Shen Congyun put down his phone and focused his attention on the computer screen.
When you open your browser, you see a person's LinkedIn profile.
Songran Jiang's resume doesn't even have a photo, and it's very basic.
From 2011 to 2020, I worked at Google as a Staff Engineer (L6) in the TPU team, which corresponds to the Dingsheng of P8, a senior expert.
In 2020, I switched jobs to a Chip startup as Head of Infrastructure, where I remain to this day. There's only one position available, and the specific job duties are completely undefined.
After closing the browser, Shen Congyun opened a folder on the desktop. This folder contained information about Jiang Songran that Shen Congyun had collected from other sources.
Jiang Songran was in the TPU team at Google, where he was involved in both software and Hardware, led several large projects, and rose through the ranks at an astonishing speed.
After switching jobs, he became the first employee of that startup, and he built a training cluster with a thousand-card-level graphics card from scratch. There are no more than a dozen people in the world who have done this kind of work, and he is one of them.
In short, he is an extremely low-profile industry leader. He has achieved remarkable success in the field of AI infrastructure, but is not well-known outside of that field.
The departure and return of such a person to China has indeed stirred the hearts of bosses and headhunters throughout the AI industry; they don't need Shen Congyun's departure. Not to mention Dingsheng, as far as he knows, people from Pengcheng have even flown over specifically to persuade Jiang Songran join them.
Han Luyi certainly wouldn't want to miss out on such a talent, but Shen Congyun had already exhausted all his options.
You can't get through on the phone, can't add him on WeChat, and he doesn't reply to emails. What can you do? You can't exactly go to his house and wait for him, can you? The problem is, you'd need to know where he lives, wouldn't you?
Shen Congyun packed up the materials and sent them to Han Luyi Lark.
Han Luyi quickly replied, "Come to the thirteenth floor."
Shen Congyun quickly gave the message a thumbs up, put away his laptop, and left the office.
Upon reaching the thirteenth floor, Han Luyi and Zhao Wenyuan were sitting side by side in front of a computer, with the information that Shen Congyun had just sent on the screen.
Han Luyi first waved for Shen Congyun to sit in the empty chair in front of him, then asked bluntly, "What's wrong? Not going well?"
"Boss Han, he's not answering the phone, not replying to emails, we can't even get in touch with him." Shen Congyun's voice was low; he felt it was his own fault for not handling things properly.
Han Luyi turned to Zhao Wenyuan and asked, "Wenyuan, do you know him?"
Zhao Wenyuan shook his head. Jiang Songran had already left when he joined Google. The two also worked in different fields; one focused on theoretical research, and the other on engineering practice. Although they were both in Silicon Valley and might have mutual friends, they didn't have any direct interaction.
Suddenly he remembered something, took out WeChat and searched for Jiang Songran's name.
Zhao Wenyuan has several WeChat groups for Chinese people working at Google, including groups for eating, drinking, having fun, hiking, and cherry picking. He joined these groups gradually during his years at Google, and Jiang Songran also be in them.
The search actually found it; it was a group called "[Haicheng People at Google]", and no one spoke in it after Zhao Wenyuan joined.
Zhao Wenyuan clicked on Jiang Songran's profile and then opened his Moments, which only showed the last six months. There was only one post, a photo of a hospital bed, not the person, with the caption. "You never know which will come first, tomorrow or an accident."
Han Luyi watched the whole thing from the side without saying a word.
Shen Congyun spoke up, "Boss Han, I've also found Jiang Songran's WeChat, but he hasn't accepted my friend request."
Zhao Wenyuan asked, "Should I add him from the group? He might accept."
Han Luyi didn't say yes or no, but took Zhao Wenyuan's phone and looked at the post on WeChat Moments again.
It was just past 10 a. m.
Neurosurgery ward, a general ward of a top-tier hospital in Haicheng.
It's very quiet here.
Jiang Songran sat on a folding chair next to the hospital bed, with a laptop on his lap.
He had short hair, a thin face, and was expressionless as he stared silently at the screen, his eyes bloodshot.
Lying on the bed next to him was an elderly man who bore a five or six-point resemblance to Jiang Songran, only thinner and more aged. An IV line was inserted into the old man's left hand, and his breathing was very light. The monitoring device by the bedside beeped regularly at intervals, as if reminding the people in the ward of the passage of time.
The day my father suffered a stroke, he was working overtime in his office in Silicon Valley.
He suddenly received a call from his mother overseas, but he couldn't recall what he was thinking or what he did at the time.
It was as if I had entered an autopilot mode.
Jiang Songran bought the earliest available plane ticket back to China, carrying only a small suitcase with a few clothes and a personal computer.
He didn't know at the time if he would be able to see his father one last time.
After more than ten hours of travel, the moment he saw his father outside the glass door of the ICU, tears welled up in his eyes.
Thankfully, he was saved.
After spending more than ten days in the ICU and being out of danger, my father was transferred to a regular ward, but he still spent more time sleeping than waking up.
The doctor advised that I rest first, and if the recovery goes well, I can be discharged after another month of observation.
After talking with the doctor, Jiang Songran made a decision without consulting anyone.
He's not going back to the United States.
After making the decision, he made a video call with the company's CEO to explain his thoughts.
The CEO remained silent for a long time on the other end of the video call, and in the end, he didn't even say anything to persuade him to stay. He simply wished his father a speedy recovery and said that he was welcome to come back anytime he wanted in the future.
Jiang Songran noticed that the old white man who had started the business with him in the garage and worked with him for several years had tears in his eyes.
Then I asked a friend in the United States to help manage the house and keep an eye out for real estate agents.
From the moment he received his mother's call, his life seemed to be divided in two.
The first half of his life is one of frantic efforts to prove himself and make his parents proud.
He didn't know what the second half was.
Jiang Songran was waiting for his mother to come and relieve him.
Jiang Songran idly scrolling through the news when he came across one article—
The Domestic Computing Power is the biggest variable in the next five years.
Upon seeing this title Jiang Songran smiled inwardly, a smile that hadn't been seen in a long time.
These media outlets probably don't even know what they're talking about.
"Domestic Computing Power is a national strategic resource"—of course, everyone would say that.
But for someone like him who specializes in this field, it's a joke.
There's no other reason than that the entry barrier is too high.
This is like saying we should solve the climate problem—a correct but meaningless statement. What exactly needs to be done? How? How much will it cost? How long will it take? What will the consequences be? Nobody can say for sure.
After closing the news, Jiang Songran opened a new page on his computer: arXiv.
This is a website for publishing papers for free, especially in the AI and computer systems field, where almost all new research is first published. There's no need to wait for journal review or conference acceptance; once you've written it, you can publish it and the whole world can read it that same day.
People in the industry call this "hanging it up".
Checking arXiv every morning to see if there's anything new under the categories you follow is a basic habit for most researchers in this field.
Jiang Songran has maintained this habit for more than a decade, from his school days to Google and then to his startup.
He has been back in China for more than two months accompanying his family to the hospital. He has put aside all his work, but he still maintains the habit of browsing arXiv every day.
His gaze swept quickly across the list, and the title of a newly published paper caught his attention.
<On Adapting Operator Adaptation -Efficient Attention to Non-CUDA Hardware Accelerators: A Single -Operators Case Study>
The author has only one name: Wenyuan Zhao.
Jiang Songran went through the name in his mind, but couldn't recall it.
He originally intended to swipe away directly.
There are tons of non-CUDA adaptations on arXiv, most of them are just projects paid for by domestic Chip manufacturers to have their own engineers do, with no real substance.
But he clicked on it anyway.
Because it is a single-author work.
For non-CUDA adaptation work, there are usually at least four or five people listed as contributors: Hardware manufacturers, software teams, algorithm engineers, and those doing evaluations.
A single author implies two possibilities: either there is no team, or the author has suppressed everyone else and only credited themselves.
Both possibilities are quite interesting and worth reading this paper about.
The PDF loads.
The first sentence of the abstract is:
This paper reports an end-to-end implementation of a memory-efficient attention mechanism on a domestically developed non-CUDA accelerator: with a single-node 8-card configuration, the throughput reaches 83% of the corresponding cuDNN reference implementation, with a numerical error within 2.3e-6.
Jiang Songran increased the font size on the screen and then scrolled down.
The background text is concise and not padded.
The implementation section includes specific optimization methods and code snippets, rather than just general discussions.
The evaluation section includes a complete comparison, throughput curves for different context lengths, and a separate quantification of the contribution of each optimization. These are some of the areas where evaluators are most likely to cut corners, but he didn't.
The section on limitations was the one that Jiang Songran read twice more than the other.
The authors themselves wrote, "This work only completes the single Operators (SDPA) and does not involve dozens of other Operators such as LayerNorm, Softmax, and various Optimizers; it does not involve multi-GPU communication libraries; nor does it involve backpropagation and mixed precision in training scenarios, and does not constitute a usable training stack."
Very honest, doesn't seem like someone who's just trying to gain attention.
He flipped up to look at the team: Source Intelligence AI.
It's a company I've never heard of.
Jiang Songran leaned back in the folding chair and rubbed his eyes.
Surprisingly, there are actually people in China doing this kind of thing.
This is a thankless and arduous task.
Even if you create something, you won't get published in top conferences, you won't gain recognition, and you won't be able to raise funds. While you're working on it, nobody knows what you're doing or how difficult it is.
Most importantly, since there are no other supporting components, making one on its own is really not very useful.
He's good at this kind of work; that's how the Google TPU team got through its first two years.
However, Google had several hundred engineers back then.
Eight cards, one Operators.
Jiang Songran mentally assessed what was still missing between here and "being able to run a usable training system".
The remaining dozens or hundreds of Operators, multi-GPU communication libraries, distributed training frameworks, compilers, fault tolerance, and checkpoints.
Each project requires an engineering team of 30 to 50 people and takes two to three years to complete.
He worked on this system at that company in Silicon Valley for five years.
If there were hundreds of teams in China, all like this one, quietly doing this kind of thing that no one applauds.
So, domestic production might not be out of the question.
Unfortunately, there aren't hundreds of such teams.
Just then, my mother came in from outside and started nagging as soon as she entered.
"Songran, you must be exhausted. Go back and get some sleep. You need to rest while you're at the hospital with your dad. Don't get sick before he even wakes up..."
Jiang Songran was indeed exhausted.
He said a few words to his mother, tidied up a bit, picked up his laptop, and left the ward.
As he walked to the first floor of the inpatient department, he saw a tall, thin, and handsome young man standing at the entrance, carrying a backpack, seemingly looking for someone.
Jiang Songran was watching when the young man also looked over. He seemed to have noticed him, smiled and nodded at him, and walked towards him.
The man's eyes were incredibly bright.
Jiang Songran wasn't sure if he was coming for him, but he nodded in response.