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

Wednesday's demo was no surprise; Chen Yiran completely won over Professor Wu's experimental group.

What is GPUmat? Simply put, it's a GPU plugin for MATLAB.

For each step of the calculation, it has to copy the data from the CPU memory to the GPU memory, and then copy it back after the calculation is complete.

A four-layer autoencoder needs to perform twenty back-and-forth passes to complete one forward and backward propagation cycle. In 2010, the theoretical bandwidth of PCIe 2.0 was eight gigabits per second, but in practice, it was less than five gigabits per second. This means that four Tesla C2050 processors plugged into the machine were mostly not performing computations, but simply waiting for data.

DeepFlow's framework comes in handy here. After the computation graph is compiled, all the data remains in GPU memory and is never copied out. Adjacent operators are merged whenever possible, and the four GPUs each receive a quarter of the data and run in parallel. Finally, the gradients are summed. The hard drive and CPU only move when loading data at the beginning and when backfeeding the loss function at the end; otherwise, the PCIe is completely idle.

Therefore, for the same model and the same STL-10 dataset, GPUmat takes an hour to run once, while DeepFlow only takes three minutes. GPU utilization is increased from over ten percent to ninety percent.

Andrew Ng had expected the results to be good, but he hadn't anticipated them to be this good; even his doctoral students were stunned. After repeatedly confirming the various results, he began directing his research staff to work and, together with Chen Yiran, confirmed that they could upload the paper to the online placeholder platform arXiv within two days.

Chen Yiran was very cooperative. Two days later, after receiving notification that Professor Wu's paper had been published online, he updated all the new operators on GitHub.

After that, Chen Yiran didn't pay any more attention to it. After all, it was Friday again. He went to the company for a walk and encouraged the employees.

On Saturday and Sunday, I took Pidan to the company to visit and show my appreciation to the employees.

During the two days that flowEDA worked overtime, Microsoft's senior management also started a closed-door meeting. The topic of the meeting was the same as before: how much money to report to Chen Yiran.

As a top tech giant, Microsoft Research has established a dedicated group in Redmond, even though deep learning isn't yet a mainstream field. Andrew Ng, a leading figure in the industry, is naturally a focus of their attention. Checking online updates and reading research papers are essential daily tasks for them.

So when they saw Andrew Ng's paper on arXiv on Friday morning, they immediately paid close attention to it and found Chen Yiran's Deepflow in the notes.

Just then, Chen Yiran updated the operator, and they hurriedly conducted tests. The amazing results demonstrated an efficiency that was unusual for Microsoft.

After a day of tinkering and running various tests all night, the performance comparison report was on the desk of Avner, Vice President of Mergers and Acquisitions at Corporate Development, on Saturday afternoon. The astonishing data comparison shocked even Avner, whose technical skills were not particularly strong. Following a full-scale investigation by Corporate Development, they quickly discovered something that caused Microsoft some embarrassment.

Avner immediately reported this to his immediate supervisor, Bob Mullia, the president of corporate development.

Half an hour later, Mullia walked into Ballmer's office with two documents in his hand.

Ballmer looked up at the person who had come in and asked with a smile, "Why are you coming to work on a Saturday? That's not like you at all."

"Something big has happened, that's why they called me over. Andrew Ng, a leading figure in deep learning, just published a preprint in his lab, using an open-source framework called DeepFlow to run experiments. It's twenty times faster than GPUmat. But that's not the key point. The key point is that the author of this framework is Chen Yiran."

Ballmer stopped.

"The one who sold us the editor a few months ago?"

"Yes. It cost 200 million yuan at the time, a pure asset acquisition. After the purchase, he took the money and left, turning to chip design. Now he runs a company called FlowEDA. DeepFlow was originally an internal tool that he casually open-sourced. Now Andrew Ng's experimental group has proven that the scalability of this framework is incredibly exaggerated."

Ballmer flipped through the two documents Mulia had brought, then placed his hand on the call button: "Have someone from the legal department come over."

Although it was Saturday, companies like Microsoft clearly have staff on duty in key positions. A short while later, a bald man in his forties walked in.

After looking at the document, he made a few points.

"This time might be very different from last year."

"How are they different?"

"In last year's deal, Chen Yiran used a foreign identity to join the Delaware shell company. We negotiated the whole thing through Fenwick & West. We didn't even talk to him on the phone, let alone meet him in person."

This is specifically for tax purposes. They probably want to apply the US-China tax treaty that Chinese citizens enjoy in the US, which should allow them to avoid federal taxes and state taxes. However, if they were smart, they would pay the taxes.

"Is this time different?"

"FlowEDA is a legitimate Delaware C-Corp. This kind of corporate structure is not the simple LLC structure. It has already penetrated the tax system, leaving no easy way to avoid federal taxes. In this respect, this is the first time he has truly run a company properly. If you are interested in mergers and acquisitions, we can contact him directly. However, this also means that he is more serious this time and may not sell it casually like last time."

Ballmer leaned back in his chair, remained silent for a moment, and then said somewhat reluctantly.

"We spent 200 million to 200 million, and now it seems we didn't buy the most crucial asset. Chen Yiran is the most crucial asset. Soon we'll become a laughing stock. Microsoft is a sucker, only knowing how to buy patents and not valuing technical personnel. I can already imagine how Google will publicize this everywhere."

Neither Mullia nor the legal counsel responded. How could they possibly respond? Should they blame Ballmer for being a fool who made the initial decision and got ripped off?

"No, we have to make a move. We need to acquire him as soon as possible, and this time we have to buy him outright, including the people."

Mullia disagreed somewhat, not because he didn't think highly of Chen Yiran, but because he thought too highly of him. He felt there was no hope of buying him now, and that even discussing investing was a blessing. Although Ballmer seemed very determined, Mullia still offered a symbolic rebuttal, just in case there was a possibility of failure later.

"But DeepFlow is open source. It's under the MIT license, so we can use it directly; we can negotiate the licensing price."

"No, he's too young, and his company has a lot of potential. Also, there's that problem: we can't let people say we're technically incompetent. We have to make a statement; we have to propose an acquisition, and we have to let everyone know we're going to make one."

"If he actually agrees to this, we'll be at a disadvantage in terms of price."

Ballmer thought for a few seconds and decisively said, "We won't lose much. He's only 18. Who knows what he'll achieve later? It's settled then. Let's just ask for a price and see. His company, his employees, himself, and this patent—let's package it up and offer $5 million to get started."

This time, we can't just contact him directly. We need to find his phone number and talk to him directly. An 18-year-old techie should be easier to fool than the old foxes at Fenwick & West. We're Microsoft; try offering us that price with our reputation first.

Mullia could only nod helplessly. He was so annoyed. Microsoft was known in Silicon Valley as a company that was rich but stupid, and that reputation was absolutely not helpful for any acquisition.

As soon as Mulia left his office, he started getting his men to work.

Microsoft's corporate development department often suffers losses, spending a lot of money to buy a bunch of obsolete stuff.

However, most of the problems stem from the incompetence of higher-ups; their actual work abilities are still quite strong.

After employing various advanced methods, including asking alumni, making a scene at law firms, and extensive Google searches,

They finally confirmed that Chen Yiran had gone to Stanford for an exchange program, and there must be a phone number or something similar in the Stanford admission record.

At 6 p.m. that evening, Chen Yiran's phone vibrated. It was an unfamiliar number, with the area code 206, calling from Seattle.

For some inexplicable reason, Chen Yiran's first thought was, "This must be a call from Microsoft. Are they coming to deliver money?"

"Hello, this is Chen Yiran."

"Hello Mr. Chen, this is Ivner from Microsoft's mergers and acquisitions department. We completed a transaction through Fenwick & West a few months ago, and now we have a new collaboration that we'd like to discuss with you."

"What's the matter? You've seen Professor Wu's paper, haven't you?"

"Yes, yes, we saw the preprint paper from Professor Andrew Ng's lab. We also tested DeepFlow's performance, and it was very impressive."

"Thanks."

"Mr. Chen, let me be frank. We want to discuss a new business deal. Microsoft wants to acquire all related assets of FlowEDA and DeepFlow, including patents, as well as all equity in your current company, employees, and of course, terms for your continued employment, for a total of $500 million."

Chen Yiran silently repeated three times in his mind: I have no interest in money, I have no interest in money, I have no interest in money.

"Okay, I'll think about it carefully and get back to you."

"OK."

After hanging up the phone, Chen Yiran jumped up and down excitedly a few times, then pulled Pidan over and gave him a good rough time.

Whether you buy it or not, it's still nice to have someone eagerly giving you money.

So, are we planning to sell the company? Do I need to think about a new technology roadmap?

"You see, even if you use my brain, you still can't match my emotional intelligence. Of course you won't buy it, but it's not good to directly refuse, after all, it's Microsoft. Let's give them a couple of days of silence, and then do some big publicity, and our reputation will rise, and Deepflow will get more attention. Isn't this a better way to implement our previous idea of ​​making this framework the standard?"

This isn't emotional intelligence, it's cunning. Humans are truly ruthless.

Just as Chen Yiran was about to retort, her phone rang again. When she picked it up, she saw that it was Huang Shengbin calling.