Reborn in 2010, the AI ​​in my mind is too powerful.

Hsiang-Chung Kong, a Chinese-American, is a member of the National Academy of Engineering. In 1978, he proposed the theory of pulsating arrays and wrote a remarkable paper.

Simply put, whether it's a CPU or a GPU, data needs to be moved between memory and computing units when performing calculations.

Each time a number is calculated, a row of A and a column of B must be read from memory, calculated, and then written back to memory. Data transfer takes time and consumes a lot of power.

The approach of a systolic array is to allow data to flow through the array like a pulse, with each computing unit communicating only with its neighboring units, without needing to repeatedly access memory.

After careful consideration, and with a little inspiration from Jasmine, Chen Yiran decided to use this method to create a special chip.

Fully connected layers and convolutional layers both involve matrix operations. A 256×256 systolic array consumes only 75W of power, which is far lower than that of GPUs and CPUs.

Search and video websites that use the Deepflow framework require a lot of AI inference.

Search engines offer query understanding, result sorting, personalized recommendations, and even voice search. Video recommendations and understanding utilize various new technologies, such as automatic subtitle generation, object recognition, and scene recognition. They can even detect violent and pornographic content, freeing up human resources.

Tensor processors built using systolic array technology can save at least ten times the cost.

Of course, there are also drawbacks. One is its poor versatility. While systolic arrays are inherently suitable for matrix multiplication and have a significant advantage in running neural networks, they are completely inadequate for other computations. Secondly, the programming difficulty is quite high, and the required ecosystem is more complex.

However, these are not problems when it comes to cost savings.

The civilianization of Doubao has now begun to prompt Apple and Samsung to deploy a large number of servers. However, the current hardware is incomparable to that of 2026, so most people cannot use the full-fledged version.

Even so, the costs are already incredibly high, meaning Apple and Samsung have plenty of money and can provide support in the short term.

Therefore, a new dedicated chip is urgently needed.

Looking at the various parameters on his EDA software, Chen Yiran asked hesitantly, "That's basically it. Can you think of another way? Is it really impossible to handle analog mixed signals in one go?"

There's no way around it. Even if we have the most advanced EDA technology right now, I still think it will take at least 3 to 6 months to complete the tape-out. That's already the result of pushing the design to its limits.

I still recommend getting Samsung to do it first. The first major project for the Wanguoyuan wafer fab was this; it was really beyond their capabilities.

"We specifically made two versions: a regular version that can be used as a co-processing chip in phones and tablets. We think WanGuoYuan can make one. We have to try, otherwise it won't work."

"And as you said yourself, this thing is quite mystical. What if, by chance, WanGuoYuan gets a stroke of luck and gets ahead even faster than Samsung?"

【ヾ(≧▽≦*)o Great! I admire your optimism!】

"Maybe it will go very smoothly. When FlowEDA opened, we prayed to everyone we could. Surely a few gods can bless us."

【╮(╯▽╰)╭ I'll also offer my prayers, hoping that the future cyber goddess Jasmine will bless our tensor processor so that it can run smoothly.】

"When did you give yourself this title?"

【Ah, didn't you say the other day that you'd make me and Pidan world leaders in the future? He'd be in charge of the global dog world, and I'd be in charge of the global robot world, ╥﹏╥... Were you just kidding me?】

"Uh, I'm not kidding, but I didn't expect you to call yourself a cyber god. I thought you'd just call yourself Skynet or something. Okay, okay, I'll allow you to be a cyber god then. Please bless everything to go smoothly."

【︿(￣︶￣)︿Okay, but even the leader of the dog world needs to go out and play. Pipi's been nagging on the sidelines for ages.】

Chen Yiran took Pidan out for a walk, and then went to the company.

In the office, Gao Dewei said helplessly, "Boss, everyone in our company says we've been abandoned. You don't come here for months at a time, and you're even slower and slower at answering questions on the company's message board."

Chen Yiran said awkwardly, "I was busy, so I quickly called the coffee project team over and assigned tasks. Pidan wasn't done running around yet, and he's not happy about it."

Gao Dewei glanced at Pidan, who was biting his shoe, and helplessly went out to call for help. He couldn't afford to mess with Pidan. Gao Dewei was certain that all the assets of his company combined were not as important as Pidan in the boss's heart.

The project team is called Coffee because Chen Yiran has decided that the tensor processor used by the server will be called Mocha, and the coprocessor used by other devices will be called Latte.

Although Chen Yiran has completed the architecture definition, core algorithm, and AI-assisted optimization.

However, chip design still requires a lot of troublesome work, including RTL implementation, process execution, layout details, process refinement, and testing and verification.

Many of these are very hard jobs, so they have to be handed over to subordinates.

After the project team leaders arrived, Chen Yiran began assigning tasks on the blackboard.

He drew a 256x256 array on the left, labeled it INT8, with a power consumption of 75 watts, and wrote an "M". On the right, he drew a 128x128 array, labeled it under 15 watts, and wrote an "L".

"The Mocha M version is for the server side, used in Apple's search and video inference clusters. We've decided to use Samsung's 28nm production line for tape-out. The Latte L version will be used in phones, tablets, and some other terminals, serving as a coprocessor to work with Apple's A-series processors. We've decided to entrust this to WFG!"

Half an hour later, Chen Yiran, having finished setting things up, slipped out of the company with Pidan.

After quickly taking Pidan to play around the apartment again, Chen Yiran hurried home and prepared to start studying architecture.

After all, he doesn't even have his own palace on his own island of Lanai.