Skip to main content
  • Home
  • Library
  • Read
    • Popular
    • Charts
    • Browse
    • Categories
    • Wikis
    • Glossary
  • Request
  • Activity
  • Leaderboard
  • Help
Our Apps
  • DaoSearch
  • Narratorr
Social
  • Discord
  • Patreon
The AI Wave Tycoon's Rise
Chapter 273

One Talent Could Secure the Nation

2,556 words 9 min read 1 view 1mo ago Source ​
Skip to the end of the chapter
Prev
Next

"Impossible," Xue Zhaoheng blurted out.

He designed the L100 and knows all its specifications inside and out.

He had never seen or used the closed-source model of Glutinous Rice Ball.

However, to achieve the top spot on the leaderboard, it must at least be no smaller than those open-source models.

Rumors suggest that OpenAI's latest model, GPT-V, has trillions of parameters, while Glutinous Rice Ball Energy would have at least hundreds of billions.

As for L100, there's a small cluster of L100 running in the lab next door, and it has only been able to run 7B smoothly at most.

An open-source model with seven billion parameters is so large that even a 21-billion-parameter model would experience severe lag.

This is still two orders of magnitude less than the number of parameters needed to run a closed-source model at the level of Glutinous Rice Ball.

Now you're saying Glutinous Rice Ball is running on an L100 cluster?

How was the communication efficiency issue resolved? His software team spent over a year working on the L100 Operators, and only managed to cover less than five percent of the usage scenarios. How were the remaining parts filled in?

In a flash, Xue Zhaoheng thought of many things, then his face darkened, and he slowly said, "Boss Han, I don't know what you mean by that."

Only Han Luyi and Xue Zhaoheng were in the office. He looked up at the door and continued, "If you're trying to use this line of reasoning to deceive people, saying that Ximing has broken through Nvidia's ecosystem blockade, and then using your Source Intelligence name to package Ximing, then I'm sorry, I can't cooperate with you."

Han Luyi knew that Xue Zhaohuan was wary of him, but he still hadn't expected him to make this guess.

Isn't it possible that the workload for Operator Adaptation is quite large, but your Ximing team couldn't write it, while my team with the Source Code has already finished it?

This possibility clearly never crossed Xue Zhaoheng's mind from beginning to end.

He wouldn't believe anyone.

Both Lu Zhengping and Zhou Han are laymen. After hearing about the domestic adaptation, they only thought that Han Luyi had made a technological breakthrough. However, China has made many technological breakthroughs in recent years.

However, Xue Zhaoheng Supreme Purity of the amount of engineering work required to write these Operator Adaptation and communication protocols.

If you say it took four or five years to polish it and finally you got results, that's more believable.

However, the L100 has only been in mass production for a year, and those 500 cards were delivered to Source Intelligence only two months ago.

Two months is hardly enough time to write anything.

Han Luyi originally planned to use the fact that "Source Intelligence has completed the inference adaptation of L100, and Glutinous Rice Ball is running on the L100 Cluster" as a bargaining chip to win over Xue Zhaoheng and the Ximing technology team.

However, he did not expect that Xue Zhaoheng would not believe him at all.

Having no other choice, Han Luyi took out his phone and, in front of Xue Zhaoheng, opened the backend panel of the Glutinous Rice Ball Inference Cluster in the Cloudrim Data Center, showing him the backend directly.

He placed his phone in the middle of the table, turned it so that it was facing Xue Zhaoheng.

"President Xue, this is the real-time data, please take a look."

Xue Zhaoheng leaned forward, glanced at the backstage panel, and then reached out and swiped it.

The above contains various data related to cluster load, including real-time call volume, average latency, GPU memory utilization, throughput, error rate, and task distribution.

Xue Zhaoheng stared intently at several key indicators, comparing them with the data in his mind as he looked at them.

He quickly came to a conclusion.

Fake.

The L100 is not that efficient.

"In that case, Boss Han, there's no need for us to continue this discussion."

This time, Xue Zhaoheng's tone was no longer polite as before; in fact, he was suppressing his anger.

He had already decided that Han Luyi an ambitious man who wanted to use Ximing to swindle and deceive, and he was already plotting how to sabotage the deal—although, to be honest, as a minority shareholder, there wasn't much he could do.

But he would rather Ximing go bankrupt than watch Ximing become a tool for others to swindle money.

Proving you haven't fabricated anything is very difficult, because once someone suspects you of fabrication, any evidence will appear to them as forged.

Han Luyi had experienced this many times before, but he didn't expect Xue Zhaoheng to be so stubborn, refusing to believe it even with the panel right in front of him.

He disregarded the offensiveness in Xue Zhaoheng's words and proposed one last method.

"President Xue, does Ximing have a ready-made testing environment?"

Xue Zhaoheng was visibly taken aback upon hearing this.

Aren't you going to give up yet?

"Have."

What size?

"There's a small cluster in the Data Center next door, with thirty-two servers; that's for our internal verification."

Han Luyi nodded and said, "Thirty-two is enough, President Xue. Would you mind taking me to the Data Center?"

Xue Zhaoheng knew that Han Luyi was still going to stick to his previous explanation.

He had no reason to refuse, and at this point, a slight wavering had begun to creep into his heart.

Is the L100 really that amazing?

The two stood up and left the office together.

Ximing's Data Center is right next door, but strictly speaking, it can't be considered Data Center in the standard sense.

It was just a room in an office building that was converted.

There are no hot or cold air channels, no humidity control, and no electrostatic shielding.

Several Server Cabinet stood in the middle of the room, their fans humming.

There was no one in the Data Center.

At this time, the team had all left work, but Xue Zhaoheng had waited here specifically to see Han Luyi.

However, he has full access to the Data Center and Server cluster.

--

"Boss Han, how do you plan to verify this?" Xue Zhaoheng asked.

"As you yourself said, L100 Cluster needs to solve at least two major problems to run Large Model with hundreds of billions of nodes."

"Yes, communication efficiency, and Operators." Xue Zhaoheng nodded.

Although there are only two major categories, the technical challenges involved are numerous.

Take the Operators library as an example. Nvidia's ecosystem is the result of years of meticulous work and refinement. It's not just about having a complete Operators; they're also constantly optimizing it. When performing the same operation on a model, CUDA on a Nvidia graphics card is fast; but on a different ecosystem, it's not as fast.

For L100 to achieve the same result, engineers from multiple teams working together on software Hardware architecture would need to find the most efficient parameters.

It can be said that the foundation of the CUDA ecosystem lies here.

"We've solved those two problems." Han Luyi said this casually, but those who knew him well knew he was doing it on purpose.

Xue Zhaoheng was not familiar with Han Luyi's style, but he also felt that he had been ridiculed.

Do you know what you're saying?

Han Luyi ignored Xue Zhaoheng's reaction and took out a USB flash drive from his pocket.

He had just asked Liu Huo for this USB drive outside. It was empty, but while they were talking, he had already written the adaptation files from the Source Intelligence private code library into it using the Sight.

Xue Zhaoheng thought that Han Luyi had prepared this in advance.

"This is the runtime and Operator Adaptation rewritten by the Source Intelligence team. You can try using it before deploying an open-source model."

Upon hearing this, Xue Zhaoheng's face showed a skeptical expression. "You mean, Source Intelligence completely rewrote it, ignoring the software layer provided by Ximing? And it can still achieve a level of performance close to CUDA?"

Han Luyi didn't say anything, but handed the USB drive forward again.

Xue Zhaoheng took it and plugged it into a computer in the Data Center that was connected to the Server cluster.

The Server in the entire Data Center are all on an internal local area network, so he wasn't too worried about the USB drive containing a virus.

But he really didn't believe that Source Intelligence was capable of doing what Han Luyi had just said.

Wouldn't that really turn China into Nvidia?

The USB drive contains several compiled files with well-defined file names, which are just the right size to replace the Operators and runtime components in the current L100 software stack.

Without even needing Han Luyi's guidance, Xue Zhaoheng knew where to copy the file and then restarted the Server.

During the reboot process, these newly stored binary files on the hard drive are loaded into memory and used by the Server cluster.

The console terminal begins scrolling the log.

L100 backend detected.

runtime initialized.

kernelfusion enabled. (Operators fusion is enabled)

memoryplanner initialized.

Multi-card scheduler ready.

Service loading complete.

Model 21B has begun running.

Xue Zhaoheng's breathing became slightly rapid.

He didn't manually type in the question to see the model's response.

What the model responds to is a question that the person who builds the model should consider.

He only needs to consider the performance and utilization efficiency of the Hardware.

The Server contains scripts that Ximing usually uses for testing, which can generate a large number of model inputs and perform stress tests on the model's load-bearing capacity.

Xue Zhaoheng did not hesitate and started running the test script.

The values on the monitor began to change.

The first thing to jump is the video memory utilization rate.

Previously, when running open-source small models, Ximing's own Operators library couldn't even allocate GPU memory properly, leaving 70% of the GPU memory unused.

This number now far exceeds 30%.

75% 8% 88% --

It eventually stabilized at around 95%.

The load testing script was increasing concurrency before detecting any backfeed pressure.

The current real-time throughput has reached 250,000 Token per second.

Xue Zhaoheng knew that the last time he ran the same test, he was stuck at 20,000 points and couldn't go any higher.

On Computing Power Chip, a key metric that truly measures software performance is Model FLOP Utilization).

FLOPUtilization (MFU). However, this metric is not directly displayed on the panel because calculating it requires using the Chip's maximum Computing Power.

If the Chip is theoretically capable of running at 1000 TFLOPS, but your model only actually uses 200 TFLOPS, then the MFU (Mean Functional Fusion) is 20%.

Running the 21B model requires 221 bytes to generate each Token, which is approximately 42 GFLOPs.

250,000 Token per second corresponds to 2,500,0042 GFLOPs = 10,500 TFL0PS.

The theoretical Computing Power of L100 is 600 TFLOPS, so a 32-card cluster would have 32600 = 19200 TFLOPS.

10500/19200 = 54.7%

54.7%?

After doing the mental calculations, Xue Zhaoheng always felt that he had miscalculated the number of digits.

He opened the calculator on the computer and did the calculation again.

That really is 54.7%.

It's worth noting that even Nvidia's most advanced H200 graphics card only achieves a maximum MFU (Mean Functionality Rate) of around 60%.

At this moment, Xue Zhaoheng's heart was beating as fast as it had been on his wedding night more than twenty years ago.

He always believed that L100's approach was flawed, and that Ximing's current predicament was due to a failure in Chip design.

But at that moment, he realized that all he was missing was a software program.

With this software, Ximing has truly become the Nvidia of China.

Xue Zhaoheng slowly stood up and turned to look at Han Luyi, only to see Han Luyi looking at him with a faint smile on his face.

"President Xue, I didn't lie to you, did I?" Han Luyi said.

Xue Zhaoheng could still hear his own heart pounding in his chest.

This was the closest he had ever come to his dream in all these years.

The Domestic Computing Power has truly been achieved under his leadership.

After taking a few deep breaths, Xue Zhaoheng finally spoke, "Boss Han, I'm sorry."

Before Han Luyi could speak, Xue Zhaoheng continued, "I apologize for my arrogance just now. This software is more valuable than L100, more valuable than Ximing. I actually thought you were a fraud."

Han Luyi stepped forward, patted Xue Zhaoheng on the shoulder, and said with a smile, "President Xue, you flatter me. Your rigorous approach is exactly what Chip R&D needs most. I'm counting on you to lead the team to make even better Chip in the future."

"Besides, software and Hardware are both indispensable, and it's hard to say which is more valuable. Without L100, the software layer is just a castle in the air." Han Luyi offered Xue Zhaoheng a way out.

However, Xue Zhaoheng knew in his heart that this software adaptation layer was not developed through the joint efforts of the two companies, but rather by Source Intelligence's team through reverse engineering. If they could make Ximing's software, they could naturally make software for other companies' graphics cards as well.

The real scarcity isn't in graphics cards, but in the ability to quickly adapt software layers.

Ximing was lucky to be chosen.

With just a few hundred megabytes of the adapter layer, transferring data from a USB drive takes only a few seconds. But with these few hundred megabytes of files, Ximing's valuation of two billion dollars would probably increase more than a hundredfold.

It's important to know that the whole of China is looking for a domestic alternative to Nvidia's system.

The size of China's AI Chip market reached 300 billion yuan three years ago.

Even if we only capture 10% of that, it would still be 30 billion in revenue, enough to support a company with a market value of several hundred billion.

Although Xue Zhaoheng is not a successful businessman, he has worked in the Chip industry for many years and is very familiar with these figures.

However, he did not have any ulterior motives regarding the documents provided by Source Intelligence. On the contrary, he was filled with gratitude for Han Luyi's trust. After all, he could have secretly kept a backup and then told the investors about it.

However, Xue Zhaoheng knew that the technical capabilities of the Source Intelligence team were far more valuable than these few hundred megabytes of files.

Only through collaboration between Ximing and Source Intelligence can we continue to explore the limits of domestically produced Computing Power.

Of course, what Xue Zhaoheng didn't know was that the reason Han Luyi entrusted the USB drive to him so confidently was because all the electronic equipment in the Data Center was under the surveillance of Sight. Even if he really wanted to secretly back up a backup file, he would be discovered and deleted immediately.

Fortunately, he didn't have a chance to discover this magical defense mechanism.

After Xue Zhaoheng deleted the adaptation files and restarted the cluster, he and Han Luyi walked out of the Data Center one after the other.

It was already 8 PM.

"Boss Han, it's getting late, and you've had a long journey today, so I won't keep you for dinner." Xue Zhaoheng said with an embarrassed chuckle. "You should get some rest. After we check the inventory tomorrow, I'll treat you to Pengcheng specialties."

Speaking of which, Xue Zhaoheng, the CEO of this startup, lives a much more frugal life than Han Luyi. He doesn't have a secretary or assistant, and he doesn't even have a driver.

His eagerness to see Han Luyi was already impolite, and the fact that it was so late, coupled with the fact that he had no one with him to arrange a banquet, left him with no choice but to awkwardly see his future boss off.

Fortunately, Han Luyi didn't care about these things.

In his view, work comes first, and if he can bring Xue Zhaoheng into the team and have him wholeheartedly dedicate himself to research and development in the future, then these few hours will not have been wasted.

With the successful transaction with Ximing, Glutinous Rice Ball's reasoning Computing Power is at least solved in the short term.

Xue Zhaoheng escorted Han Luyi the car. Before the car door closed, Han Luyi glanced at Xue Zhaoheng's attribute panel one last time.

[Xue Zhaoheng]

[R&D (Chip) SS | Management D Stability SS]

[Special Skills: Chip Development | Supply Chain Management]

[Entry: From Nothing to Something - Phoenix Chick]

"Boss Han, shall we head back to the hotel?" Liu Huo asked, turning around from the passenger seat.

Han Luyi turned to Lu Mingzhou and said, "Mingzhou, you've been here for two days. Have you eaten at any good restaurants? Let's go have dinner first."

After saying that, Han Luyi looked out the window.

The Crouching Dragon and the Young Phoenix—possessing either one is enough to bring peace to the world.

Now that I have both of them in my hands, Nvidia, how will you respond?

End of Chapter
Prev
NovelSummaries
Next

Explore the wiki

Han Luyi

Comments

No comments yet.
AboutStatsAPIChangelogStatusContactPrivacyTermsDMCA
© 2026 RandomTranslator·v3.26·RSS