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 201

Five hundred cards? Just the starting point

1,755 words 6 min read 2 views 1mo ago Source ​
Skip to the end of the chapter
Prev
Next

In the afternoon, Han Luyi finally met Jiang Songran was on his first day at the company.

The two found an empty conference room on the thirteenth floor.

After a few brief pleasantries, Jiang Songran couldn't help but ask about work, "Boss Han, this morning President Zhao showed me the progress, and to be honest, I was shocked. It's even more amazing than the results you showed me at the hospital."

Han Luyi knew that Jiang Songran had already seen what kind of efficiency Source Intelligence Technology had achieved by using Agent.

This is the key to catching up with Nvidia's CUDA ecosystem in a short period of time and turning the impossible into possible.

Providing a few Operators adaptation results is impressive, but it can't change the industry.

Having a system that can stably and quickly adapt Operator Adaptation would be a different story.

Jiang Songran eye, could see what this meant.

However, Han Luyi was also curious about what Jiang Songran, as an AI Infra expert, could bring to Source Intelligence Technology. He had already thought about the general direction and wanted to achieve two things.

First, he was uneasy about having his own training and inference stack data running on someone else's cloud.

Secondly, using domestic alternatives is a win-win situation. Firstly, due to various reasons, he simply cannot buy Nvidia cards; secondly, he has the ability and motivation to take the lead in completing the domestic substitution of Computing Power. If he can do it, not only will his own Data Center problem be solved, but he will also be able to open up a broader domestic government and enterprise market.

However, once the direction is set, the actual implementation still requires talent with practical experience in this industry.

This is why Jiang Songran so important to Source Intelligence Technology.

Jiang Songran an industry leader with profound technical expertise, but his greatest value to Source Intelligence Technology lies in another aspect: Jiang Songran was the number one employee and senior executive in his previous startup company. He has experience in building Data Center from scratch, and his experience spans both software and Hardware.

Such talent is far rarer than a technical genius.

It can be said that his importance to Source Intelligence far exceeds his importance to Dingsheng and Kunlun.

Those major cloud service providers already have extensive experience in Data Center management. At most, Jiang Songran's trip there will bring some new perspectives and improvements in certain aspects, but it won't lead to any fundamental breakthroughs.

However, Jiang Songran came to Source Intelligence Technology, he directly filled the gap in Source Intelligence Technology's Hardware management.

That's why Han Luyi was so generous, directly giving out a 2% Stock Options reward.

This is a legendary gold medal certified by Brother Sight!

Jiang Songran may have made a profit, but Han Luyi definitely didn't lose out.

Since Jiang Songran brought it up, Han Luyi asked directly, "Dr. Jiang, what we need to do now is to have our own reasoning stack and training stack as soon as possible. My initial plan is to purchase 500 cards. This purchase quantity is not large among domestic cards, and any company can afford it. But I would like to hear your opinion on where to buy them and to what extent they can be used."

Han Luyi had already revealed this number to Jiang Songran while they were in the hospital. Jiang Songran was prepared; he knew Han Luyi would definitely ask about it after he joined the company, and he wouldn't fight an unprepared battle.

"Boss Han, I've checked. The most advanced domestically produced cards on the market right now can achieve 400 TFLOPS. If it's just for inference, 500 cards are more than enough to support 100,000 Daily Active Users," Jiang Songran said. "Post-training based on Annotated Data, while barely feasible, is still possible. But for pre-training, it's far from sufficient."

TFLOPS (TeraFLOPS, Trillion Floating-point Operations Per Second) is a unit of measurement for computing Computing Power, which means "one trillion floating-point operations per second". Because GPUs excel at performing floating-point mathematical operations, if each GPU is likened to a sports car, TFLOPS is its top speed on the dashboard.

Using NVIDIA graphics cards as an example, the RTX 5090, a dream card for gamers, can achieve around 200 TFLOPS in practical AI training scenarios, while the Nvidia B200 series graphics cards, specifically designed for AI training, can reach 2250 TFLOPS per card.

The terrifying Computing Power of TFLOPS.

Han Luyi nodded; this was exactly what he had expected.

"Furthermore, Boss Han, I don't know if anyone in our team has practical experience in managing GPU clusters. However, the difficulty of managing a non-CUDA ecosystem GPU cluster is completely different from that of the CUDA ecosystem. Out of 500 domestically produced GPUs, no more than three companies nationwide can get the training stack running."

After saying that, Jiang Songran revealed a slightly proud expression. "With me here, I'm confident that Source Intelligence Technology will be one of them."

Han Luyi had no doubt about this, after all, he recruited Jiang Songran precisely to do this job.

He then asked, "Which company is the best to buy from? Do you have any suggestions?"

Jiang Songran hesitated for a moment, then did not answer immediately.

This is actually a somewhat sensitive issue. Jiang Songran, having been a manager, knows that procurement is a sensitive topic for any company, especially since a single graphics card used in AI costs over 100,000 yuan. Five hundred cards would amount to a purchase of 50 million yuan, which is no small sum.

For safety reasons, he would be better off recommending a leading domestic manufacturer. Large manufacturers have strict internal controls, so the risk of him receiving kickbacks is small, and Han Luyi would be less likely to suspect him.

However, after conducting his investigation, Jiang Songran did find a better option, but revealing it might not be a good thing for him personally.

He didn't hesitate for long, gritted his teeth, and gave a more risky version. "There's a company called Sili-Lumin Semiconductor..."

Sili-Lumin, headquartered in Pengcheng, is also a startup. Last year, they began mass production of their first product, the L100, with Computing Power of 600 TFLOPS and Inter-Die Interconnect bandwidth of 600 GB/s, comparable to Nvidia's H100. The Hardware is robust, but there's a drawback that's both an advantage and a disadvantage for us: their software is poorly developed, completely lacking any training deployment support; buyers are only using it for inference.

"So Ximing is particularly suitable for us. An order worth over 50 million is a lifeline for them, and since their sales channels are poor, they have stock on hand." Jiang Songran explained his thoughts. "In addition, their software is weak, but their Hardware is strong, making their cost-effectiveness exceptionally high based on the Hardware unit price. And we don't even need their supporting software, which means all our procurement money is spent on Hardware."

After saying this, Jiang Songran felt uneasy.

He explicitly stated that this was a manufacturer he had investigated beforehand. Would Han Luyi suspect that he had taken a kickback?

He silently thought to himself: He treated me like a national hero, and I should repay him as a national hero.

Han Luyi outwardly calm, but his eyes gleamed.

Without hesitation, he said directly, "Dr. Jiang, you've gone to great lengths. It's indeed very suitable for Source Intelligence's current needs. Since it's inconvenient for you to travel, I'll have my assistant come to find you in a bit. Please explain to him that he should make a trip to Pengcheng."

Jiang Songran breathed a sigh of relief.

He was quite relieved that he wasn't involved, at least it would avoid any appearance of impropriety.

After finishing their business, Han Luyi added, "Since you say you're confident, I look forward to the day we build our own cluster."

In fact, Jiang Songran was being conservative; he wasn't just confident. He had done this once before, so doing it again was a sure thing.

Having witnessed the technological potential of Source Intelligence Technology, he felt that clusters of 500 or 1,000 kilowatts were not out of the question, and even 10,000 kilowatts were not impossible to try.

It's important to understand that scaling up a graphics card cluster involves an exponential increase in technical difficulty.

The real challenge lies in the communication between the nodes in the cluster during use. To train Large Model, the data between the nodes must be completely synchronized. If there are only ten people, it is not too difficult for them to exchange information; however, if the number is expanded to one hundred or one thousand people, each doing their own thing, even simple information synchronization becomes very difficult.

Moreover, a 10,000-card cluster does not offer a tenfold increase in computing speed compared to Thousand-Card Cluster; due to limitations in communication bandwidth, it is only about seven or eight times faster.

The costs are extremely high, and the benefits are minimal. Given this, why bother with a multi-card cluster? Why not just create many smaller clusters?

That's because a 10,000-card cluster can do what Thousand-Card Cluster cannot: train Large Model.

Based on current Large Model theory, the larger the model and the more parameters it has, the stronger its capabilities. However, training so much information into a single model requires memory capable of simultaneously accommodating all the original information. Only a cluster of the scale of 10,000 kilobytes (kcal) is capable of training such a Large Model.

It can be said that the NVIDIA cluster is the crown jewel of the Nvidia ecosystem.

Now, Jiang Songran has ambitions to use domestically produced graphics cards to challenge the multi-card cluster at Source Intelligence Technology.

Of course, even though domestically produced graphics cards are much cheaper, the cost of a single card for AI training is still over 100,000 RMB. To build a cluster of 10,000 graphics cards, the cost of the graphics cards alone would be 1 billion RMB. Not to mention that for a cluster of this size, the cost of supporting facilities such as data Data Center, power, cooling, and storage is about double that of the graphics cards themselves. The total investment for a cluster of 10,000 graphics cards would be over 2 billion RMB.

But Jiang Songran wasn't worried about any of these things. Wasn't that something the CEO should be concerned about? He believed Han Luyi would handle it.

These thoughts swirled in Jiang Songran's mind, but he didn't say anything aloud.

But his heart, which had become dormant due to the changes in his family, suddenly burned with passion again.

End of Chapter
Prev
NovelSummaries
Next

Explore the wiki

Han Luyi

Comments

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