Chinese semiconductor thread II

[tphuang]

Please, Log in or Register to view URLs content!

there is actually a few things in that article which is believable.

Between 2020 and 2024, Cambricon invested Rmb5.6bn in research and development, which it ploughed into software improvements to make it easier to deploy models originally trained on Nvidia’s GPUs to run on its rival Siyuan chips. Several people, including a ByteDance AI engineer, said Cambricon’s software compatibility made its products easier to use than Huawei’s Ascend. “Cambricon struggled to gain traction until the end of 2024, when it collaborated with ByteDance to make its chip more compatible with algorithms trained on Nvidia’s ecosystem,” said Lin Qingyuan, semiconductor analyst at Bernstein.

The part important here is that Cambricon has been around longer than some of the other AI chip designers so their software is likely better. They've also had more time to work with AI labs, especially bigger players like ByteDance and Tencent.

Keep in mind that ByteDance was against working with Huawei for a long time but also needed a domestic chip option.

Many of China’s leading purchasers of AI chips, including China Telecom, Alibaba, Tencent, and Baidu, also compete with Huawei in other businesses, such as network equipment, cloud computing, and self-driving cars, creating a strong incentive for them to support alternatives to the sprawling tech conglomerate.

yes, this is true

Cambricon has struggled to meet demand after selling most of its chips in the first quarter to internet giant ByteDance, according to two people familiar with the matter, including one executive at the internet giant.

most of its sales is to ByteDance, but keep in mind that Nvidia gets 40% of its revenue from 2 customers. So this is not abnormal

“Our customers have a very high rating for its products after testing,” said one salesperson at Inspur, one of China’s biggest data centre assemblers. “There just isn’t enough supply.”

Interesting to hear this from Inspur since they are in charge of building the servers that customers use.

 

[Moonscape]

Language Processing Units for LLMs.
View attachment 159519
View attachment 159520

View attachment 159521

Please, Log in or Register to view URLs content!

This is very interesting

Fits well with China's physical applications first approach to AI

Lets you embed a 100 billion parameter-scale inference chip onto a robot for local inference

 

[tphuang]

Please, Log in or Register to view URLs content!

The latest example: China’s biggest cloud-computing company, Alibaba, has developed a new chip that is more versatile than its older chips.
Alibaba was long one of the biggest customers of American AI-chip leader Nvidia. Now it and other chip designers are filling the void left after Nvidia

Please, Log in or Register to view URLs content!

to selling its products in China.

Alibaba, founded by internet pioneer Jack Ma, is sometimes compared with Amazon.com because its biggest business is e-commerce, but it makes much of its money from the lower-profile business of cloud-computing services—running applications and storing data for customers on remote computers. Alibaba competes with Amazon Web Services, Microsoft and Google for cloud business,

Please, Log in or Register to view URLs content!

.
Alibaba said Friday that cloud revenue grew 26% in the April-June quarter on the back of surging demand for AI services. Chief Executive Eddie Wu has said “AI plus cloud” is one of Alibaba’s two engines of growth alongside e-commerce. In February, Alibaba said it would invest at least $53 billion over the next three years in the area. It also has one of the world’s highest-rated AI models, called Qwen.

The rapid adoption of AI across China’s economy is creating a

Please, Log in or Register to view URLs content!

—when AI programs tap their training to deliver output such as a smartphone voice assistant’s answers. Inference typically doesn’t require the most advanced chips.
Previous cloud-computing chips developed by Alibaba have mostly been designed for specific applications. The new chip, now in testing, is meant to serve a broader range of AI inference tasks, said people familiar with it.

The chip is manufactured by a Chinese company, they said, in contrast to an earlier Alibaba AI processor that was fabricated by Taiwan Semiconductor Manufacturing. Washington has blocked TSMC from

Please, Log in or Register to view URLs content!

that use leading-edge technology.
One challenge for Alibaba and other local players relying on Chinese chip factories is getting enough supply. These factories, which use older foreign machines and less powerful homegrown equipment, have struggled to increase capacity.

Huawei, subject to U.S. sanctions, didn’t design its chips to work with the Nvidia platform, whereas Alibaba’s new chip will be compatible with it, meaning engineers can repurpose programs they wrote for Nvidia chips.
Private-sector cloud companies including Alibaba have refrained from bulk orders of Huawei’s chips, resisting official suggestions that they should help the national champion, because they consider Huawei a direct rival in cloud services, people close to the firms said.

this is very interesting. A few years ago Alibaba T-Head designed Yitian-710 (with 128-core using ARM) and they've followed up with numerous RISC-V designs. It'd be interesting to see what the new chip will be designed as.

It does seem like Alibaba would have all the tools to develop its own chip.

 

[sunnymaxi]

Please, Log in or Register to view URLs content!

this is very interesting. A few years ago Alibaba T-Head designed Yitian-710 (with 128-core using ARM) and they've followed up with numerous RISC-V designs. It'd be interesting to see what the new chip will be designed as.

It does seem like Alibaba would have all the tools to develop its own chip.

Alibaba announced to invest 380 billion RMB over 3 years in Ai and Ai chips.

According to the changes in global AI chip supply and policy, Ali developed backup plans with different partners jointly developing different forms of supply chain reserves.

 

[tphuang]

https://twitter.com/i/web/status/1961480668795547983

My extended thoughts on Alibaba and AI. I think they had something finished designing that would've used TSMC or Samsung process, but the October surprise made that void basically. So, they had to go back to drawing board and get it designed around SMIC process.

I'm just wondering if this will be use those RISC-V core that they've designed which has strong vector and matrix support. C930 seems like a top notch core to be used in a large AI chip. If you have like a 128 core AI-chip. That's 1024 TOPS.

 

[gadgetcool5]

Please, Log in or Register to view URLs content!

this is very interesting. A few years ago Alibaba T-Head designed Yitian-710 (with 128-core using ARM) and they've followed up with numerous RISC-V designs. It'd be interesting to see what the new chip will be designed as.

It does seem like Alibaba would have all the tools to develop its own chip.

Wait, but isn't Alibaba a cloud provider too, and thus a rival to other cloud providers? You don't want to get a situation where every cloud provider has to create its own chips/ecosystem and they all refuse to cooperate with each other. China needs a single major AI chip champion to take on NVIDIA using the combined scale of China's market behind it.

 

[HighGround]

Wait, but isn't Alibaba a cloud provider too, and thus a rival to other cloud providers? You don't want to get a situation where every cloud provider has to create its own chips/ecosystem and they all refuse to cooperate with each other. China needs a single major AI chip champion to take on NVIDIA using the combined scale of China's market behind it.

This is normal.

Google and Amazon are doing the same thing to optimize on costs. It'll most likely be used by Alibaba only.