rpbooks Chapter 109 Alliance
Chapter 109 Alliance
The Cambrian people arrived ten minutes earlier than expected.
When Han Lu led the three guests into the conference room, Zuo Cheng was talking quietly with Yu Ying. Yu Ying was wearing a light blue shirt today, her hair was tied in a ponytail, and she was sitting next to Zuo Cheng with a notebook open in front of her.
"Mr. Zuo, this is Zhang Zhiyuan, the Business Director of Cambricon, Sun Mingchen, the Technical Architect, and Zhou Ting, the Product Manager." After introducing them, Han Lu left the conference room and closed the door.
Zhang Zhiyuan, in his early thirties, wore a well-fitting dark blue suit, and his handshake was firm but controlled. Sun Mingchen, a typical tech guy, wore a plaid shirt and glasses, and was surveying the equipment in office 402 before even entering. Zhou Ting, quite young, carried a thin laptop and wore a professional smile.
"Mr. Zuo, it's an honor to meet you." After sitting down, Zhao Mingyuan glanced around the conference room. "402's rapid development is evident to everyone in the industry. From satellite communication to the Internet of Things platform, we've achieved the connection of over 100,000 devices in just over two years. This is a rare feat among startups in China."
"Mr. Zhang, you flatter me," Zuo Cheng said. "Cambricon is the one that truly commands industry attention. Its MLU series chips are a benchmark in China's AI chip field."
Zhang Zhiyuan smiled and got to the point: "We're here today to discuss a potential collaboration with 402. Cambricon is currently designing the architecture for its next-generation edge AI chip and needs real-world edge scenario testing. 402's IoT platform runs over 100,000 devices, which is exactly the testing environment we need."
"How exactly will we cooperate?" Zuo Cheng asked.
"Cambricon provided 50 MLU270 engineering samples for 402 to test free of charge. Simultaneously, our technical team provided remote support to help 402 port its AI models onto our chips. In return, 402 provided test data from edge AI scenarios to help Cambricon optimize its chip architecture."
Zuo Cheng nodded. This cooperation model was very advantageous to him: free samples plus technical support, almost zero cost. But he needed to confirm a few key issues.
"I have a question," Yu Ying suddenly spoke up, "The MLU270's inference framework is Cambricon's self-developed BANG architecture, which is not fully compatible with mainstream frameworks like TensorFlow and PyTorch. If the 402 model needs to be migrated from mainstream frameworks to the BANG architecture, how high would the migration cost be?"
Sun Mingchen adjusted his glasses and glanced at Yu Ying. He clearly hadn't expected anyone in room 402 to ask such a professional question. He put down the documents in his hand, leaned forward slightly, and spoke with a more serious tone.
"Migration does indeed have costs," said Sun Mingchen. "We provide the CNToolkit toolchain, which can convert models from mainstream frameworks to the BANG architecture. However, for custom operators and complex model structures, manual adaptation is required. In our experience, the migration cycle for common visual models is about three days, while natural language processing models may take up to a week."
"What about the workload of adapting custom operators?" Yu Ying pressed. "Part of the gradient aggregation logic in our federated learning framework is implemented using custom operators. Can this part run on the BANG architecture?"
"It depends on the complexity of the operator. A simple convolution operator might take half a day to a day, while a complex attention mechanism might take two to three days."
Yu Ying jotted down a few words in her notebook, then turned to Zuo Cheng and whispered, "The migration costs are manageable, but we need to allow at least two weeks for adaptation."
Zuo Cheng nodded slightly, then said to Zhang Zhiyuan, "We have no objection to the cooperation framework. However, there are two conditions. First, the intellectual property rights to the test data belong to 402, and Cambricon only has the right to use it. Second, if optimization directions for the chip architecture are discovered during the testing, 402 has the priority right to participate in the joint design of the next-generation chip."
Zhang Zhiyuan and Sun Mingchen exchanged a glance. The first condition was normal, but the second condition was somewhat unexpected. Priority participation in joint design meant that 402 was not content with being a chip verification provider; they also wanted to be a chip designer.
"The second condition needs to be reported back," Zhang Zhiyuan said. "But personally, I think it's a win-win situation. 402 has deep experience in edge AI scenarios, and if they can participate in chip architecture design, it would be good for Cambricon. To be honest, we are also looking for a transformation path from simply selling chips to providing solutions, and cooperation with 402 may be a good entry point."
Zuo Cheng's heart skipped a beat. Cambricon also wanted to provide a solution? This meant they didn't just need 402 for testing, but hoped 402 would become their ecosystem partner. This bargaining chip was bigger than he had imagined.
"Then we'll wait for Mr. Zhang's good news," Zuo Cheng said.
Zhang Zhiyuan stood up and extended his hand, saying, "It's a pleasure working with you."
After the meeting, Zuo Cheng escorted Yu Ying downstairs.
"You did a good job today," Zuo Cheng said. "You stumped Zhang Zhiyuan and the other technical architects."
Yu Ying rolled her eyes at him: "I didn't want to come in the first place, but you insisted on dragging me in as a technical consultant."
"I mainly miss you."
"You said it twice." Yu Ying couldn't help but laugh and gently pushed him.
Zuo Cheng took her hand and stood at the entrance of the office building for a while. The autumn wind was already a bit chilly, but Yu Ying's hand was warm.
"Kongkong, the collaboration with Cambricon is just the beginning." Zuo Cheng looked into the distance. "402 will eventually have to make its own AI chips. Now, by accumulating scenarios and experience, when the time is right, we will be the best partners for chip companies, or even become chip companies ourselves."
Yu Ying tilted her head and looked at him: "Your ambitions seem to have grown even bigger."
"It can't be too small." Zuo Cheng smiled. "Go back, and be careful on the road."
After Yu Ying's car drove away, Zuo Cheng returned to his office. He closed the door and opened the system panel.
Two hundred and ninety points remaining. The technology radar is still cooling down. The nine leaves of the AI branch are quietly shimmering, with the leaf for model compression optimization seemingly shining brighter than the others, as if hinting at something.
Zuo Cheng stared at the leaf for a long time. Model compression optimization, Shen Yiming's federated learning framework, and Cambricon's hardware support—these three lines converged, and 402's edge AI capabilities were taking shape.
But that's not enough. He needs more talent, more scenarios, and more technological accumulation.
Zuo Cheng turned off the control panel, turned on his computer, and began writing a three-month plan for the AI division. Phase One: Talent and Chips. Phase Two: Algorithms and Applications. Phase Three: Products and Commercialization.
Take it one step at a time.