Reborn in 2010, the AI ​​in my mind is too powerful.

After a pleasant trip to Los Angeles, Chen Yiran didn't wait until Monday; instead, she started working over the weekend.

Previously, I had compiled all the information on the ISPD 2006 placement benchmark, as well as netlist files for several test cases, LEF format descriptions of the standard cell library, and official documentation on line length and density constraints.

The data input format is already understood. LEF describes the physical dimensions and pin locations of a standard cell, while netlist provides the connectivity between cells. HPWL is the half-perimeter line length model. The width plus height of the bounding box of each net approximates the line length.

"You mean, you can design the parser directly?"

Yes, but this is only the first step. The format of the benchmark itself is not complicated; the complicated part is how to feed the parsed data into DeepFlow. The editor's AST is a one-dimensional tree structure, while the placement is a two-dimensional space with three-dimensional connectivity.

The core change is that the trigger condition for incremental updates has changed from text changes to spatial position changes. When a cell moves, the affected nodes are no longer sibling nodes on the syntax tree, but rather spatially adjacent cells and cells connected on the netlist.

We need an adapter whose function is to receive the coordinate changes after each cell movement by the placement engine, convert them into a delta description that DeepFlow can understand, and then feed the gradients calculated by DeepFlow back to the placement engine.

Chen Yiran started drawing in his notebook, and Jasmine analyzed the diagrams while asking him to make corrections.

Despite the long-absent low-grade fever, Chen Yiran managed to come for two days.

The final output was a 13-page design document, which included: the benchmark parser architecture, the interface definition of the DeepFlow adapter layer, and the data flow conventions between the placement engine and the GPU training cluster. The input and output formats, data structures, and call order of each part were all clearly written out.

On Monday morning, Chen Yiran printed out the design documents and sent them to everyone. Catherine also wanted a copy, but Chen Yiran refused, saying it would be a waste of paper.

Gao Dewei looked at the document for five minutes, then put it down and looked at Chen Yiran.

"Boss, no wonder you're the kind of person who can sell 200 million on your very first project."

Gao Dewei stood up and walked to the whiteboard. He redrawn the architecture diagram from Chen Yiran's document. After finishing, he took two steps back and looked at it for a while.

"The design of this adaptation layer is correct. It changes the triggering condition for incremental updates from the tree structure of the AST to a hybrid index of spatial and connectivity aspects. This is the most fundamental difference between placement and the editor."

Zhou Tianyou saw the data flow conventions between the GPU training cluster and the placement engine and said that he had just gotten the CUDA environment and MPI working on three R710s over the weekend. He originally thought it would take another week or two to use them, but the architecture document had already defined the interface.

The others each took a piece of the pie chart. Zhang Minghui and Liu Yi were in charge of the benchmark parser, Sang Jie and Wang Xue were in charge of loading data for the LEF/Standard Cell Library, and Amit and Lin Zhiwei began translating the physical constraints of ISPD into a format that DeepFlow could process.

The whiteboard was full of work assignments. Chen Yiran glanced at it and realized that she didn't have much to do.

"I won't be coming over starting tomorrow. You guys do it yourselves. If you have any problems, contact me on MSN."

Gao Dewei glanced at him: "Boss, you're staying home now?"

"I'm more efficient at home."

Gao Dewei agreed with this statement, as he was the most senior and experienced person in the company.

He was actually a little worried a few days ago. After all, Chen Yiran took everyone out to play on the second day of the business opening. Although this behavior would make the employees happy, it was not helpful in terms of the company's progress.

What's even more terrifying is that although he assigned people on the opening day and the route was actually quite reasonable, Chen Yiran didn't give a rough timetable at all.

Gao Dewei himself calculated that if it were normal and relatively smooth...

It takes 1 to 2 weeks for the team to gel and for the toolchain to be set up before the ISPD 2006 benchmark can be compiled and run.

It will also take 2 to 3 months for DeepFlow to adapt to placement before it can adapt to incremental computing engines for 2D spatial data.

Finally, the first benchmark results will be available in March or April, and then they will need to be compared with preliminary data from traditional methods.

It will take at least six months for the RL training framework to be fully operational.

Then, after about a year, the company can produce a usable commercial prototype with a stable, comparable, and demonstrable engine.

This is all assuming the best; any problem that arises along the way could cause a long delay. Without a detailed project timeline, the entire team could become somewhat disorganized.

On the day of the opening, Chen Yiran didn't mention any of these things at all, and Gao Dewei sensed that he probably had no concept of them whatsoever.

That makes sense, considering that Chen Yiran was previously an independent developer and had absolutely no experience in overall project management.

Gao Dewei originally thought that after Chen Yiran discovered the problem, he might suggest gradually introducing someone to be the overall project manager.

Gao Dewei could do this job himself, but he didn't really want to because he's too old and doesn't want to be so tired.

Chen Yiran doesn't think he's a diligent person. Among the team right now, Raj is the only one who seems suitable.

However, in just two days, Chen Yiran showed Gao Dewei what a genius was.

He almost single-handedly solved all the problems within the visible range and designed everything that needed to be designed.

This kind of person doesn't need to work; a few online directives are enough to propel a project forward at breakneck speed.

For the next few days, Chen Yiran basically slept until he naturally woke up, and after walking Pidan, he would watch dramas and communicate with his employees online.

There were quite a few problems, but most of them could be solved by figuring them out on their own after a while. The things that really needed his help were things that Jasmine could see at a glance.

Lin Zhiwei encountered a format version issue while parsing LEF files. The data from ISPD 2006 was still in the old LEF 5.6 format, and his parser, written according to version 5.7, reported an error when reading the OBS segment. Jasmine glanced at it and knew that the layer of the OBS segment had been modified to be nested between 5.6 and 5.7, so she added backward compatibility.

When Zhang Minghui was running a large case study, the memory crashed. He habitually read the entire netlist into memory before building the structure. The BigBlue4, with two million cells, consumed 16GB in one go.

Chen Yiran instructed him to switch to streaming parsing, building the graph as it reads, eliminating the need for a complete copy in the middle. Zhang Minghui, having come from EA, was familiar with streaming resource loading in game engines, so this task should be easy for him. Sure enough, the refactoring was completed the next day, reducing memory usage by half.

Amit encountered a rather tricky problem: the fixed position constraints of the macro cell and the density constraints of the standard cell conflicted in the same region.

A constraint resolution layer was needed, prioritizing macros, with standard cells only effective in the space not occupied by macros. Jasmine tidied up the logic, wrote a few lines, and sent them over. After Amit added this layer, all constraints passed, and then he sent over a bunch of compliments.

Apart from that, the team didn't encounter any major problems. It can be said that with Jasmine's one-on-one guidance, the company's integration was very fast, and they basically completed the task that Gao Dewei had predicted would take more than half a month in just one week.

On Friday afternoon, Chen Yiran drove to the company.

Gao Dewei ran the first integration test in front of the terminal. The log on the screen refreshed for about thirty seconds, and the last line showed: ISPD 2006 adaptec1 loaded and validated.

There was a few scattered claps in the office.

Chen Yiran walked to the whiteboard. The architecture diagram drawn on Monday was still on it, but now it was covered with arrows and annotations, all added by everyone over the past week.

"Next week we'll start the first placement, Adaptec 1, 200,000 cells. We don't need a long production line, just a smooth workflow."

Lin Zhiwei leaned back in his chair and thought for a moment: "For the first run, I suggest using the simplest random initialization as a baseline. Don't do any AI training, just randomly place cells and let DeepFlow calculate the line length. First, verify the data throughout the entire process, from parsing to the engine to line length evaluation, to ensure that the data is correct at every step. Only after the baseline is confirmed to be fine should we implement RL."

Gao Dewei nodded: "Agreed. And this baseline itself is useful. You can run the simulated annealing on Adeptec1 at the same time, using the same benchmark conditions. Once the RL is trained, these two baselines will be the anchor points for comparison."

Zhou Tianyou, still holding half a roll of network cable, said, "There's a problem with the training cluster. The R710 only has CPUs, no GPUs. If placement RL training runs purely on the CPU, the iteration speed for 200,000 cells will be very slow. DeepFlow's incremental computing engine can significantly reduce the computational load per iteration, but the large amount of concurrent exploration required for RL still needs GPU acceleration."

Amit asked, "Didn't you say before that all three R710 units were working properly?"

"It's working, but the R710 is a server-grade Xeon, without a GPU card. Running deep learning training on the CPU alone is only about one-tenth as efficient as using a GPU."

Gao Dewei turned to look at Chen Yiran: "Let's run some smaller tests on the CPU this week. Once the benchmarks are running smoothly, we can consider upgrading the GPU?"

Chen Yiran glanced at Jasmine's statement and replied, "Let's first look at the single iteration time of adaptec1 on a pure CPU. DeepFlow's incremental updates may be more computationally efficient than you think. If it can run tens of thousands of iterations in one night, that's enough for the initial prototype with 200,000 cells. We can then add GPUs for large-scale training later."

Raj said somewhat helplessly, "Hurry up, I feel like this is all I can do right now."

Amid nodded and said, "Next week I can finish translating all the constraints for the eight benchmark cases. This week I only finished adaptec1."

"How long will it take to complete all eight?"

"Two weeks. The later ones are more complex; the Big Blue series has additional density constraints due to macro corners."

"Focus on getting the adaptec1 running smoothly first; we can add other benchmarks later."

After another round of discussion, the whiteboard was filled with more arrows and TODOs.

That's how it is in Silicon Valley, especially in startups. Everyone's very enthusiastic at that time. When things get excited, some of the single guys even start suggesting that they don't go home tonight and just stay at the office.

Chen Yiran quickly retreated. Today was Friday, and she needed to go home to video chat with her parents. They should be able to finalize the purchase of the house in China this week.