Chinese semiconductor thread II

clnt

Just Hatched
Registered Member
Its going to take a while to move off Nvidia, it is a huge sunk cost and as long as they can keep using it, they will squeeze whatever they can out of what they have. Switching is not a simple porting over exercise from CUDA to the Huawei framework, a complete code base re-design and overhaul would be needed. It goes far beyond refactoring because your code logic, systems design, architecture, basically everything was built with the assumption of using CUDA libraries, extensions, etc. Even if Huawei provided an easy toolkit for porting such as an interpreter to make it more adaptable, unless their architecture is similar to CUDA or they provide cross-compatibility your code would not work properly without big changes or would have large quality issues. We are talking millions of lines of code that would need to examined, documented, re-designed and re-implemented which is a huge sustaining engineering effort. As an example, a game that runs on PS4/PS5 versus one that runs on PC would need two different codebases because the Playstation hardware architecture requires far more SW optimization than their PC counterparts. Switching to a domestic system would have to be a gradual parallel effort because you simply cannot afford to drop everything cold turkey without losing months of work at a minimum.
Even in the west people are using cuda less and less. Cuda got bloated over the years and it’s inconvenient to use. Most ai developers who write te models use pytorch or triton nowdays. I don’t think lacking cuda would be an insourmountable obstacle for the chinese. Pretty sure they have enough developers to write the low level kernels.
 

Michael90

Junior Member
Registered Member
Its going to take a while to move off Nvidia, it is a huge sunk cost and as long as they can keep using it, they will squeeze whatever they can out of what they have. Switching is not a simple porting over exercise from CUDA to the Huawei framework, a complete code base re-design and overhaul would be needed. It goes far beyond refactoring because your code logic, systems design, architecture, basically everything was built with the assumption of using CUDA libraries, extensions, etc. Even if Huawei provided an easy toolkit for porting such as an interpreter to make it more adaptable, unless their architecture is similar to CUDA or they provide cross-compatibility your code would not work properly without big changes or would have large quality issues. We are talking millions of lines of code that would need to examined, documented, re-designed and re-implemented which is a huge sustaining engineering effort. As an example, a game that runs on PS4/PS5 versus one that runs on PC would need two different codebases because the Playstation hardware architecture requires far more SW optimization than their PC counterparts. Switching to a domestic system would have to be a gradual parallel effort because you simply cannot afford to drop everything cold turkey without losing months of work at a minimum.
I think this should be another warning for China not to really on any US/foriegn technology/product for its critical needs in her ecosystem. Since they make themselves very vulnerable..they should always invest and put maximum focus in having an alternative. I think before the government didn't believe things could change so drastically. So they were not involved in building such alternatives on a war footing. If there is one thing US sanctions have done is to wake many of them up which long term is a good thing better to suffer short term draw backs for long term security/profit. Instead of relying on US technology forever.
 

xypher

Senior Member
Registered Member
Even in the west people are using cuda less and less. Cuda got bloated over the years and it’s inconvenient to use. Most ai developers who write te models use pytorch or triton nowdays. I don’t think lacking cuda would be an insourmountable obstacle for the chinese. Pretty sure they have enough developers to write the low level kernels.
PyTorch is a general framework for building neural networks in Python, it can use many backends and the most supported one is CUDA, so that statement does not make any sense in this context - the era of writing NNs in C++ and using CUDA calls directly has ended like 15 years ago if you are talking about that, people have been writing using TensorFlow, Caffe, PyTorch, Jax, etc. for a long time now, all of them have CUDA backends as it is still the fastest for GPU-based inference. The only time when you might need to implement the network on C++ is during deployment and even there most use NVIDIA's TensorRT if they need GPU inference.

Triton is a backend, yes, but most of its optimizations (at least for now) are NVIDIA-only as well, so it's not much better than CUDA in that regard - you still need NVIDIA cards for maximum performance.
 

tokenanalyst

Brigadier
Registered Member

Huahai Qingke's first 12-inch package thinning and film bonding machine Versatile-GM300 is launched.​


Huahai Qingke Co., Ltd. shipped its first 12-inch packaging thinning and film laminating machine Versatile-GM300 to a leading domestic packaging and testing company . This product is another key core product launched by Huahai Qingke for the packaging field after the launch of the mass production model thinning and polishing machine (Versatile-GP300) in the advanced packaging field, injecting new impetus into the company's diversified business growth.

1718201940516.png
The Versatile-GM300, an integrated packaging thinning and film bonding equipment , adopts a new layout, which can flexibly realize ultra-precision grinding and dry polishing stress removal on the back of thin wafers. When used in conjunction with a wafer bonding machine, it can safely and reliably provide full-process automatic operations from precision thinning, cleaning and drying to bonding of material rings and back film peeling for 8/12-inch wafers, meeting the production process technology requirements of thin wafers in the high-end packaging field.


Versatile-GM300 ultra-precision wafer thinning machine has technically broken through the technical barriers of ultra-thin wafer thinning process . Relying on advanced thickness uniformity control technology, it can achieve intra-wafer uniformity TTV <1.0μm , reaching the domestic leading and international advanced level. In addition, this product has the advantages of high precision, high rigidity, and high flexibility in process development . It can be widely used in wafer back thinning, BG/DC film inversion and other processes in the packaging field . The launch of the 12-inch packaging thinning and film mounting all-in-one machine will help further consolidate and enhance the company's core competitiveness.

Please, Log in or Register to view URLs content!
 

tokenanalyst

Brigadier
Registered Member

Urepu successfully delivered SICE200 equipment to customers​


Recently, Shanghai Unipu Semiconductor Equipment Co., Ltd. ("Uripu") successfully delivered a wafer edge inspection device SICE200 to its customers. The device can be used for edge defect detection of silicon-based and compound semiconductor substrates and epitaxial wafers.

According to Dr. Tang Deming, General Manager of Europ, the SICE200 device launched by Europ has the following technical features :
  • Compatible with 6&8 inch SiC&Si substrates and epitaxial wafer edge detection, also suitable for edge defect detection of other compound substrates and epitaxial wafers
  • Can simultaneously realize 360° inspection of wafers (defect inspection on the front, back and edge of wafers)
  • Can simultaneously measure wafer chamfer and diameter accurately (optional)
  • Optical-mechanical system with independent intellectual property rights can achieve high resolution, high detection rate and high detection speed
  • Wafer thickness, TTV/Warp/ Bow and other parameter measurements (optional)
1718202086770.png

Dr. Tang Deming said that SICE200 has the following technical advantages in terms of overall software and defect detection algorithms :
  • Image enhancement technology: highlight edge collapse, cracks, surface scratches, stains and other defect features
  • A variety of conditional combination judgment parameterized detection tools: A variety of high-precision detection algorithm tools with high adaptability, which can accurately judge and identify defect categories and accurately classify them
  • Wafer chamfer and diameter measurement technology: Accurately measure, fit, and calculate the chamfer and diameter dimensions of the edge profile and wafer diameter. Functional modularization: The detection software functions are modularized, and the detection recipe can be quickly configured to meet the customer's personalized detection needs (optional)
  • Defect thumbnails: data retention for easy defect review (Defect Review)

Please, Log in or Register to view URLs content!
 

tokenanalyst

Brigadier
Registered Member

Tongchao Precision Holding Company Xintong Semiconductor and China-Japan-Korea Semiconductor Industrial Park reached a strategic cooperation​

with the care and support of relevant leaders of Shaoxing Xinlian Integrated Circuit, Wuhu Xintong Semiconductor Materials Co., Ltd. and China-Japan-Korea Semiconductor Industrial Park signed a strategic cooperation agreement in Shaoxing . The first phase plans to invest 25 million yuan to build a 300mm high-end cleaning factory above 14nm . Zhou Peiqi, general manager of China-Japan-Korea Semiconductor Industrial Park, and Han Yuhua, deputy general manager, Si Qifeng, chairman of Wuhu Tongchao Precision Machinery Co., Ltd., and Zhang Xixian, deputy general manager, Wu Xiaojie, general manager of Wuhu Xintong Semiconductor, and Jin Yongkuan, general manager of Shaoxing Xintong Semiconductor, attended the signing ceremony.

Xintong Semiconductor Co., Ltd. independently develops precision cleaning, ceramic spraying and anodizing processes, and can provide customers with professional PVD/CVD/ETCH process equipment parts precision cleaning, etching equipment parts Al 2 O 3 /Y2O 3 /YF 3 spraying, semiconductor/optoelectronic display etching equipment parts anodizing regeneration and new product processing services. It has the integrated surface treatment capabilities of semiconductor equipment components. With years of deep industry experience and industrial strength, it ensures the stable supply and continuous innovation of products. It has currently reached long-term cooperation with domestic semiconductor leading customers such as Yangtze Memory, Changxin Memory, Xinlian Integration, SMIC, Zhuzhou CRRC, Zhansheng Micro, Silan Jike, Fujian Jinhua, etc.

The cooperation agreement signed with the China-Japan-Korea Semiconductor Industrial Park aims to share resources with customers in the Yangtze River Delta region, build a solid strategic partnership, and promote comprehensive cooperation between the company and its strategic partners on the basis of full trust. It is a breakthrough for the company to deepen cooperation with customers. Xintong Semiconductor will take this strategic cooperation as an opportunity to give full play to its business and resource advantages, target the semiconductor market in the Yangtze River Delta region, and take advantage of the concentrated layout of semiconductor power devices, third-generation semiconductors and other industries in the region. It will provide professional equipment parts precision cleaning services and domestic substitution processing for some core parts nearby, and expand the company's comprehensive service capabilities in the field of integrated surface treatment of semiconductor equipment parts.

Please, Log in or Register to view URLs content!
Please, Log in or Register to view URLs content!

 

tphuang

Lieutenant General
Staff member
Super Moderator
VIP Professional
Registered Member
Make a note of what Huawei Whisper says here
Screen Shot 2024-06-12 at 12.23.33 PM.png
lithography machine is starting mass production. Many chipmaking related stuff is going faster than he expected, but doesn't dare to say. I mean his source is clearly people inside Huawei that would know about this kind of stuff.

So I think people shouldn't worry too much about EUV, HBM, GAA or anything else. Maybe something will take a year longer than expected, but other things could be faster
 
Top