Chinese semiconductor thread II

[tokenanalyst]

UNeXt-ILT: fast and global context-aware inverse lithography solution​

Zhejiang Univ. (China)
Zhejiang ICsprout Semiconductor Co., Ltd.. (China)
Zhejiang Technology Innovation Ctr. of CMOS IC Manufacturing Process and Design, (China)
​

Background​ Due to the limitations of optimization degrees of freedom in traditional optical proximity correction, it cannot meet the mask optimization requirements of advanced technology nodes. Inverse lithography technology (ILT) is considered the most promising resolution enhancement technique, and the trade-off between mask optimization quality and computation time is a challenge.​ Aim​ The biggest limitation of ILT is its high computational complexity, which requires exploring an ILT algorithm that can ensure the fidelity of lithography patterns and the process variation (PV) band while also having a short computation time.​ Approach​ We propose UNeXt-ILT, a deep learning–based ILT technology. The UNeXt model is adopted as the backbone model, and its multi-layer perceptron structure ensures the lightweight of the model while having global context-awareness capability, thus quickly providing a high-quality initial mask and accelerating the overall computation time. In addition, the addition of mask regularization and mask filtering techniques enhances the robustness of gradient descent–based ILT algorithms and further improves the quality of mask optimization.​ Results​ Compared with the most advanced deep learning–based ILT algorithm, UNeXt-ILT reduces L2 error by 17.83%, reduces PV band by 8.76%, and shortens turnaround time by 34.48%. Please, Log in or Register to view URLs content!

 

[tphuang]

Please, Log in or Register to view URLs content!

Another article from the Koreans on how Samsung and SK are concerned about the DDR5 from CXMT. It's quite clear by this point that YMTC has already disrupted by 3D NAND market, so the main area that the memory chip makers are resting their hopes on is HBM3/HBM3E and such. And clearly, you can get to HBM3 with D1z as shown by Samsung.

 

[gotodistance]

Please, Log in or Register to view URLs content!

Another article from the Koreans on how Samsung and SK are concerned about the DDR5 from CXMT. It's quite clear by this point that YMTC has already disrupted by 3D NAND market, so the main area that the memory chip makers are resting their hopes on is HBM3/HBM3E and such. And clearly, you can get to HBM3 with D1z as shown by Samsung.

Korean media is trash media. Don't trust Korean media articles too much. Make sure they are verified. Koreans don't trust Korean news either. There is too much fake news.

 

[tonyget]

Please, Log in or Register to view URLs content!

Another article from the Koreans on how Samsung and SK are concerned about the DDR5 from CXMT. It's quite clear by this point that YMTC has already disrupted by 3D NAND market, so the main area that the memory chip makers are resting their hopes on is HBM3/HBM3E and such. And clearly, you can get to HBM3 with D1z as shown by Samsung.

There is a pic in that news about production capacity

 

[AETHER]

interesting post on reddit regarding Huawei AI chips

https://www.reddit.com/r/LocalLLaMA/comments/1iadomi

Note I have no proof of this other than my word.

Recently met with a Huawei employee who was pitching their 910B chips for GenAI. We didn't end up going with them, but in the process I learned some interesting tidbits of information:

• Huawei 910C is the same architecture as 910B
• The 910C is aiming for 800 TFLOPS of fp16 (unclear if fp32 accumulate, or fp16) -- it was mentioned that their goal is around Nvidia H200 NVL
• The 910C is on a Chinese 7nm process
• The 910C aims to use Chinese HBM2e, they provided no comment regarding capacity or bandwidth
• The 910C aims to resolve serious cross-card interconnect issues present in the 910B, which rendered the 910B unsuitable for training LLMs
• They mentioned that the chief designer of Huawei Ascend chips, who did the first Ascend design was a Chinese student educated in the USA. No details provided on if he was undergrad or PhD educated in the US. But mentioned his initial design focus was edge/low-power inference. They mentioned that a significant part of their EDA & compiler teams had undergrad/PhD US educations.
• They are aiming for an exact silicon doubling of the 910B. They suggested this was done via chiplets, but were evasive when I pushed for details and tried to confirm this
• Their goal is public sampling in 2025 Q1 or Q2
• They claimed better Pytorch compatibility than AMD, and said it was comparable to Intel's current GPU compatibility
• They claimed significant PyTorch compatibility improvements since 2024 Q1, since the 910B launched. And mentioned that a large effort was put into Pytorch operator compatibility/accuracy under fp16, and their own NPU API called ACL
• They grumbled about 910B being prioritized to some "cloud" infrastructure customers who didn't have a viable cloud business, and required significant on-site ecosystem support. They liked working with the GenAI startups who had the skills for scale out infrastructure
• They mentioned that demand outstripped supply as a whole
• They grumbled about certain customers still preferring to use smuggled Nvidia chips rather than their solution
• They grumbled about having to be bug compatible with Nvidia, and efforts to resolve accuracy issues
• They are aiming for a new architecture for whatever succeededs 910C

 

[tphuang]

interesting post on reddit regarding Huawei AI chips

https://www.reddit.com/r/LocalLLaMA/comments/1iadomi

Let's keep our sources high and not post random "trust me" post out of reddit please?

Nothing important here are things we don't already know. They've delivered quite a few Ascend-910C to big tech in China last year. I don't know what is meant by public sampling. People are already using it.
I also don't know why it's surprising Ascend what would better PyTorch compatibility than others. They've been working intimately with PyTorch for a couple of years now.
In fact, people that tested both Ascend and Cuda with PyTorch said Ascend works better with PyTorch than Cuda. But guess what, incumbency matters.

Korean media is trash media. Don't trust Korean media articles too much. Make sure they are verified. Koreans don't trust Korean news either. There is too much fake news.

Can you point out to which part of the article is trash or fake? I didn't see anything out of ordinary for me.

 

[gotodistance]

Can you point out to which part of the article is trash or fake? I didn't see anything out of ordinary for me.

=> The Korean media is all trash media that shouts anti-Chinese sentiment and anti-communism. They make news about China facts, but end with negative news about China in the middle or at the end. That's why when you watch Korean media news, you end up with a negative perception of China.

“SK하이닉스 중국 공장에서 생산한 D램으로 장난친 것”
=> Playing with DRAM produced at SK Hynix’s Chinese factory

The above Korean article is a news article based on TechInsights. It is wiser to read the original TechInsights article.

 

[tokenanalyst]

Mingga Semiconductor successfully prepared 4-inch gallium oxide crystal for the first time.​

Beijing Mingga Semiconductor Co., Ltd. has successfully produced a 4-inch gallium oxide crystal blank for the first time in the world. The successful gallium oxide crystal blank grew to a diameter of 4 inches and a thickness of 55 mm. After processing, the usable size is 3 inches and the thickness is up to 40 mm. Under the same process conditions, the more device chips are manufactured with gallium oxide, the lower the unit device chip manufacturing cost, and the more obvious the cost-effectiveness of gallium oxide.

This technological breakthrough not only optimizes the crystal growth environment, but also makes it possible to produce a crystal blank with a usable thickness greater than 40 mm. This is not only a technological breakthrough, but also promotes the development of the industry. Large-sized gallium oxide substrates can effectively reduce production costs, improve production efficiency, and accelerate the widespread application of gallium oxide materials in various fields.

As a fourth-generation semiconductor material, the cost of gallium oxide is about one-third of that of the third-generation semiconductor material silicon carbide. Its basic power devices have the characteristics of high voltage resistance and low loss. The products made from it have significantly improved energy efficiency and endurance performance. For example, in the field of new energy vehicles, the driving range of a car using gallium oxide devices on a single charge is 10% to 20% higher than that of a car using traditional silicon devices.

Please, Log in or Register to view URLs content!