this graphics card is much better than APO4500. Although the price is a bit more expensive, it is still slightly cheaper if the comprehensive example cost is calculated.

Therefore, many large high-tech companies continue to place orders with Zhiyun Semiconductor, with orders of tens of thousands or hundreds of thousands of units.

American companies such as Google, Microsoft, Apple, and Tesla, as well as many domestic high-tech companies, such as those large Internet companies, are placing orders one after another.

Even so, Zhiyun Group still felt that the computing power card market was not lively enough... and released a new generation of general terminal computing power platform PX3 a few days later.

The PX series universal terminal computing power platform can be applied to the fields of robotics, intelligent driving, drones, or other medium and large devices that require local computing power.

The PX3 chip, which uses a 60-nanometer process, has a computing power of 480TOPS for a single chip. The most important feature is the high interconnection bandwidth. It supports the interconnection of up to eight chips on a single motherboard, providing a maximum comprehensive computing power of TOPS.

Once released, this computing power platform also attracted the attention of many companies in the industry.

A large number of new car manufacturers at home and abroad have entered the field of smart electric vehicles, and some traditional car companies that are ambitious to develop their own assisted driving platforms have come to consult.

Because the computing power provided by this chip can actually meet the needs of L3 level assisted driving... As for whether it can be done, it depends on the technical level of each manufacturer in artificial intelligence assisted driving... competing in algorithms and big data training.

PX3 is a terminal computing chip that can cover small to large devices, together with the lightweight, low-power general-purpose terminal computing chip LC1 released at the beginning of the year.

The semiconductor department of Zhiyun Group has completed the layout in three major areas: server computing chips, medium and large terminal computing chips, and small computing chips.

Whether the customer wants to train artificial intelligence or run large-scale artificial intelligence.

Or should we start from small scale or even run local AI in large devices?

Or running local artificial intelligence on tiny devices.

Zhiyun Semiconductor, a subsidiary of Zhiyun Group, can provide corresponding computing chip products with excellent performance... Of course, the price is as expensive as ever!

At the same time as the release of these computing chips, Zhiyun AI Research Institute under Zhiyun Group released a new generation of open source AI GTAI2, including regular upgrades to several open source AIs released in the first half of the year. Although people found after testing that this GTAI2 open source model has made significant improvements compared to the previous generation, it is still much worse than Yun AI.

But it is open source... We won't charge you a penny. If you have the ability, you can deploy GTA2 on your own or even use it in commercial areas. Zhiyun Group will not care about you.

Zhiyun Group released the GTAI2 large-scale generative AI model with a simple purpose, which is to further promote the popularity of the artificial intelligence market... The previous GTAI1 was released more than half a year ago, but major manufacturers were too slow to move forward. Apart from using it as is, no manufacturer has upgraded it based on GTAI1.

When Zhiyun Group released GTAI, it was just an attempt to stimulate further development...but it didn't expect that after the brick was thrown, the jade was still nowhere to be found...The management of Zhiyun Group overestimated the R&D capabilities of other companies in the field of artificial intelligence.

Even if there is an open source model in front of them, with various public paper materials attached, allowing them to read and study it at will... but only a few can truly understand it and continue to conduct further research and development based on this foundation, and it would be even better if they can produce results.

So Zhiyun Group is anxious...

Your technology is advancing so slowly, how can I sell chips?

So, they picked out some of the old technologies that were no longer used by the company, integrated them into GTAI, and released the so-called GTAi2.

If you don't develop it, I will help you... Now the more advanced Kaiyuan model GTAI2 is already here. If you don't know how to tinker with it, then don't bother. You can just use my Kaiyuan model. You can even give it a name and say it's your own development.

Then quickly bring it out for commercial use. Artificial intelligence is such a great thing, you must make it available to your own massive customers... Even if the computing cost is high, you have to grit your teeth and persevere!

The open source model of artificial intelligence in the field of robotics is also based on a similar idea... open source a robot model, and then many companies will see the opportunity and devote themselves to various artificial intelligence product applications, so that they can sell various computing chips.

In order to sell chips, Zhiyun Group has been working hard... The only thing left is to stuff artificial intelligence directly into the mouths of high-tech companies in various countries!
If you don’t play with artificial intelligence, how can I sell chips?
If I don’t sell chips, how can I recover my huge investment in the semiconductor field?

You should know that Zhiyun Microelectronics has just put into operation a CoWoS packaging factory, increasing the CoWoS packaging capacity to 50,000 pieces per month. At the same time, it is also building two new CoWoS packaging factories. After the entire plan is completed, Zhiyun Microelectronics' CoWoS packaging capacity is expected to reach 100,000 pieces per month, which can not only meet its own computing chip needs, but also supply large quantities to the outside world!

The investment in a CoWoS packaging factory is also huge.

CoWoS packaging, which can also be translated as 2.5D packaging or 3D packaging, is simply an advanced packaging technology that can stack multiple chips of different types together.

The previous chip development was from single-core to multi-core...but a multi-core chip is still a chip, it’s just that multiple core areas were planned during the chip design.

The 3D packaging technology is to simply package multiple chips of different types together... For example, the GPU, CPU and even memory chips are packaged together... For example, the APO or AI series GPU core produced by Zhiyun Group looks like just a GPU chip, but in fact, in addition to the GPU core, there are also a bunch of memory chips.

3D packaging technology is also a necessary packaging technology for producing high-performance computing chips... because this packaging technology can effectively improve the connection speed between chips, thereby improving the overall performance of the computing card.

Zhiyun Group started using this packaging technology in the field of computing power cards very early on, and has very strong technical accumulation in this field... because Zhiyun Group itself was the first manufacturer to play with computing power cards and the first to apply 3D packaging technology on a large scale.

Of course, Zhiyun Microelectronics is not the only company that has this 3D packaging technology. Some other semiconductor manufacturers also have similar technologies. The technical implementation paths may be different, and the performance effects may be different, but the basic principles and purposes are the same.

This is one of the reasons why AMD can produce computing power cards with just okay hardware performance. Its foundry manufacturer TSMC also has 3D packaging technology...it’s just that AMD’s hardware design capabilities are still a bit lacking.

More importantly, AMD’s computing power card ecosystem is not good, so after the computing power cards are developed, they cannot be sold, which in turn leads to huge losses.

Since AMD entered the computing chip field a few years ago, its cumulative losses in this business have exceeded US$1.5 billion... If it continues like this, the losses will be even greater. The management is terrified and the board of directors is heartbroken.

So starting this year, AMD has actually significantly slowed down its R&D investment in the computing chip field... If it continues like this, the little money it earns from selling CPUs will be all gone.

In the field of computing power cards, Zhiyun Group has such a huge advantage that AMD simply cannot compete... After several years of working on computing power cards, they have gained nothing except a loss of more than a billion US dollars!

So AMD backed off!

Qualcomm and Intel, on the other hand, have not given up and are still trying to develop computing power cards, but it is more difficult for them to develop computing power cards... They are not even as good as AMD.

AMD's failure in the computing power card field is mainly due to ecological problems. Speaking purely of hardware, their latest flagship computing power card is actually quite good in terms of hardware performance. Even though it is not as good as the APO4500 graphics card under Zhiyun Group, it barely reaches a level of about %. If you really want to use it, it can actually be used.

The problem is that the cost-effectiveness of the overall computing power is too low, and the ecosystem does not support it...so few people use it.

Qualcomm and Intel, in addition to facing the same cost-effectiveness and ecological issues in computing power, are also facing huge troubles in the field of hardware design... Without some experience and technology, it is impossible to produce high-performance computing chips.

After all, AMD has the foundation of independent graphics cards, so it can be considered a professional match.

However, in the field of GPU, Qualcomm and Intel have only experience in CPU integrated graphics and SOC integrated graphics respectively. As for high-performance independent GPUs, to be honest, they have never done it!

This is a completely unfamiliar area for them!
Not to mention, server GPU is a step further than regular consumer-grade GPU!

There are still several levels in between.

----

The three major chip giants in the United States have all made no progress in the field of computing power cards and are facing various huge problems.

On the contrary, Tesla's self-developed computing chips are making good progress... However, Tesla's self-developed chips are all for its own use, self-production and self-sales, and even the software development ecosystem is designed by itself. In fact, it is following a similar path as Zhiyun Group in its early days... The ecological issue is not a big problem for them, it is mainly a hardware design issue, but Tesla is developing terminal computing chips, which is very different from server GPU chips.

In addition, for ADM, Intel, and Qualcomm, they develop computing chips in order to sell them...and if they sell them to the outside world, they need to get customers to accept their new ecosystem, which is extremely difficult.

AMD has been promoting its own ecosystem for several years but has not been successful. The reason for their failure is not only the ecological competition with Zhiyun Group, but also partly due to Intel and Qualcomm, because these two companies are also promoting their own new computing chip ecosystems.

In the field of computing chips, even Xu Shenxue did not expect that in the face of their own monopoly position, the three major chip manufacturers in the United States would fight among themselves... They were all competing for the second ecological niche, but in the end no one could get it.

At the same time, the three major chip manufacturers in the United States, as well as Tesla, Apple and other companies that want to develop their own computing chips are all making considerable moves, which has also made Zhiyun Group feel some competitive pressure.

That’s why they have been so active in the past year, not only engaging in open source artificial intelligence, but also opening up to sell various computing chips… This can actually be considered a form of product dumping.

Zhiyun Group is saying to Google, Microsoft and other manufacturers that need a large number of server GPUs: Don’t wait for AMD, Qualcomm or Intel’s computing power cards. Why bother? I will just sell you the APO4500/4600 graphics cards at a discount and provide a full set of free development software support.

Not only do I sell chips, I also open-source the GTAI2 large-scale generative AI...

After you deploy it locally, you can change the name and the skin and commercialize it directly... You can even call it your own, I have no objection.

When Zhiyun Group was selling computing power cards, it provided all kinds of thoughtful services, making all major artificial intelligence companies feel comfortable... Then AMD's computing power cards couldn't be sold, and it suffered huge losses for several years. Now it is about to cut off this business.

At the same time, the technological gap between the two is getting bigger and bigger. In this case, the more time goes by, the more perfect Zhiyun's computing power card ecosystem will be established, and the more serious the dependence of artificial intelligence companies in various countries in the world on Zhiyun's computing power card ecosystem will be.

When that time comes, AMD will have even less chance.

As for Qualcomm and Intel... because of their professional mismatch, although they are still insisting on continuing to develop computing power cards, they are far inferior to AMD's level, not to mention seeing the taillights of Zhiyun Semiconductor, the gap is as huge as a huge distance... So these two companies are not a big threat for the time being.

However, even if there is no obvious threat, Xu Shenxue intends to maintain his current strong attitude and continue to suppress several American chip design companies.

If you want to suppress it, the best way is not to stop selling high-performance server GPUs, but to sell them on a large scale!
Nothing works better than dumping!

When we completely squeeze out their server GPU field and make them lose the courage to even take a look at server GPUs, it will be time to reap the rewards!

(End of this chapter)