The phone kept vibrating next to it, but it was better to be as if it wasn't felt it.

Recently, he has almost completely let go of the turbulence algorithm and devoted himself to the EDA software.

It is no problem to go to the turbulence algorithm. In fact, the reason why this project has been conducting internal testing in the laboratory and has not yet conducted large-scale public testing of external contacts is also because it has encountered technical bottlenecks.

At present, the erroneous judgment rate of the turbulence algorithm in the laboratory is stable at around 100,000 and cannot be further reduced.

The error rate of 1.8 in 100,000 does not sound high, but it is terrible if applied to ticketing websites and APP servers like 12306.

The user base is too large.

Every 100 million people use 12306 ticket bookings, more than 1,800 people may be misjudged as malicious crawler programs and punished. This is a big defeat for people. Not to mention that more than 100 million people use the 12306 ticket booking system every year?

There is more than a long holiday!

Ning Wei and the three auxiliary researchers have thought of ways. Several people have brainstormed together and have gone through two iterations of the algorithm, but they have never been able to reduce the error rate of laboratory tests.

This made Ning Wei feel that further breakthroughs in the turbulence algorithm may require his understanding of the entire system to be further deepened. This is too urgent. He simply stayed in the laboratory first and did an internal test for a period of time. There was enough data available for analysis, and maybe some clues could be found from these data.

So Ning Wei no longer pays attention to the turbulence algorithm. After all, no one in the EDA software project team knows the structure and difficulties of this software better than him.

The biggest difficulty is actually to simplify chip design.

Let's give a simple example.

When using mainstream EDA software for back-end design, the first step is generally to set up the standard library.

The required data can be roughly divided into three categories, namely logic layer, physical layer and design.

These three categories can be subdivided into six types of files, including process files, unit libraries, sub-stores, integrated netlist files, TDF files that constrain Pad locations and timing constraint files given after integration.

Then there is the overall layout of the chip, which is an important step that directly affects the chip area, speed signal integrity and design cycle. The corresponding positions must be assigned to the module, input and output interface, power pad, etc.

Next is clock tree synthesis, which is also a big challenge. Think about it, after gathering tens of millions or even hundreds of millions of transistors, it is not easy to transmit the clock signal to every local area of ​​the chip with zero deviation.

Next is overall wiring, which is also the most important physical practice process in the entire chip design.

Then, the design rules check will be carried out.

After all these are done, there is no problem, and then enter the simulation stage...

According to Ning Wei's idea, the above steps must be completely reduced to one step, and integrate the first-level chip architecture that completes a specific type of task in the object library, that is, the container, and then the second-level architecture is the group, and then divide it down in turn...

The operator only needs to make simple combinations and after simple patching and deleting of the existing structure, he can directly enter the simulation stage.

This is also the reason why this EDA software is simply designed to design various AI chips.

If you want to design a general-purpose chip in this fool-like way, it would be better to use up the problem of fool-like operation even if you really use up your hair.

But it can be done in the field of AI chips.

As for the current situation in the AI ​​field, the so-called intelligent algorithm is a process of quickly seeking the optimal solution. These algorithms solve complex engineering problems by simulating some natural processes.

There is also data-based machine learning, which mainly starts from known data and is used to predict unknown data, etc.

So Ning Wei’s design concept is very simple.

For example, first, it is equipped with a container for standard convolutional neural network learning, and then it integrates groups of various machine vision categories under the subclass, such as picture recognition, face recognition, advanced expression recognition, gesture recognition, etc.

Let group designs be arranged in groups relatively scientifically to solve specific tasks.

The operator only needs to enter the chip type he wants to design, and first it will jump out of the container framework in the object library, and then the operator will choose the functions to be implemented under this framework.

It is difficult for software designers to make the operator feel simple.

Ning Wei's goal is to design at least five container structures and corresponding fusible group structures in the object library before the launch of this EDA.

It is simple after being introduced to the market. If someone uses it, a high-level design team can design containers and supporting groups, groups and components by themselves after learning various standards and instructions of the software, or use existing defined components to redesign container structures and supporting groups and groups. After uploading, after being verified by the audit department many times, it will be included in the object library.

But at this stage, he has to do these tasks by himself. After all, the professors or programming skills assigned by the laboratory are good, but the mathematics is still a little worse.

Finally, after three hours of busyness, I used the components I defined earlier to build a group structure under a container. I would rather breathe a sigh of relief when leaning on the chair. It is really tiring to do scientific research projects, especially when I encounter difficult problems.

The turbulence algorithm encountered a bottleneck, and EDA was inseparable from others. Ning Wei began to reflect on whether he was really too greedy and thought about scientific research issues too simply.

I picked up my phone casually and found that the dormitory group was particularly lively today.

Dozens of messages have been browsed early in the morning.

My mind was a little bored, just to see what these people were making trouble and take a break.

Then I saw the complete steps of the classic face change.

At first, several people took turns to @ him in the group, and the threat was very strong.

"Ning Wawa, you haven't returned to the dormitory yet, we have decided to expel your dormitory!"

"Ning Doll, dads miss you."

"Xiao Ning, your wings are hard, right?"

Perhaps seeing that he hadn't responded, his tone began to soften.

"Brother, stop making trouble. If you have urgent matters, please reply to the news."

Later, it was completely softened.

"Brother Ning, Mr. Ning, you can't really stop saving you from death."

In the end, they were completely obedient, and some even lost their moral integrity.

"Dad, Dad Ning, please reply to the news!"

It's pretty good. It's hard to describe the psychological satisfaction brought by seeing the roommates calling him dad on their own initiative.

So he went back to "?"

Mr. Xuan first reacted: "Fuck, Ning Dowager, you finally didn't pretend to be dead? What should we do with our paper? You're giving an idea!"

"I'll tell you, don't make trouble!" Ning Wei replied with a panda's head expression.

"Dad Ning, please give me some advice on the children's papers!"

"You can teach me!" Ning Wei returned with a panda head expression.
Chapter completed!