Chapter 910 AI Symbiotic Development Technology

The deep learning framework is naturally followed by the AI ​​chip, and the host is the first to pave the way, "The explosion of deep learning cannot be separated from the deep learning framework, and it is also inseparable from the AI ​​chip.

It can be said that the root of the explosion of deep learning is the deep learning framework, and the beginning of the explosion of deep learning is the GPU platform established by Dafeng Group, which pushed GPU into the field of deep learning.

Needless to say, the advantages of GPU in the field of deep learning. Today, GPU is still the most commonly used deep learning computing unit.

But we have also seen that with the continuous development of deep learning, the limitations of GPU are becoming more and more prominent. Deep learning includes two computing links, training and application. GPU is very efficient in deep learning algorithm training, but it cannot show the advantages of parallelism in application. At the same time, GPU cannot flexibly configure the hardware structure, and the market needs new development.

This leads to the current two new fields of AI chips, FPGA and ASIC. FPGA is mainly used on the device side due to its programmable specificity and high performance, while ASIC is mainly used on the device side due to its small size, low power consumption, and low cost. Features such as high confidentiality are more used in consumer terminals.

Judging from the current corporate structure, Galaxy Technology is the largest player in the GPU field, followed by Nvidia.

The leading ASIC company is Dafeng Semiconductor, and the main players include Huawei, Google, Sony, Ali, etc.

The FPGA leader is Intel, and major players include ACTEL, Xilinx, and China Microelectronics. "

"It seems that GPU and ASIC are both led by China, and the United States currently only has an advantage in FPGA." The audience began to discuss again.

"GPU is the current mainstream, ASIC is the future trend, and FPGA feels caught in the middle."

"ASIC still has a long way to go, and FPGA's offensive in the field of cloud computing is very fierce. If you grab the market first, you will be able to monopolize it later."

"The biggest problem with ASICs is the limitation of special use, but as long as Huaxia enterprises make great efforts and achieve a detailed division of labor like the deep learning framework, this problem can be solved with the most primitive method.

Don’t forget, although the United States is the core player of FPGA technology, China is the largest market for FPGAs. As long as Chinese companies can replace chips in various fields, the entire market can completely abandon FPGAs. Of course, the premise is that if it is necessary if. "

The audience was discussing, and people on the stage began to change. Just like the introduction of the deep learning framework, the host finished the preparations, and representatives of various companies began to speak on their own products and industry development.

Huaxia enterprises still occupy the home field advantage. Most of the representatives on stage are representatives of Huaxia enterprises. Everyone also revealed some common information in their speeches. For example, AI chips will rely heavily on technological development.

Those who understand will naturally understand.

The common point that is more concerned by everyone is the collective explosion of Chinese companies in multiple AI chip fields.

Among them, the most close to the public life is tantamount to the outbreak of wearable AI chips in China. Nine Chinese companies including Gale, Huawei, Xiaomi, ZTE, and Haier have launched their own wearable AI chips in 2018.

"Does the outbreak of wearable AI chips indicate that Dafeng Group's AR glasses have taken another step towards the consumer market?"

"It's more than promoting AR glasses. The growth rate of wearable smart devices has exceeded 30% for three consecutive years. This is a huge future trend. Now wearable AI chips are actually covered by Chinese companies."

"Hasn't any American company entered the game in this field?"

"It seems that there is no such thing as Huaxia has absolutely taken the lead."

"There is another important point that cannot be ignored. All wearable AI chips use the Hongmeng architecture, which is paving the way for the creation of an ecology. After such an explosion of wearable AI chips from Huaxia Enterprises, after AI chips have been installed in terminal products, This ecology completely falls on Huaxia's side."

"If the Dafeng Group works harder, when it can compete with Intel in traditional CPU technology, and with the advantages of the Dafeng Group's process, it will really be able to completely defeat Intel."

"Dafeng Group has already caught up with or even surpassed Intel by relying on its technological advantages, right? Isn't Intel's 10nm+ market response not very good?"

"There should still be a gap in the overall process, right? No one can tell now. What the two companies say to the outside world is full of water."

Amidst the discussions in the audience, Liang Mengsong of Gale Semiconductor stepped onto the stage.

Liang Mengsong first introduced the main chip development of Dafeng Group in the field of AI semiconductors, and then started the last topic of today, "The development of artificial intelligence, especially deep learning, in recent years has further confirmed a thesis of our company many years ago. The development of artificial intelligence requires a new hardware architecture to meet the exponentially growing demand for computing power, which is the hardware revolution we mentioned earlier.

This hardware revolution has obviously kicked off, and in the process of driving this hardware revolution forward, everyone has encountered a problem.

For a long time, the design of chips is time-consuming and labor-intensive, which seriously hinders the iteration speed of chips. This is why semiconductor companies will have a five-year, seven-year or even ten-year development blueprint. "

Speaking of this, Liang Mengsong suddenly paused before continuing, "One of the important reasons why Dafeng Semiconductor has made such achievements in the semiconductor field is our accurate prediction of the development of the semiconductor industry.

And this is why some semiconductor companies have always wanted to monopolize this industry, because the more monopolized, the more accurately they can predict the future. After all, for monopolies, future development can be controlled.

But for the rapid explosion of artificial intelligence, firstly, there is no so-called monopoly semiconductor company, and secondly, the product iteration is too fast, and the five-year, ten-year plan cannot adapt to the field of artificial intelligence at all.

Therefore, the only way for the development of AI chips is to find a way to compress the design time, but for human experts, in the face of increasingly complex chip evolution, there is no ability to compress the design time until we find an interesting one. Technology development direction.

That is to use deep learning to promote the hardware revolution.

Yes, you heard me right, the hardware revolution drives the development of deep learning, and the development of deep learning in turn drives the hardware revolution, which is why I say interesting.

We call it: AI Symbiotic Development Technology. "

Liang Mengsong began to show a document, "We use deep learning to let machines participate in the update and optimization of chips by learning from past chip design experience.

Our earliest experimental goal is global routing, because this is the most complex and time-consuming stage in chip design. We use deep learning algorithms to learn from past layout netlists and successfully create new netlists that have never been seen before. Generate an optimized chip design, and more importantly, we compress the time by 10 times. "

Liang Mengsong began to interpret the document in detail, using a lot of detailed data to show the feasibility and efficiency of using deep learning to reverse the development of chips.

"In the near future, we are likely to see a scenario where we just heard about 3nm mass production in the first half of the year, 2nm mass production in the second half of the year, and 1nm mass production in the next year. I believe in this future. will come soon.

Moreover, this technological breakthrough can not only compress the chip design cycle, but also help small and beautiful companies develop dedicated AI chips at a relatively low cost, further promoting the explosion of dedicated chips.

We will provide this technical support to global enterprises from now on, and contribute to the hardware revolution and the development of AI chips. "

"Let me take a look at the current situation. From the underlying chip to the deep learning framework, to the general AI technology, and finally to the application of artificial intelligence, Huaxia has opened up this line in the development of artificial intelligence."

"And if you just pick up an industry, it becomes an industry chain. Take chips as an example, design, manufacturing, materials, technology, and now there is AI symbiotic development technology, and the hardware revolution is just around the corner.

Intel's manufacturing process is afraid that it will never be possible to catch up with Dafeng Group. "

"Don't talk about catching up, Intel should be lucky not to be left behind."

"This trick is a bit powerful. The hardware revolution will be promoted through AI chips, which will directly end Moore's Law, and then end Intel's long-standing monopoly."

"I'm a little confused now. As far as China's technological level and market share are concerned, how can American companies have the confidence to restrict them?

Huaxia Enterprise doesn't restrict them, they should be thankful. "