bigabreads Chapter 15 Leaving the Laboratory
Chapter 15 Leaving the Laboratory
In late October, Zuo Cheng stepped into Lanwan Communications' R&D center for the first time.
The location was in the High-tech Zone of Huaxia City, a 40-minute subway ride from the school. The R&D center occupied a 26-story office building, and access to the core area required swiping three cards. Zuo Cheng picked up a temporary employee badge at the front desk, and Han Zhe's assistant led him through a long corridor with laboratories separated by transparent glass on both sides, filled with testing equipment he had only seen in textbooks.
Han Zhe was waiting for him in the conference room on the seventeenth floor.
"Sit down." Han Zhe gestured to the chair opposite him, in front of him lay Zuo Cheng's research proposal, several parts marked with red pen. "Let me explain the situation in Part Two to you first."
He opened a network topology diagram of Blue Bay Communications' base stations and pointed to a node in the southern district of Huaxia City.
"This is a 5G experimental base station we deployed in the southern area, with a coverage radius of two kilometers and approximately three thousand daily users. Your algorithm needs to run on the signal processing unit of this base station, processing real user signals in real time, running continuously for seventy-two hours without failure, and achieving the performance levels promised in the solution."
Zuo Cheng looked at the topology diagram and quickly calculated the workload in his mind.
There is a huge gap between laboratory simulation and real base stations. The signals in the simulation environment are clean, controllable, and repeatable; real base stations face three thousand living users, each with different signal characteristics, unpredictable movement trajectories, and ever-changing environmental interference.
The hardware is even more challenging. In the lab, algorithms run on high-performance servers with virtually unlimited computing power. However, base stations use custom-designed embedded platforms with limited computing power, memory, and power consumption. The algorithm must achieve the same performance under these constraints—it's like putting an athlete who excels on a wide track into a narrow alley full of obstacles and still maintaining the same speed.
"Could you give me a copy of the technical specifications for the hardware platform?" Zuo Cheng asked.
"It's all ready." Han Zhe pushed a USB drive over. "It contains the complete hardware manual, interface documentation, and development environment configuration guide for the base station signal processing unit. I've also granted you remote access; you can connect to the base station's development and testing environment via VPN from school, so you don't have to come all the way here every time."
"Thank you, Mr. Han."
"There's one more thing," Han Zhe said, leaning back in his chair, his tone becoming more casual. "President Zhou asked me to tell you that he'll be at the R&D center one day next week, and if it's convenient for you, he'd like to speak with you in person."
Zhou Henian. CTO of Blue Bay Communications.
"It's convenient for me," Zuo Cheng said.
After leaving Lanwan Communications, Zuo Cheng opened the hardware manual on the USB drive on the subway.
The more he looked, the more his brows furrowed.
The base station signal processing unit uses a custom-designed embedded processor with a low clock speed, but it integrates a dedicated signal processing acceleration module. The problem is that the instruction set of this acceleration module is completely different from that of the general-purpose processor. Zuo Cheng's algorithm involves a large number of matrix operations, which can be directly called from the standard math library on a general-purpose processor, but on this custom chip, the standard library cannot run and must be reimplemented using a dedicated instruction set.
This is Fang Ze's battlefield.
That evening, Zuo Cheng sent the hardware manual and his preliminary analysis to Fang Ze.
Fang Ze spent two hours looking at it before replying with a very long message.
"I've roughly understood the processor architecture. The good news: this chip's acceleration module has dedicated hardware support for fixed-point matrix operations. If you convert your floating-point algorithm to a fixed-point implementation, the speed could theoretically increase by three to five times. The bad news: converting floating-point to fixed-point introduces quantization errors. Your algorithm requires high precision, and poor quantization error control will cause performance to plummet. This is a trade-off between precision and speed, and we'll need to work on it together."
Zuo Cheng stared at the three words "Let's gnaw together" with a slight smile.
Fang Ze never says "I'll handle it" or "You solve it." He says "Let's tackle it together." This person is a collaborator at heart, not a lone wolf—provided he recognizes your abilities.
For the next three days, Zuo Cheng and Fang Ze met in the lab every night to focus on the precision control problem of converting floating-point to fixed-point.
It's an uphill battle.
The core challenge lies in the recursive update step within Zuo Cheng's adaptive tracking algorithm—each iteration makes a minor correction based on the previous result. Floating-point numbers have sufficient precision, allowing for accurate execution of these minor corrections. However, fixed-point numbers have limited precision, introducing a small quantization error with each iteration. Over hundreds of iterations, this error can accumulate to an unacceptable level.
Zuo Cheng tried three conventional error compensation methods, all of which failed—either the error was not reduced, or the speed dropped even though it was reduced, resulting in more harm than good.
On the fourth day, he sat in the lab staring blankly at a screen full of data.
Fang Ze walked over and placed a cup of iced Americano next to him. This was the first time Zuo Cheng had ever seen Fang Ze buy coffee for someone else.
"Let's try a different approach," Fang Ze said, leaning against the table with his arms crossed. "You've been thinking about how to reduce quantization error, but have you ever considered—making the algorithm itself insensitive to quantization error?"
Zuo Cheng raised his head.
"The reason your recursive update step amplifies the error is because there's a division operation in the update formula—fixed-point division has the greatest precision loss." Fang Ze picked up a pen and drew a simple diagram on the paper. "If we change the update formula from multiplication and division to addition and subtraction, and use logarithmic field transformation to convert multiplication and division into addition and subtraction, the precision loss of fixed-point addition and subtraction is two orders of magnitude smaller than that of multiplication and division."
Zuo Cheng stared at the simple diagram on the paper, his mind reeling as if struck by lightning.
Logarithmic field transformation.
Transform multiplication and division into addition and subtraction.
He hadn't considered this approach at all—because logarithmic field transformations are not commonly used in signal processing; they are techniques used in digital communication coding. Zuo Cheng's knowledge system grew from the tree of signal processing, naturally creating blind spots. Fang Ze, on the other hand, did embedded low-level development, dealing with fixed-point arithmetic every day, and was adept at these kinds of techniques.
This is the meaning of a team—one person's blind spot is precisely another person's strength.
"Fang Ze, you are a genius," Zuo Cheng said. This wasn't just polite talk; it came from the bottom of his heart.
Fang Ze took a sip of his coffee expressionlessly: "Not a genius, just someone who's been tortured by fixed-point calculations too many times."
That evening, Zuo Cheng rewrote the recursive update module, replacing all multiplication and division operations with logarithmic field transformations. Fang Ze simultaneously ported the new code to the simulator on the custom chip for testing.
The test results came out at 2 a.m.
The quantization error was reduced to one-fiftieth of its original value. The algorithm runs four times faster on custom chips than the floating-point version. Performance metrics are almost identical to those from lab simulations.
Fang Ze looked at the data on the screen and, for the first time ever, uttered a swear word.
It's the good kind.
Zuo Cheng let out a long sigh, leaned back in his chair, and stared at the ceiling, laughing for a long while.
A line of text silently popped up on the screen in my consciousness:
[Host detected to have resolved a key technical bottleneck; Main quest chain, stage two, progress updated: 35%]
Only 35%. Algorithm porting is only the first step; there are still system integration, joint debugging and testing, and 72 hours of continuous operation verification to follow.
But the most difficult hurdle has been overcome.
Zuo Cheng closed the control panel and glanced at his phone. At 2:15 AM, there was an unread message.
It was sent by Han Zhe at 11 PM.
"President Zhou has confirmed he'll be at the R&D center next Wednesday at 3 PM. Just come directly to the 17th floor. Also, he asked me to let you know in advance—he wants to discuss more than just the research project."
It's not just about research topics.
Zuo Cheng put down his phone and looked out the window at the quiet campus.
What could be the topic that Blue Bay Communications' CTO wants to discuss that's "more than just research topics"?
He couldn't think of an answer yet, but his intuition told him that the conversation next Wednesday might be more important than winning the entire project.