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

In the following days, Chen Yiran, who was a little impatient, began to harass TED intensively.

"How's your green card coming along?"

"Seriously, sir? I'm a lawyer, not the President of the United States! I've submitted both the I-485 and I-140, and my fingerprints have been taken. The immigration office hasn't asked for any additional documents, and there's no waiting list, but that's just how inefficient the US government is. I can only wait. I've been calling me every day from Monday to Wednesday this week, but it's no use."

"Making a phone call on a rainy day is just something to do anyway."

"My Chinese is excellent! I know this proverb! Sir, please spare me. If you're so eager to buy a gun, you can just go to a shooting range and have some fun first. There's an indoor shooting range in Sunnyvale run by a Chinese person, and I know the owner well. I'll give you a discount, just stop rushing me, okay?"

"Give me the address, I'll go take a look when I have time."

Despite saying that, Chen Yiran still called TED every day and then started being sarcastic again.

Such a big law firm, yet they don't have any special channels. In movies, people just make a few phone calls and contact a few big shots and it's all settled. Fenwick & West's capabilities and connections seem to be just so-so.

On Friday morning, TED was in a terrible state and promised Chen Yiran that he could get it done within the next two weeks, begging Chen Yiran to let him off the hook next week.

Chen Yiran verbally agreed, but thought she should see how he felt next week; if he wasn't in a good mood, she could still call him. Last time, a case had cost the law firm $30, and to make the bill look better, TED had charged him for the car maintenance, claiming that it was because he frequently drove to see Chen Yiran that the car needed maintenance, so that time should be counted.

However, once she arrived at the company, Chen Yiran's mood improved considerably.

Gao Dewei reported to him that the team had completed two baseline tests and then gave him a piece of paper.

This is a complete data table compiled by Lin Zhiwei.

adaptec1, 200,000 cells.

The random baseline, HPWL, is 2.3 times the optimal solution officially published by ISPD 2006.

Simulated annealing baseline: After Amit's parameter tuning, it ran all night. HPWL finally converged, successfully achieving 1.18 times the official optimal solution.

It can be said that the first test exceeded expectations.

While reviewing the data, Chen Yiran asked Jasmine to fill in any gaps or omissions.

"Lin Zhiwei, when you were at Synopsys, what was the first result when you first ran adaptec1 using IC Compiler?"

"Between 1.15 and 1.2, a lot of parameters need to be adjusted to get close to 1.05."

"Are we now getting simulation annealing results that are close to the beginner settings for Synopsys' commercial tools in a week?"

Lin Zhiwei leaned back in his chair, somewhat excitedly saying, "This is just simulated annealing. Our real focus is on reinforcement learning training. Boss, you're amazing!"

Gao Dewei turned to look at Chen Yiran: "You're right. Simulated annealing is just an anchor point. When can reinforcement learning training begin? I'm getting impatient. I didn't expect DeepFlow's incremental computation to be so powerful!"

"Zhou Tianyou, how's the cluster status on your end?"

"The three R710s have been running at full load for two days now. Because of the incremental computation of DeepFlow as the underlying layer, we don't need to recalculate the entire graph line length every time a cell moves. Traditional simulated annealing takes tens of thousands of perturbations overnight, but we've run it nearly two million times. So, the adaptec1 only took six to eight hours to achieve a complete SA convergence."

This is just using DeepFlow as an accelerator. The situation is different during reinforcement learning training. According to our plan, DeepFlow will not only be fast in computation, but it can also explore different placement schemes simultaneously on dozens of concurrent threads and then summarize the experience.

Chen Yiran nodded and said, "I understand what you mean, so you're saying that GPUs are urgently needed, right?"

Gao Dewei nodded and said, "Plans can't keep up with changes. Boss, your DeepFlow is excellent. I didn't expect it to work so well. My original plan was to verify that the baseline was running before considering GPU expansion. According to my initial estimate, the toolchain setup and baseline would take at least three or four weeks, and if things didn't go smoothly, it might take one or two months."

Zhou Tianyou agreed, saying, "Reinforcement learning is different from simulated annealing. Simulated annealing is a traditional algorithm, and a CPU is enough. But reinforcement learning requires training neural networks, and a single iteration on a pure CPU takes hours. I suggest we buy GPUs; each R710 can hold two."

Raj, feeling he finally had a chance to use his skills, quickly spoke up: "The C2050 will do. NVIDIA internally uses the Fermi architecture test card for CUDA verification, which is from the same generation as the C2050. When I was on the verification team, I ran many computationally intensive workloads, and the double-precision floating-point performance is sufficient for placement neural networks."

"Okay, this is probably not something we can buy at Best Buy."

Chen Yiran took out her phone and called Linda.

"Linda, please help me order six NVIDIA Tesla C2050 motherboards and corresponding Dell PCIe expansion kits for my R710. They should arrive today at the earliest, and Monday at the latest."

"Okay, Mr. Chen, I can help you contact them. Dell and Nvidia have very convenient shipping options in Silicon Valley, and they should be able to deliver everything to you this afternoon."

After hanging up the phone, everyone in the company stared at Chen Yiran in disbelief.

Gao Dewei immediately launched into a rant: "Is this what rich people are like? We can just contact Dell ourselves, why do you have to order others around?"

"That's called efficiency, saving time."

Chen Yiran then looked at Zhou Tianyou and asked, "Enterprise-grade GPUs shouldn't be too difficult to install, right? I've never actually seen one before. Will it be installed as soon as it arrives this afternoon?"

Zhou Tianyou nodded: "Easy to set up, it was done in no time. We should be able to run the first round this afternoon. We'll start with Adaptec1, on a small scale, to get the end-to-end reinforcement learning process running. If it's just for validating the process, it'll take a day or two, but to see line-length data that's significantly better than simulated annealing, it'll probably take another week. I need to carefully adjust the hyperparameters of the neural network."

"Great, you see, because we found someone, the goods will arrive this afternoon and we can start working. If we had ordered it ourselves, we might not even get it until Monday. Alright, meeting over, everyone back to your workstations."

Instead of leaving immediately, the group began to check each other's information, change the to-dos on the whiteboard, and archive previous project progress.

Chen Yiran felt that the company atmosphere was pretty good. Perhaps all startups are like this, and the employees are still very motivated.

To boost morale, Chen Yiran decided to treat everyone to lunch at work today.

Although there weren't many good takeout options nearby, I ended up ordering Papa John's.

But the way a group of people chatted about future plans and technology while waiting for the equipment, and shared photos of pizzas, really resembled a scene from a movie about a startup struggling to survive.

JPMorgan Chase usually keeps its word, and at 1 p.m., a Dell delivery truck pulled up in front of the office. They also brought some Nvidia goods, since they were all complementary items.

Zhou Tianyou, along with Sangjie and Liu Yi, unpacked and inspected the several boxes to confirm that all the goods were there before moving them inside to begin loading the machine.

Opening the R710 and adding the card required disassembling the air duct, but fortunately Raj was very skilled at it, and the other tech guys were also quite capable. They managed to get it done in about twenty minutes by fiddling around the rack.

At 2 PM, the first line of CUDA code ran successfully, and by 3 PM, the reinforcement learning training environment was fully configured.

This is a milestone event for the company.

Surrounded by the crowd, Chen Yiran took a deep breath and typed the first training command on the terminal. The log on the screen began to refresh: matrix multiplication, gradient update, and loss descent curve.

The loss decreases very slowly after it starts, which is normal. Neural networks don't know anything at the beginning; they're basically completely unintelligent.

It takes at least two or three days to see results.

That evening, Chen Yiran wanted to organize a team-building activity again, but Hannah called him and said that her aunt was supposed to come and visit their family. However, her aunt changed her mind at the last minute and had to go to a music festival, so she stood them up, which made the little girl, who was very close to her aunt, very unhappy.

Since the little girl came to complain, Chen Yiran definitely had to comfort her young heart.

Moreover, Chen Yiran suddenly remembered many funny jokes she had seen in her past life, where the boss always inexplicably liked team building activities, especially after get off work hours.

"Terrifying!" Chen Yiran broke out in a cold sweat. "Is this the curse of being a boss?" Becoming a boss has made her crave all sorts of team-building activities.

After saying goodbye to everyone, Chen Yiran was the first to leave work, chatting with Jasmine as she went.

"Jasmine, next time I want to organize a team building activity, you have to remind me so I can see if it's appropriate."

Okay, I got it. But actually, I think it's perfectly reasonable to have a team-building event tonight, given it's a significant company event. ╮(╯▽╰)╭

"Alright, just remember to remind me."

We drove all the way to Santa Clara University and picked up the indignant young girl.

"The music festival is at the end of the month! I think my aunt just wants to hang out with her new boyfriend, so she's not coming to see me."

"Yeah, yeah, that's terrible."

"She doesn't care about family ties at all; her relationship with this boyfriend definitely won't last."

"Exactly, it won't last."

"I'm so angry, let's go to the music festival too!"

"Yeah, yeah... wait a minute, what music festival?"

"The Montreal International Jazz Festival seems to be the world's largest jazz festival."

Chen Yiran started racking her brains, trying to find a reason to refuse, and quickly came up with a very strong one.

"Sigh, Montreal is in Canada, right? I'd love to go, but unfortunately, I don't have a Canadian visa."

"Is that so? What a pity."

At that moment, the music playing in the car stopped, and a phone call came in.

TED's voice came through the car stereo: "Wow, great news! You don't have to wait until next week, your green card has been issued. You're a permanent resident now!"