Rebirth: Starting from Lighting Up the Tech Tree

Chapter 112_Technical Breakdown

After leaving Lanwan Communications, Zuo Cheng did not go straight back to the company, but instead asked Shen Yiming and Fang Ze to go back first.

He walked alone on the street; the autumn wind was a bit chilly.

5 watts.

This number kept replaying in his mind. The satellite's power consumption budget was only 5 watts, while they were running a complete AI scheduling system.

At nine o'clock in the evening, Zuo Cheng pushed open the door to office 402. Shen Yiming and Fang Ze were both there, with a thick stack of technical documents in front of them.

"Brother Cheng, you're here." Shen Yiming looked up, a hint of weariness in his eyes. "Fang Ze and I have already started breaking down the technical requirements."

Zuo Cheng nodded and sat down opposite them.

Shen Yiming pushed over a hand-drawn architecture diagram: "The core requirement of the Sky Dome Phase IV project is autonomous collaborative networking among satellites. Each satellite must have edge AI capabilities, enabling it to dynamically adjust its behavior based on the status of surrounding satellites."

"The key challenge isn't the algorithm, it's the computing power," Fang Ze continued. "The satellite's power consumption budget is only 5 watts, while the minimum power consumption of a traditional AI model is over 30 watts."

30 watts.

Exceeding the standard by six times.

Zuo Cheng was silent for a moment: "Is it possible to use multi-card collaboration?"

Shen Yiming was taken aback: "Multi-card collaboration?"

"We break down a complex AI task into multiple simpler sub-tasks and distribute them across different satellites," Zuo Cheng said. "Each satellite is responsible for only a small part, and communication and coordination are carried out through inter-satellite links."

Shen Yiming's eyes lit up. He strode to the whiteboard and began drawing a new architecture diagram.

"If we use multi-GPU collaboration, the computing power requirements of a single satellite can be significantly reduced," he said while drawing. "For example, what used to require a 30-watt GPU might now only require several 5-watt low-power chips."

Fang Ze frowned: "But this would place very high demands on the bandwidth and latency of inter-satellite communication."

"That's right." Shen Yiming nodded, "Therefore, we need an efficient communication protocol to ensure synchronization and coordination between distributed tasks."

Zuo Cheng looked at the new structure on the whiteboard and already had an idea in mind.

"Yiming, do you think this direction is feasible?"

Shen Yiming pondered for a moment: "Theoretically feasible, and..." He paused, "the technology we accumulated during our studies in the Federation can be used here."

Federal learning.

That's a technology they just validated in a smart city project. It optimizes the model through distributed training without exposing the raw data.

The same approach can now be applied to satellite collaboration.

"However, there's a problem," Fang Ze began, "Even if we distribute the computing power, each satellite still needs to run a simplified version of the AI ​​model. What compression ratio can we achieve for this model?"

Shen Yiming pulled up the test data: "Current compression technology can reduce the model size to one-thirtieth of its original size, but the accuracy will drop to around 70%."

"70% isn't enough," Zuo Cheng shook his head. "Satellite control requires high reliability; the accuracy must be at least 90%."

The meeting room was silent for a few seconds.

Shen Yiming rubbed his temples: "I need to think of another way."

Zuo Cheng stood up and walked to the window. Outside the window was the night view of Hangzhou, brightly lit.

He recalled the scene in Zhou Henian's office just now. The difficulty of Phase Four of the Sky Dome was indeed more than an order of magnitude higher than that of Phase Three.

But he has a system.

He has advantages that others don't.

The key is how to leverage this advantage without exposing the system's existence.

"Brother Cheng," Shen Yiming's voice interrupted his thoughts, "I have an idea."

Zuo Cheng turned around.

"Federated learning isn't just for training," Shen Yiming said. "If we expand it and turn it into a framework for 'collaborative reasoning,' the results might be even better."

Collaborative reasoning?

"Each satellite not only runs its own model, but also shares its reasoning results with other satellites," Shen Yiming explained. "After receiving the results, the other satellites will combine them with their own observations to make a final decision. This is equivalent to bringing together the 'wisdom' of multiple satellites."

Fang Ze's eyes lit up: "Isn't this distributed computing?"

"It goes a step further than distributed computing," Shen Yiming said. "Distributed computing distributes tasks, while collaborative reasoning aggregates wisdom. The effects are completely different."

Zuo Cheng understood.

This is an upgraded version of federated learning. It includes not only distributed training but also distributed inference.

"How much can the accuracy be improved if we use collaborative reasoning?" he asked.

沈一鸣在白板上快速计算着："保守估计，可以提升15到20个百分点。70%加20%，就是90%。刚好达标。"

90%.

Just meets the standard.

Zuo Cheng remained silent. He knew that in a real review meeting, a 90% accuracy rate was far from sufficient. Those traditional experts would seize upon every detail and question it.

"It must be over 95%," he said.

Shen Yiming was stunned for a moment: "95%? This..."

"Lin Jianhua is at the review meeting," Zuo Cheng said calmly. "He won't give us any room for error."

The meeting room fell silent once again.

Fang Ze broke the silence: "Actually, there's another direction we can try."

Zuo Cheng and Shen Yiming both looked at him at the same time.

"Chips," Fang Ze said. "If we can improve the energy efficiency of chips, we won't need to make so many compromises on algorithms."

"You mean custom-made chips?" Shen Yiming asked.

"It's not customization, it's optimization," Fang Ze said. "Cambricon's MLU270 has sufficient performance, but its power consumption is 30 watts. If we can do deeper optimization at the software level, the power consumption should be able to be reduced further."

Zuo Cheng nodded. Fang Ze was right. Only by addressing both hardware and software issues could the problem be truly solved.

"Starting tomorrow, we'll split into two groups," he decided. "One group will be led by Yiming to continue optimizing the collaborative reasoning framework; the other group will be led by Fang Ze to see if we can make a breakthrough at the chip level."

"Okay," they both replied in unison.

The meeting ended late at night. Zuo Cheng told Shen Yiming and Fang Ze to go back first, while he stayed in his office.

He opened the system panel to check the current status.

[Technology Tree Status]

- Branches: Communications Engineering, Aerospace Communications, Internet of Things, AI (4/7)

Points: 277

- Leaves: 39 (including 4 fused leaves)

- Tech Radar: Cooling down (12 hours remaining)

277 points.

Enough for one technological integration, or several scans and copies.

But the technology radar is still cooling down, and it cannot directly scan new technologies.

We still need to wait.

Zuo Cheng closed the system panel and leaned back in his chair.

5 watts power consumption, 90% accuracy, two weeks.

Three rigid conditions loomed before us like three towering mountains.

But he didn't panic.

Because he knew that as long as they could get through this period, 402's AI capabilities would experience a qualitative leap.

It's not just for Sky Dome Phase IV, but for the even more distant future.

He stood up, turned off the light, and left the office.

The corridor was quiet, with only the echoing sound of footsteps.

Tomorrow will be another battle.