Rebirth: Starting from Lighting Up the Tech Tree

Chapter 119 Solution Formulation

At 8:30 a.m., Shen Yiming pushed open the door to office 402 and found Zuo Cheng already sitting at his workstation.

There were two cups of coffee on the table, one of which was clearly reserved for him.

"What time did you go to sleep last night?" Shen Yiming asked, moving his coffee cup closer to himself as he sat down.

"Around two o'clock," Zuo Cheng said. "Come on, take a look at this."

He turned the screen around.

Shen Yiming leaned closer, took one look, and froze.

97.1%.

He thought he was seeing things, so he leaned forward again, but the numbers hadn't changed. Accuracy 97.1%, power consumption 4.2 watts, latency 11 milliseconds. He recognized these three metrics; they were the acceptance requirements for Sky Dome Phase IV, and all of them had been exceeded.

"This is impossible." His voice was a little hoarse. "How did you do it?"

"I tested a new algorithm idea last night, and it just so happened to run successfully," Zuo Cheng said casually, as if he were talking about taking a detour due to traffic congestion this morning.

Shen Yiming stared at him for two seconds without asking any further questions. People in the tech field often have this intuition: some breakthroughs are hard to explain, but the results are real, and that's enough.

"So what do we do now?" Shen Yiming asked.

"You lead the verification, I'll organize the solution documentation," Zuo Cheng said. "There are still two days left, enough time."

Shen Yiming glanced at the time, took a sip of his coffee, then turned on his computer, carefully reviewed the test log, and ran the test himself. The result was exactly the same as what was displayed on the screen; there was no element of luck involved.

When Fang Ze arrived, it was already past nine o'clock. He saw that both of them were at their workstations. Shen Yiming was still looking at the test logs, and Zuo Cheng was writing documents. The atmosphere was neither as heavy as yesterday nor as noisy as it usually is after a successful breakthrough. It was strangely quiet.

"What happened?" Fang Ze asked.

"Come and take a look," Shen Yiming said.

Fang Ze walked over, glanced at the screen, then looked at it again. His expression didn't change much, except for a slight raise of his eyebrows: "97.1%?"

"Yes," Zuo Cheng said. "Power consumption 4.2 watts, latency 11 milliseconds. Verify it again from the chip's perspective."

Without wasting words, Fang Ze sat down, put aside his work, and began to check.

Over the next two days, 402 became a high-speed machine. Shen Yiming was responsible for the complete process verification, running six rounds of stress tests to simulate the scenario of 480 satellites working concurrently, with each round's results consistently above 97%. Fang Ze checked the power consumption distribution at the chip level and found two areas for further optimization, reducing the actual power consumption to 4.1 watts. Zuo Cheng was organizing the solution documentation, writing the entire technical architecture in simple and easy-to-understand language to ensure that even committee members who didn't understand the technology could comprehend it during the review meeting.

Han Lu was responsible for liaising with Lanwan Communications and arranging the submission time and demonstration environment for the proposal.

After the first round of testing, Shen Yiming compared the results with his expectations: "Based on my previous calculations, the best-case scenario was 93% to 95%, but I didn't expect it to jump directly to over 97%. From which direction did you approach the compiler's algorithm?"

"Compilation graph optimization," Zuo Cheng said. "It involves reorganizing the model's computational paths, removing redundant nodes, and retaining branches with high information density."

Shen Yiming thought for a moment, nodded, and didn't press further. He had considered this direction, but manually optimizing the computational graph required a very deep understanding of the target hardware, which couldn't be done in a short time. The fact that Zuo Cheng could pull it off overnight meant that his experience in this area was deeper than he had imagined.

Scholars can sometimes convince others, and Shen Yiming is like that. He won't deny the results just because he doesn't understand them. The data is real, and the conclusion is valid, which is enough.

Fang Ze's verification report came out earlier. When he handed it to Zuo Cheng, he said, "Cambricon's MLU270 achieved about 78% of its performance in this solution. There's still room for improvement, but it's sufficient for now."

Zuo Cheng accepted the report and made a note of it. This figure meant that as subsequent chip versions were iterated, the solution still had room for further optimization; it wasn't a one-off deal.

The night before the proposal was to be submitted, Han Lu ordered takeout, and everyone gathered in the conference room to eat. Compared to the somber atmosphere of a few days ago, the atmosphere was much more relaxed now. Plastic bags were scattered on the table, and the paper sleeves of disposable chopsticks were crumpled up and thrown into the trash can. Chen Hao, holding a can of cola, leaned back in his chair.

"Tomorrow is the final battle." Han Lu held up a can of beverage. "Wish us luck."

"Good luck?" Chen Hao smiled. "I think victory is inevitable."

Shen Yiming placed his chopsticks on the lid of the box and nodded: "An accuracy rate of 97.1%—there shouldn't be another company in the world that can produce this data."

Fang Ze picked up a piece of meat without saying a word, but there was a slight curve at the corner of his mouth, which was equivalent to a big laugh on someone like him.

Zuo Cheng didn't speak; he was looking out the window.

The nights in Hangzhou are always like this, with the lights of the high-rise buildings layered one on top of another, like countless worlds in motion stacked together. He thought of what Zhou Henian had said, of Lin Jianhua's expression at the last review meeting, and of the Tianqiong Phase IV project—480 satellites autonomously coordinating, a feat unprecedented globally.

Then he realized that 402 now had the answer.

"Brother Cheng," Han Lu suddenly said, "what are you thinking about?"

Zuo Cheng looked away, glanced at the takeout on the table, and picked up his chopsticks: "Thinking about the presentation order for tomorrow."

"Everything is arranged," Han Lu said. "Huaxin will present first, then we'll go second. This is advantageous for us; we can strike later and gain the upper hand."

"Okay," Zuo Cheng said, "then let's wait until tomorrow."

There was a subtle silence in the air, not one of unease, but more of a solemn atmosphere before battle. Everyone knew what was going on, but no one spoke of it aloud.

Then Han Lu's phone vibrated. She looked down at it and her expression changed slightly.

"What's wrong?" Chen Hao asked.

Han Lu turned her phone screen over to show everyone a message from the contact person at Lanwan Communications: "For tomorrow's review, Huaxin has added a representative: Lin Jianhua."

The meeting room fell silent for a moment.

Shen Yiming frowned slightly, Fang Ze put down his chopsticks, and Chen Hao remained silent.

After reading the message, Zuo Cheng closed the lid of his lunchbox, stood up, and threw the empty Coke can into the trash can.

"Come on then," he said. "The outcome won't change."

He didn't say anything more and sat back down to continue eating.

Lin Jianhua's arrival was unplanned, but not unexpected either. Last time during the review, he failed to stop 402 despite using his veto power. This time, he must be here to find an angle to challenge them on the spot.

Zuo Cheng was all too familiar with this tactic.

In his previous life, he stumbled because of such details. A technical detail that was not properly prepared, or a data loophole that was seized upon by the opponent, was enough to turn the tide of the battle.

This time it won't happen again. 97.1% accuracy, 4.1 watts power consumption, six rounds of stress testing—behind every number lies solid verification. Lin Jianhua tried to find a loophole, but couldn't.