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 475

The next wave of computing power

2,438 words 8 min read 1 view 23d ago Source ​
Skip to the end of the chapter
Prev
Next

Zhao Wenyuan stared at the Research Institute's eyes, which were filled with a mixture of curiosity and a reluctance to look, and carried the cardboard box back to his office.

When Fang Fang, the secretary, saw him walk in, she was startled and almost thought her boss had been laid off.

Then, after a moment's thought, I realized that the person carrying the cardboard box had come back from outside, not gone out.

In addition, Boss Han didn't come either.

Fang Fang tapped her head, thinking to herself, "Looks like I've been watching too many workplace dramas lately, and the mindless plots have made me brain-dead."

She quickly got up, took the cardboard box, and glanced inside. There was a white helmet, which looked quite exquisite, and a silver mechanical hand, reflecting the luster of metal.

Is this a newly purchased figurine? It looks quite similar to the gauntlet from the Avengers.

It definitely feels like the style that President Zhao likes.

"President Zhao, shall I put it on the bookshelf for you?" Fang Fang asked.

Zhao Wenyuan gave her a strange look.

"Why put it on the bookshelf? I'm connecting it to my computer later." Zhao Wenyuan said, glancing at his watch. "Just leave it here for now. I have to go to my meeting." His next meeting was Computing Power with Jiang Songran, where they would discuss the application of the Douzhi Model.

Upon entering the conference room, Little Niu and the other researchers in charge of the Douzhi Model were already seated. Seeing Zhao Wenyuan enter, they all stood up and greeted him, "Hello, President Zhao."

Zhao Wenyuan nodded slightly to them and said, "Please sit down, Little Niu, turn on your computer and connect to the online meeting."

The video call connected quickly, and the other end was also a conference room.

To everyone's surprise, in addition to Jiang Songran and his directors and senior experts in charge of Computing Infrastructure, Han Luyi was also in the conference room.

"Hello Boss Han," Zhao Wenyuan quickly greeted him.

Han Luyi nodded and said, "I came to see Old Jiang just now, and I happened to run into your meeting. You don't mind if I sit in on it, do you?"

Upon hearing this, everyone except Zhao Wenyuan visibly tensed up.

Cross-departmental collaboration is inherently difficult, and when the big boss suddenly appears in a meeting, all the prepared presentations and exchanges suddenly feel inadequate.

Long-time employees like Little Niu, who have been with the company since the Haicheng era, are fine, but researchers who joined in Beijing, who don't have many opportunities to interact with Han Luyi to begin with, are inevitably more uneasy.

At the start of the meeting, Zhao Wenyuan set the agenda.

"Today's meeting is mainly to set the tone for both teams, to explain what our Research Institute's latest douzhi (fermented Douzhi Model can and cannot do, and what it is suitable for. I already demonstrated its functions to President Jiang this morning, and Little Niu will give everyone a brief review later. Then, we'll mainly discuss how the two sides can cooperate going forward, and how to integrate this Douzhi Model into the scheduling layer. Finally, we'll make a simple schedule." After Zhao Wenyuan finished speaking, he asked Han Luyi, "Boss Han, is there anything else you'd like to add?"

Han Luyi and said, "Let me say this: meetings should be focused on the essentials. If anyone here feels they can't contribute their opinions in this meeting and only wants to keep up with the progress, they can leave now. They can review the meeting minutes later; there's no need to waste time here."

Han Luyi finished speaking, he looked around and saw that no one had left.

Then he continued, "Also, every meeting should have expected outputs before it begins. I just overheard that after this meeting, the cooperation plan and schedule will be finalized, right? Remember to summarize at the end of the meeting."

After finishing speaking, Han Luyi nodded at the screen and said, "Little Niu, let's begin."

Little Niu stood up, cleared his throat, and said, "Boss Han, President Jiang, let me give you a demonstration first."

Little Niu shared its screen and gave a demonstration, which was basically the same as the demonstration Zhao Wenyuan gave at the executive meeting. In the Haicheng conference room, the employees of the Infrastructure Department were seeing the results for the first time. They were initially satisfied, but they couldn't help but have some doubts.

It performs quickly and well with individual requests, but I don't know how it will perform under high concurrency and how much resources it will consume.

For Douzhi to replace Glutinous Rice Ball as Dispatch Assistance, besides ensuring its accuracy, another issue is that it cannot consume too much Computing Power. Otherwise, the saved Computing Power would be used to run the scheduler, which would be completely meaningless.

Jiang Songran the first to raise this question, "How much Computing Power and resources would be needed to support 100,000 QPS for this Douzhi Model?"

Currently, the throughput of all the Computing Power owned by Origin is around 100,000 QPS, which is why Jiang Songran asked this question.

This shows the resources needed to connect the Douzhi to each scheduling request. Little Niu pulled up an icon and said, "President Jiang, we've already done load testing. Our eight L100 cards can achieve about 24,000 QPS."

This number shocked everyone in the Haicheng conference room.

Doesn't that mean that even with resource redundancy, only about forty L100 Computing Power are needed to serve all the inference requests in the entire Origin?

This Douzhi Model is so small and requires so little Computing Power!

Jiang Songran said to Tao Feng, the infrastructure director beside him, "Old Tao, President Zhao just said that the accuracy rate of the Douzhi Model can reach 98%. Based on this data, if we run the Douzhi Model for every step of Computing Power Scheduling now, according to our previous estimates, how much efficiency can we improve?"

Tao Feng quickly tapped a few keys on his laptop and replied, "We calculated before that about 30% of the total Computing Power is wasted due to resource misallocation. If we can achieve that level of accuracy, we should be able to recover at least 20%."

Currently, there are about 30,000 Computing Cards deployed in the two Data Centers. If we can optimize the performance by 20%, it means that we will have added 6,000 cards out of thin air.

Forty cards can be exchanged for the Computing Power of six thousand cards; that's the value of the Douzhi Model.

Han Luyi pondered to himself, realizing that there was still a gap between the training cluster Zhao Wenyuan had requested and the one that required a 10,000-card system; he needed to think of another way.

With the newly purchased cards being continuously deployed, and the adaptation work by the Domestic Computing Power Alliance being carried out by more manufacturers, the shortage of these several thousand cards may be filled in another two months.

It's just that we don't know how much the Origin for Computing Power will increase by then.

While Han Luyi was doing the mental calculations, Tao Feng and the senior experts he brought were already discussing the specific details of the joint debugging with Little Niu. "We definitely need to do a phased rollout. Given the current scale, we'll start by connecting one percent of the Servers, and then gradually increase the number once the results are stable," Tao Feng said.

"Okay," Jiang Songran nodded and said.

Product launches are generally not this complicated, but infrastructure is the cornerstone of all products in the company. If something goes wrong, all products will be affected, so extra caution is needed.

"Then all that's left is time. Let's not talk about going live yet, just making changes, how long will it take?" Jiang Songran turned to ask another subordinate attending the meeting, Tan Yingxu, a senior expert from the Infrastructure Department.

Tan Yingxu pondered for a moment and said, "This interface is very simple. Just replace the call to Glutinous Rice Ball with a call to Douzhi, and roughly change the parsing logic of the returned result. If you're in a hurry, I can finish it today."

Han Luyi chuckled as he listened from the side.

He felt like he was seeing himself back when he was doing scheduling and budgeting at Dingsheng.

"Don't just finish the changes in a day, let's make it a week," Han Luyi said. "Just finishing the changes isn't enough. Remember to get the tests running. This change has a huge impact; there's no need to rush it."

According to Han Luyi's experience, the initial estimate when making a plan is definitely inaccurate. If you double the time estimated at the beginning and add three more days, it will be about the same as the time required to achieve the final result.

The most important thing in infrastructure development is establishing rules. Once the rules are in place, the road ahead will be smooth.

Strange, why are there strange noises coming in?

Tan Yingxu was taken aback. He had worked in many large companies and had seen plenty of orders to "finish it tonight" by working overtime. This was the first time he had ever seen someone voluntarily give him extra time. However, since this was a good thing for him and his team, Tan Yingxu naturally wouldn't object.

"Yes, Boss Han."

Jiang Songran glanced at the results and summarized, "The joint debugging plan has been finalized, and the schedule is also set. Let's aim to achieve 1% grayscale coverage by this time next week. If any department has any new information, please keep everyone informed."

Just as the video conference was about to end, Zhao Wenyuan suddenly remembered something and said, "By the way, Boss Han, there was also some new progress from Mengnao Technology this morning. When should I report to you?"

What he really wanted to show off was the box of treasures he had just brought back to the office.

Han Luyi replied, "Old Jiang and I need to discuss other matters further. I'll contact you later."

Han Luyi and Jiang Songran sat on the sofas on either side of the coffee table in the Chairman's Office. Han Luyi asked Liu Huo to bring over two cups of coffee.

"How's L200 progressing?" Han Luyi asked as he took the coffee.

Jiang Songran now the Senior Vice President in charge of Computing Power, overseeing all aspects of L200 development, from Chip R&D to Data Center management and infrastructure construction. Han Luyi first asked Jiang Songran about all the progress of L200.

Chip development is one of the most confidential fields in every company. Jiang Songran had no materials on hand and relied entirely on his mental knowledge of the progress to explain.

"Things are progressing fairly smoothly on Old Xue's side. The data from our previously simulated new architecture on the simulation platform is amazing. With the help of the Icecore team, we've made adjustments to a few physical limitations. As for the Interconnect, both sides are currently developing their own systems, but we might need to bring Shen Zhonghe and his team to Yiyue Haicheng next." Han Luyi nodded and said, "When I went to Beijing to see him this time, he was also insisting on bringing his team. They're really looking forward to collaborating with Icecore."

Jiang Songran smiled and said, "It feels good to cooperate with Icecore, but I guess they feel good about cooperating with us too."

Han Luyi understood what Jiang Songran.

After receiving the latest tools shared by Origin, the Icecore team's work efficiency has improved dramatically. Even the engineers who came to exchange ideas were overjoyed, praising the new tools to everyone they Origin.

"So the next time point is--" Han Luyi said.

"It's September. The design drafts and technical integration with several teams are still underway. We'll definitely finalize the design by the end of August and be able to tap out the chip by the end of the year."

Seeing Jiang Songran's determined look, Han Luyi didn't say anything more, only instructing him, "If any unexpected situations arise, or if there are problems you can't solve, you'll have to tackle them yourself, but don't forget to let me know."

"Huh?" Jiang Songran taken aback.

Then he suddenly remembered that Han Luyi probably had a collaborative team from a national Research Institute behind him.

That's the one who wrote the Operators back then.

Could it be that Boss Han's connections have been maintained—no, expanded? Even including experts in the Chip industry?

"Okay, Boss Han, we'll let you know as soon as possible if we encounter any difficulties," Jiang Songran said. After discussing the progress of L200, Han Luyi suddenly sighed, "You really can't rush things." Jiang Songran felt the same way.

Since he joined the company, it seems that the company's Computing Power has never been surplus.

The time when we were at our richest was probably when we bought the first batch of 500 L100 coupons—back when Glutinous Rice Balls) hadn't even been released to the public yet.

Jiang Songran couldn't help but chime in, "These things can't be rushed, Boss Han. Don't be impatient. Take it one step at a time, and things will come."

Han Luyi looked at Jiang Songran and suddenly asked, "Wenyuan is planning to allocate 10,000 L100 cards to train Glutinous Rice Ball 2.0. Do you have any ideas?"

Jiang Songran taken aback again.

Huh? Is it my turn to come up with a solution?

We currently only have a little over 30,000 cards, and only 18,000 L100 cards. This still represents the highest paper Computing Power and the highest utilization rate.

Zhao Wenyuan really dared to ask for 10,000 L100 sheets for training!

Jiang Songran a moment, then said, "How about we rent Cloud Computing Power instead? It's a bit expensive though—"

The domestic Computing Power is already strained. Every major manufacturer with its own GPUs is busy developing its own Large Models, and very few are willing to rent them out.

Furthermore, the price of graphics cards has been soaring, naturally leading to higher rental costs. They had previously calculated that if they couldn't secure a discounted price and simply rented Computing Power from the market, given that Glutinous Rice Ball's pricing wasn't particularly high, they might not even be able to cover the Computing Power costs and would actually lose money.

Unless we can get cheap Computing Power.

This isn't a fatal problem for many large overseas manufacturers, but it presents significant challenges for China manufacturers.

Jiang Songran pondering this when he suddenly heard Han Luyi say, "Expensive—"

"With the addition of Douzhi) and the new Scheduling Architecture, does it still seem expensive?"

Upon hearing this, Jiang Songran startled!

Yes, the utilization improvement brought about by the new scheduling system is comprehensive, not just limited to the Free Data Center in Origin.

If Cloud Computing Power is used, efficiency is also improved.

That already fully covers the costs!

End of Chapter
Prev
NovelSummaries
Next

Explore the wiki

Han Luyi
Zhao Wenyuan

Comments

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