Future Super Intelligent System

Zhan Kaihe was unconsciously moved by Liu Fan, but he calmed down quickly, "This is indeed a beautiful vision, but this matter cannot be resolved overnight."

"Although it does take time, it may not take too long." Liu Fan took out the tablet, "With the current mainstream deep learning algorithm, if it is installed on the current general-purpose chip, there will be huge performance waste and matching problems. To solve this problem, the first step is to make a breakthrough in architecture.

When I talked to them about the chip architecture at HiSilicon today, I didn't go deep into it, but in front of the two of you, I think I can show what I really want to do. "

Ji Mushui and Zhan Kaihe approached the tablet to study for a while, and Zhan Kaihe said, "Dynamic discrete flow calculation."

"That's right, this is the core of the entire algorithm architecture." Liu Fan leaned forward, and Zhan Kaihe noticed the most important place at first glance, which made Liu Fan have a good impression of him.

Zhan Kaihe looked calm, but nodded lightly, and continued to look down at Ji Mushui, and after a long time, he took off his eyes and rubbed them, "No wonder Mr. Ji always praises you in front of us, theoretically In terms of performance, this is indeed an optimization of the existing FinFET process, and it can also greatly reduce the power consumption of the chip. Unfortunately, there is no experimental data.

In addition, there is still no way to solve the problem of matching the depth algorithm with the general-purpose chip, but it is already a great thing for you to have such a technological invention, and you are worthy of being a legend in China. "

"This plan really fails to solve the problem of how to produce AI general-purpose chips, because it is designed for AI-specific chips. I have actually made a theoretical plan for AI general-purpose chips. You can take a look at it." Liu When Fan proposed to them to make an AI general-purpose chip, he decided to reveal something a little deeper.

Both Zhan Kaihe and Ji Mushui were a little surprised. After Liu Fan opened another document on the tablet, the two concentrated their energy again. Ten minutes later, Zhan Kaihe's face was barely pasted on the tablet, obviously not as calm as before, "Xiao Liu, where did your inspiration come from?"

Liu Fan leaned forward again, "Actually, the principle of this thing is to learn from the flow of rivers on our planet into the sea, so I named this structure called Living Water Structure."

Zhan Kaihe nodded frequently, "It's a pity that there are still too many problems to be solved, just the material problem...Ji Lao, is there any nano-scale material that can withstand this kind of curvature?"

Ji Mushui shook his head, "I don't think so, not just the material, but his hook-like connection method, no one should be able to do it."

"Actually, it's not difficult to implement this linking method. Long-term training should be no problem, as long as a suitable tool can be made." Liu Fan explained, and demonstrated a general process with his hands.

Ji Mushui and Zhan Kaihe looked at each other, and Ji Mushui said, "I think the core problem of this structure should be the material. If the material problem can be solved, it is worth exploring in depth."

"So this is the chip design architecture that I can think of that is most likely to satisfy AI general-purpose chips. The development of materials has been very rapid in the past two years. I think we can be optimistic about the future of materials." Liu Fan did not think The material solution has also been brought out, but it is not necessary now, "and most importantly, this architecture can be split in the design of some dedicated chips, so that the current materials can completely drive this design.

So my current plan is like this, divided into three steps, the first step is to use the improved architecture to quickly optimize the existing AI-specific chips, and I personally think that the best way is the mobile phone chip, because this piece of HiSilicon The experience of the company is very rich, and the user coverage rate of the terminal products is high, and the social attention that can be attracted is also wider.

The second step is to try to use the living water architecture to design and produce AI-specific chips and conduct market inspections. Once successful, the efficiency of AI chips will definitely be significantly improved.

The third step is to find materials and work hard to develop the AI ​​general chip we want. "

After Liu Fan finished speaking, Ji Mushui and Zhan Kaihe fell into deep thought at the same time. After a while, Ji Mushui asked Zhan Kaihe, "Kaihe, do you think this is reliable?"

Zhan Kaihe took off his eyes and said seriously, "Everything is at the theoretical level now. I think the risk is too high and the cost of failure is too high after a rational analysis."

There was silence in the car, and a few seconds later, Zhan Kaihe said, "But I don't know why, I really want to take a risk."

Liu Fan and Ji Mushui looked at each other, and then comforted, "Actually, because we walked in three steps, the cost of trial and error is not as high as imagined. If you can't go on the first step, there is no second step." Step things up.

And the success of each step, I think the key is not the upgrade of chip performance, but the improvement of our research and development capabilities and technological level. Even if we cannot make breakthroughs in the final production of chips, the experience and progress gained in the process are believed to be beneficial to the entire China. Chip development is also a huge asset. And I think at least in the optimization of mobile phone AI chips, I have enough confidence. "

As soon as the words were finished, the car stopped at the entrance of the hotel. At this moment, Ji Mushui patted Zhan Kaihe's shoulder and smiled heartily, "Kaihe, it's time to eat."

Zhan Kaihe's heart moved, and he understood what Ji Mushui meant in seconds.

...

Two days later, in the conference room of Wood Dragon Technology, Ji Mushui and Zhan Kaihe sat on one side, Wang Tao, He Tingbo, and Zhang Rujin sat on the other side, and some backbones brought by various companies sat around the conference room.

And the eyes of these people were all focused on Liu Fan who was standing in front of the stage.

Liu Fan explained the whole plan in detail through PPT. This time, he explained the division of labor, model, budget and future planning more clearly than in the previous private chats with all parties.

Everyone sitting in the audience listened and suddenly realized a problem. From the front-end architecture design to the back-end process improvement of this chip project, Liu Fan actually brought all the core technologies to everyone?

These three parties suddenly felt inexplicably as if they were here to take advantage. At the same time, everyone couldn't help wondering, since Liu Fan has this ability, why does he need to cooperate with everyone? Wouldn't it be over if you directly recruit people to form a team? This is unreasonable.

Just as everyone was wondering about this issue, Liu Fan finally talked about the most sensitive topic today: how to cooperate...