Rebirth: Starting from Lighting Up the Tech Tree

Chapter 114_Technological Dilemma

For the next three days, Team 402 entered a state of high-intensity work.

Shen Yiming and Fang Ze practically lived in the lab. They slept no more than four hours a day, and their desks were piled high with instant noodle boxes and coffee cups.

On the morning of the fourth day, Zuo Cheng pushed open the laboratory door.

Shen Yiming was asleep in front of the computer, while Fang Ze leaned back in his chair, resting with his eyes closed. A string of glaring numbers appeared on the screen.

Zuo Cheng walked over and glanced at the data.

Accuracy: 68.3%.

It's two percentage points lower than it was three days ago.

He turned to look at the whiteboard by the window, which was covered with optimization strategies. They had tried every method, but the results were unsatisfactory.

"Brother Cheng."

Shen Yiming woke up, rubbing his eyes as he sat up.

"Are the latest results out?"

Shen Yiming nodded, his eyes filled with fatigue and frustration: "We tried twelve optimization strategies. Eight of them lowered the accuracy, three remained the same, and only one improved it by 0.7 percentage points."

"68.3% plus 0.7% equals 69%," Zuo Cheng said.

"Yes." Shen Yiming smiled wryly, "But the goal is 95%."

The meeting room was silent for a few seconds.

Fang Ze also woke up: "Actually, it's not just Shen Yiming's problem. The chip itself has also encountered a bottleneck."

He pulled a stack of test reports from the drawer: "The energy efficiency optimization of the MLU270 has been pushed to its limits. We've tried dynamic precision quantization, dual-clock domain design, and power gating."

How low can the power consumption be reduced?

"4.9 watts," Fang Ze said. "That's only 0.1 watts less than the target, with almost no margin."

0.1 watts of reserve.

In the real Sky Dome project, this 0.1 watt could become a deadly hazard.

Zuo Cheng stood by the window, looking at the sky outside. The autumn sun was warm, but his heart felt a little cold.

"The accuracy rate must be above 95%," he said. "At the review meeting, Lin Jianhua and his gang won't give us any room for error."

Shen Yiming remained silent.

He knew he couldn't do it.

It's not a matter of ability, but because it has reached the theoretical limit.

"Brother Cheng," Shen Yiming began, "I think...we might need to change our approach."

What's the thought process?

Shen Yiming hesitated for a moment: "Perhaps we shouldn't keep obsessing over compression algorithms."

"What should we do then?"

Shen Yiming glanced at the numbers on the computer screen, then at Zuo Cheng.

"The accuracy rate improved from 68% to 95%, a difference of 27 percentage points," he said. "This would have been virtually impossible to achieve using traditional methods."

"But..." he paused, "if there's a new algorithmic breakthrough, perhaps there's still hope."

Zuo Cheng turned around and looked at him.

"What kind of breakthrough?"

Shen Yiming walked to the whiteboard and picked up a pen.

"Automatic hyperparameter optimization." He wrote these five words. "The performance of an AI model largely depends on the configuration of its hyperparameters. Manual hyperparameter tuning is too inefficient. If AI can automatically search for the optimal hyperparameters, the accuracy will be significantly improved."

"How much can it improve?"

"It's hard to say," Shen Yiming said honestly. "Theoretically, it could improve by 5 to 10 percentage points. But the prerequisite is that we need to have the necessary tools."

5 to 10 percentage points.

68% plus 10% equals 78%.

not enough.

"It must be over 95%," Zuo Cheng said.

Shen Yiming fell silent.

He didn't know how to answer.

The gap between 68% and 95% is not something that can be bridged by hard work alone.

This is two orders of magnitude more difficult.

"Actually, there's another problem," Fang Ze suddenly said.

The two looked at him.

"Even if automatic hyperparameter optimization improves accuracy by 10 percentage points, the accuracy is only around 78%," Fang Ze said. "That's still 17 percentage points away from 95%."

"Where did this 17 percentage point come from?"

Fang Ze shook his head: "I don't know."

Silence fell over the meeting room once again.

Zuo Cheng looked at the sunlight outside the window, but his mind was racing.

How long will it take for the technological radar to cool down?

He opened the system panel and glanced at the countdown.

4 hours remaining.

Four hours later, he will be able to scan Shen Yiming and obtain the technology for automatic hyperparameter optimization.

But this is only the first step.

With automatic hyperparameter optimization, the accuracy may be improved to around 78%.

It's still 17 percentage points short.

How did we arrive at these 17 percentage points?

He also integrates technology.

Perhaps, the capabilities of multiple blades can be combined to create a stronger new capability.

However, the success rate of fusion is not 100%. If it fails, the 5 points will be wasted, and you will have to wait another 48 hours for the cooldown to complete.

Two-week period.

Every day wasted makes time even more precious.

"Brother Cheng."

Shen Yiming's voice interrupted his thoughts.

"What's wrong?"

Shen Yiming walked up to him, his eyes filled with apology: "I'm sorry, Brother Cheng. I... I can't do it."

Zuo Cheng was stunned for a moment.

"This isn't your problem," he said. "You've done very well."

"But we didn't reach the target." Shen Yiming lowered his head. "We only achieved 68%, while the target was 95%. The gap is too large."

Zuo Cheng looked at him without saying a word.

He remembered himself in his previous life.

He was like that back then too.

When faced with problems that cannot be solved, one may fall into self-doubt and self-criticism.

But it's different now.

He has a system.

He has advantages that others don't.

"Yiming," Zuo Cheng said.

Shen Yiming raised his head.

"As I said, this is two orders of magnitude more difficult," Zuo Cheng said calmly. "It's not a matter of your ability; it's the theoretical limit of the industry."

"but……"

"But we have another way," Zuo Cheng interrupted him. "Leave the accuracy issue to me. You and Fang Ze continue optimizing the chip design."

"You?" Shen Yiming asked, somewhat confused. "Brother Cheng, what are you planning to do?"

Zuo Cheng did not answer.

He walked to the door and glanced back at Shen Yiming and Fang Ze.

Give me four hours.

Then he pushed open the door and left the laboratory.

Shen Yiming and Fang Ze exchanged a glance, both seeing the confusion in each other's eyes.

"What is he planning to do?" Fang Ze asked.

Shen Yiming shook his head: "I don't know. But I believe him."

He believed that Zuo Cheng would definitely be able to come up with a solution.

Just like every time before.

Every time Zuo Cheng encountered a seemingly unsolvable problem, he always managed to find a breakthrough. This time, Shen Yiming believed it would be no exception.

Zuo Cheng returned to his office and gently closed the door.

He opened the system panel and looked at the line of text:

[Technology Radar is cooling down]

[Remaining time: Approximately 3 hours and 58 minutes]

Less than four hours.

He needs this technology.

Automatic hyperparameter optimization.

This is the key to breaking the deadlock.

With this technology, everything will be different.