Rebirth 08: Rise from copycat phones

Chapter 442 The Terrifying AI6000
The inspection at the 903th Factory in Tongcheng lasted a whole day. Xu Shenxue not only saw the S6000 chip that had just completed technical verification and successfully taped out, but also saw the AI graphics card that was still in the tape-out stage.

The GPU core of this graphics card is the second chip in human history to use an equivalent seven-nanometer process!

Fu Zhengyang introduced the basic situation of this graphics card to Xu Shenxue. It adopts an equivalent seven-nanometer process and has a powerful GPU core. Compared with the previous generation AI5500, the biggest change of this GPU core is that the number of transistors has increased significantly with the improvement of the process.

The number of transistors in the GPU core of AI5500 is just over 6000 billion, but the number of transistors in the GPU core of AI has been directly increased to just over billion. Not only that, the power consumption has not increased significantly.

The substantial increase in the number of transistors means a substantial improvement in performance.

Xu Shenxue attaches great importance to the AI6000 graphics card project!
After all, this is related to the subsequent development of artificial intelligence within Zhiyun Group, and the development of artificial intelligence is directly related to the development of artificial intelligence application terminals of Zhiyun Group and its sister companies and some institutions... Typical examples are the intelligent robot business and the secret project AI fighter.

These two major businesses and some other terminal projects, such as AI glasses, are heavily dependent on the development of artificial intelligence by Zhiyun Group!
For artificial intelligence to develop, in addition to the advancement of underlying core algorithms based on mathematics, the improvement of computing power is also indispensable.

This shows the importance of AI series graphics cards!
Precisely because of the importance of the AI ​​series of graphics cards, this series of graphics cards is given very high priority in design and manufacturing.

Basically, the most advanced process capacity in Zhiyun Group is first to meet the demand for S series SOC chips, and secondly to meet the production needs of GPU cores of AI series graphics cards.

In recent years, the only exception is the 10-nanometer process, mainly because the 10-nanometer process is a transitional process in Zhiyun Group's plan, and mass production is expected to take only a year... At the same time, we had considered designing an AI series graphics card based on the equivalent 10-nanometer process, but the cost-effectiveness of the equivalent 10-nanometer process is extremely low.

When it comes to the AI ​​series graphics cards, the performance improvement is not very obvious. At the same time, due to the significant increase in costs, the final comprehensive computing cost is not as good as continuing to use the AI5500 with an equivalent -nanometer process.

Therefore, there is no AI series graphics card with equivalent ten-nanometer process.

Zhiyun Semiconductor, a subsidiary of Zhiyun Group, chose to directly skip the equivalent 6000-nanometer process and has long since determined the next generation of AI series graphics cards, which will be based on the equivalent -nanometer process. This is the AI graphics card project.

This also reflects the peculiarity of the equivalent 10-nanometer process: on the surface, the performance is better than the 12-nanometer process, but the cost performance is too low!
Therefore, at present, apart from the relatively special S series chips, the new generation W1060 chips of the W series that will be launched on the market next year, and other artificial intelligence terminal chips such as the EYQ series chips and ZY series chips that occupy little production capacity, Zhiyun Group has no other large-scale equivalent ten-nanometer process chip projects.

If nothing unexpected happens, starting from the second quarter of next year, as the production of S803 chips is gradually reduced and the S903 chips are turned to large-scale mass production, although the W1060 chip will meet part of the demand, the company's own orders for the -nanometer process will still be greatly reduced.

However, there is no need to worry about insufficient capacity utilization of the equivalent 10-nanometer process, because Zhiyun Group itself does not need it, but other manufacturers will still have demand for it!

After all, for Zhiyun Group, the equivalent ten-nanometer process is a rather awkward transitional process, which is neither high nor low, has high costs and extremely low cost-performance!
But for other manufacturers, this is not a question of cost-effectiveness, but a question of whether they can obtain advanced process production capacity.

Because by next year, if nothing unexpected happens, the mobile phone SOCs of Zhiyun Group and Fruit will enter the era of equivalent seven-nanometer process.

It will be very difficult for other manufacturers to follow the 7-nanometer era in the short term because the production capacity of the equivalent 7-nanometer process will be limited, and they will not be able to obtain production capacity from TSMC or Zhiyun Microelectronics in the short term.

Even if they can obtain a small amount of production capacity, the price of the equivalent seven-nanometer process chip is too expensive, which will cause them to give up!
The overall cost of an equivalent seven-nanometer process chip produced using quadruple exposure will be much higher than that of a ten-nanometer process chip.

Taking Zhiyun Group's S series chips as an example, when calculating the cost of the chip internally, such as the development cost, the cost of the S803 chip was priced at US$903 million, but the development cost of the S chip rose to US$ million.

In comparison, it was said that there are about 100 million and 300 million respectively!
The reason why there is such a price difference in the chip development costs of the two companies is that Zhiyun Group's S series chips include a communication baseband, and the development cost of the communication baseband is also very expensive.

Apple's A series chips do not have a communication baseband. Although the development cost of the A series chips seems to be lower, if the purchased communication baseband is added, then in total, the cost of their A series chips plus the communication baseband is actually more expensive than the S series chips.

The development cost of the S series 903 chip was doubled compared to the 803 chip... and that doubled to hundreds of millions of dollars.

With such a huge increase in development costs, if the chip has a lower production and sales volume, the unit price of the chip will be greatly increased by this huge development cost.

At the same time, advanced technology not only brings about a rapid increase in R&D costs, but also brings about a substantial increase in material costs.

Taking the internal cost of the S series chips as an example, the material cost of the S803, which is equivalent to a 903-nanometer process, is about US$, but the current material cost of the S chip is estimated to be around US$, a direct increase of %.

The combined price of Fruit's A series chips of the same period plus the communication baseband they purchased would be more expensive than Zhiyun's S series chips.

Only Zhiyun's S series phones and one Apple phone dare to use this kind of expensive top-level chip. They rely on huge sales to spread the development costs and continuously reduce material costs, ultimately bringing the overall cost down to a controllable level.

If other chip manufacturers and mobile phone manufacturers dare to use chips with equivalent seven-nanometer process next year, their gross profit margin will become very ugly, and they may even suffer losses if their sales volume is lower.

Of course, they can’t use it even if they want to, because there is not enough production capacity for them!
As the equivalent of 7-nanometer process is expensive and production capacity is insufficient, many chip and mobile phone manufacturers have chosen to use the equivalent of 10-nanometer process for next year's flagship chips.

Although chips with an equivalent 12-nanometer process have a very low cost-effectiveness, they are always cheaper than those with an equivalent -nanometer process, and are also better than -nanometer process... It's better to have one than not to have one!

However, this will only be limited to some top-level chips and flagship phones. For most low-end and mid-range, and even some high-end phones in the mid-to-high-end range, the 12-nanometer process will still be the main technology used next year.

Zhiyun Semiconductor is still vigorously promoting the two chips W907 and 908 based on the -nanometer process. Qualcomm next door is no exception. It is also promoting chips based on Qualcomm's second-generation -nanometer process, which it recently changed to -nanometer process, to the market.

By the way, Zhiyun Group was the first to be shameless and directly called the second-generation 14-nanometer process the 12-nanometer process.

At the beginning, Samsung and TSMC were rather reserved and reluctant to follow suit, but later they found that consumers really bought into it and felt that the 12nm process was more advanced than the 14nm and 16nm processes!

So, starting this year, they have gradually called the improved process based on their own 14nm/16nm process the 12nm process... In fact, it is just a change of name, and there is not much actual difference.

The chip process nodes set by today's major semiconductor manufacturers have long since become terms used in commercial promotions and have nothing to do with the actual process nodes.

This is also why the term "12-nanometer process" has begun to appear on a large scale in the market. The corresponding 14-nanometer process and 16-nanometer process are rarely mentioned...

According to industry forecasts, the life cycle of the 12/14 nanometer process node will be very long. The technology is mature, the foundry fee is low, the design cost is not high, and the production capacity is relatively sufficient... and as the cost of the advanced equivalent 12-nanometer process is too expensive, a large number of mature chips will remain at the 14/ nanometer process node.

Just like the current 28-nanometer process, don’t think that the current 28-nanometer process is already outdated and has no market... In fact, a large number of chips are still manufactured using the 28-nanometer process.

The production capacity of the 28-nanometer process can still provide a large amount of revenue for Zhiyun Microelectronics.

This process node is also an important source of revenue for other chip foundries such as Quadstar, TSMC, UMC, and SMIC.

The 12/14 nanometer process market is also trending towards the nanometer process market!

----

Some mature processes are nothing to Zhiyun Group. As a smart terminal and semiconductor company pursuing the most advanced technology, their goal is always the most advanced technology.

The same is true for Xu Shenxue personally!
The AI6000, which uses the most advanced technology, is a perfect example of this!

Fu Zhengyang said: "Ai6000 graphics card, as a core product serving the group's artificial intelligence strategy, in order to improve the technical level, increase computing power and reduce computing power costs, we boldly adopted a large number of new technologies!"

"In addition to the GPU core that is manufactured using the equivalent of a 7-nanometer process and has 32 billion transistors!"

"We also used the third-generation high-bandwidth memory developed by Zhiyun Storage, namely HBM3 memory, for this graphics card."

"Compared to the second generation of high bandwidth memory, the third generation of high bandwidth memory has a higher bandwidth and a larger memory capacity per chip, and can also be packaged with a larger memory capacity."

"So with this graphics card, our video memory will reach 120GB!"

Both use 3D packaging technology, and finally package a graphics card core that looks relatively large, which is then installed on a substrate with other components, and finally becomes an AI6000 graphics card.

This will be a flagship graphics card with greatly improved performance over the previous AI series graphics cards that use the 12-nanometer process. Its huge computing power can provide Zhiyun Group with more sufficient computing resources while reducing computing costs!

Although the cost of this chip will be higher, compared with its increased computing power and subsequent operating costs, this cost increase is nothing... The continuous advancement of technology will continue to reduce the cost of computing power.

Simply put, the same computing power in the past might require five AI graphics cards, with each graphics card costing RMB 10,000, a total cost of RMB 50,000, and an operating cost of RMB 1,000 per year.

But now, you only need one of the latest AI6000 to do what five older AI graphics cards did in the past. At the same time, the operating cost of a single card is only yuan. In this case, the cost of a single card rises to yuan, which also means that the computing power cost is reduced by several times, and even in some areas the computing power cost can be reduced by more than ten times.

This is Zhiyun Group, which is highly inclined to continuously promote technological progress and is willing to work at the core of AI graphics card technology advancement... The purpose is very simple, which is to reduce computing power costs.

Of course, this is only for Zhiyun Group itself. Because Zhiyun Group is engaged in artificial intelligence, when using these AI series graphics cards, it only calculates internal research and development and material costs...so the computing power cost is originally very low.

If it is an APO graphics card sold externally, its selling price has nothing to do with cost, but is based on supply and demand.

Tsk tsk, these days even laymen know that Zhiyun Group's APO graphics cards have a super high gross profit margin.

The material cost of the APO4600 graphics card is only about RMB , but Zhiyun sells it for RMB ... and the demand is still far from meeting the supply. Even big customers like Google have to queue up to get the goods.

Many people say that compared with the huge profits of Zhiyun Group's server GPU business and terminal GPU business, all the so-called high-profit businesses in the world are dregs, such as arms, illegal drugs and various so-called illegal businesses... they are all weak!

At the same time, in order to balance production capacity and sales volume, Zhiyun Group actually also takes production capacity and computing power costs into consideration when pricing APO graphics cards. This is why the APO4500 graphics card still costs just over yuan and can still be sold.

Because in terms of computing power cost, there is not much difference between using APO4500 and APO4600. They have their own advantages in different fields. It does not mean that APO4600 is necessarily better than APO4500.

Zhiyun Group sells both types of graphics cards at the same time.

Zhiyun Group controls the computing power cost in the market within a certain range by controlling the selling price of graphics cards, and gradually reduces the computing power cost with the release of new products.

This reduction is controllable and will not be too large. It is impossible to reduce the computing power cost several times by simply releasing a new product.

If this is really done, it will mess up the market... Just imagine, the APO graphics cards that we spent $1 billion on this year have only been used for a short time, but Zhiyun will launch a new APO graphics card next year, and the new APO graphics card will only cost $200 million to achieve the same effect as last year's $1 billion.

Then the graphics card assets held by this customer would immediately plummet by 80%... This depreciation is too serious.

In other words, two companies are both engaged in artificial intelligence and have invested the same billion dollars, but because there was a one-year interval between the two companies' purchase of graphics cards, you say that the company that purchased the graphics cards first is the unlucky one and only obtained 20% of the computing power of the later purchaser. Then the shareholders who purchased the computing power first will go crazy!

This will inevitably lead to many companies adopting a wait-and-see attitude, waiting for a year or two for the technology to mature and stabilize before investing in it on a large scale.

Therefore, Zhiyun Group is actually very cautious in price control of the APO series graphics cards. Its price system is not calculated based on the simple cost or profit, but on the extent of the reduction in computing power costs... It tends to maintain a relatively stable computing power cost price system.

This will allow companies that want to engage in artificial intelligence to confidently purchase APO series graphics cards now, rather than having to wait a few years.

What we want is that the other party purchases the graphics card now, uses it for a few years, and then purchases a new graphics card for upgrading after a few years.

This will enable a relatively stable and beneficial computing power market.

However, these external market environments have nothing to do with AI graphics cards. In order to improve computing power and develop the field of artificial intelligence, Zhiyun Group has tried its best to improve the performance of AI graphics cards to an extremely terrifying level... After all, Zhiyun Group's own AI graphics cards are at cost price, which is very cheap, so there is no need to worry about the price being too expensive to afford!
(End of this chapter)