Chinese semiconductor industry
- Thread starter Hendrik_2000
- Start date
- Status
Chongqing rises to be power semiconductor hub
Avatr 12, an EV developed by Huawei, CATL and Changan Automobile, has been rolled off the production line in Chongqing, China. Huawei plans to deepen ties with carmakers with Chongqing's advantages of location and complete automotive supply chain. The city is also rising to be a new hub of power semiconductor segment.
Tan Benhong, chairman and CEO of Avatar Technology, said the company invested CNY2 billion (US$270 million) with Huawei and CATL in the mass production of Avatr 12. The vehicle is equipped with Huawei's HarmonyOS 4 and ADS 2.0, an advanced driver assistance system.
Huawei signed a strategic cooperation agreement with Chongqing at the beginning of 2023. Industry sources said Huawei will likely become a leading automotive tier-1 supplier. Since the company holds chip resources, it can grow to be competitive rapidly.
Avatr is a high-end EV brand that Avatar Technology, CATL and Changan established in Chongqing. A Huawei-backed vehicle design institute of Seres, another Chinese EV brand, is located in the city. Many power semiconductor projects have also been conducted in Chongqing.
Changan's EV brand Deep Blue formed the Chongqing Anda Semiconductor with StarPower Semiconductor to develop automotive power semiconductors.
China-based Sanan Optoelectronics plans to form a joint venture with STM in Chongqing to mass-produce 8-inch SiC devices for STM or its customers. The JV is slated to start production in the fourth quarter of 2025. The full buildout is anticipated in 2028. The total investment is about US$3.2 billion. Sanan will invest CNY7 billion in constructing a separate 8-inch SiC substrate manufacturing facility in Chongqing to supply the JV.
The power semiconductor projects of China Resources Microelectronics and Alpha and Omega Semiconductor are also located in Chongqing. CRM invested CNY7.55 billion in a 12-inch wafer fab that can produce 30,000 to 35,000 wafers monthly.
Alpha and Omega Semiconductor has received automotive-grade certification for chip production and released the mass production of IGBT. The company manufactures 10,000 to 11,000 12-inch wafers every month for industrial and consumer electronics.
Most China-based power semiconductor companies have been located in the Yangtze Delta area. As power semiconductors gain adoption, Chongqing is gearing up for more production capacity and suppliers. That is likely why Huawei chose the city to scale up its EV business.
Chongqing's government aims to form an EV component cluster by 2027, attracting 800 companies. The government targets CNY700 billion in annual revenues.
Did Huawei mention anything about Ascend 920？Like how good will it be？We don't even know if Huawei is about to release Ascend 920，or just release another version of Ascend 910
You don't need the best AI chip to train LLM，but using the best chip is the most efficient way to do the work
Huawei did release something for AI computing a couple months back during a conference. While the Ascend 920 isn't explicitly mentioned there, they have slated a new version of the Atlas 900 Pod able to match H800 performance for release. New AI hardware designed to match Nvidia's is almost certainly in the works, even if we don't have outright confirmation of a new chip.
Why “using the best chip is the most efficient way”? Isn’t the total computational power the metric that matters? You hear many AI Computational Centers being set up in China that claim 300 or 400 PetaFLOPS. Few specifies whether that’s achieved by A100 or H100 or Ascend 910, since that’s irrelevant, right?
In the case of autos China not only tried to keep up in the combustion vehicle segment, but they went ahead and researched the next thing which was EVs. You also had an already existing battery industry which started out making lithium ion batteries for consumer electronics which could be used as a base. The change in market to EVs meant that the older combustion engine technology was just baggage. That is how China could surpass Western car makers.
China needs to have an effort where they research technology to match existing technology and research the next level of technology at the same time. The current situation in semis where they are mostly researching what the West already has is not going to be good enough. It means China will be perpetually catching up. They also need something like SSMB-EUV or better.
i think we are going to have to wait a few months to find out
new Kunpeng and Ascend chips coming.
Give Hisilicon folks some time
HiSilicon: “Ascend 920 might still be a bit slower than H100 …”
Local governments building 400 PFLOPS AI centers: “Shut up and take my money!”
Being able to train a large model in 2 months instead of 3 does not change anything in the big picture.
Instead to have a 60% market share instead of 20%, it does change a lot.
NVIDIA CEO Jensen Huang warned that “If China can't buy from the United States, they'll just build it themselves” but his words fell on dwarf ears in Washington DC. This latest round of export controls are a boon to Huawei (they deserve it after many years of pain) and will also help China to get rid of US market leadership in AI hardware, at least on the China market.
Moreover next year we will see LLM ported to smartphones, like Vivo and Xiaomi and Honor are already going to do. Qualcomm already verified their latest smartphone chip can run LLama 2 with 7B parameters at 20 tokens/s (1 token is practically 1 character in Chinese). Of course these models are not as powerful as the ones on the cloud, but because they can run locally on the phone, they will allow hundreds of million of people to experience some form of AI withouth saturating the cloud data centers, that can be dedicated to more advanced AI tasks, like business tasks.
Generalized longitudinal strong focusing in a steady-state microbunching storage ring
ABSTRACT
Steady-state microbunching (SSMB) storage rings are an appealing option for high-power extreme ultraviolet (EUV) light sources. In this paper, we propose a generalized longitudinal strong focusing (GLSF) scheme to boost the average power of short-wavelength radiation from an SSMB source with a technical demand within present reach. This scheme employs transverse-longitudinal coupling dynamics to produce steady-state ultra-short bunches and reduces the required modulation} laser power significantly by exploiting the low vertical beam emittance in a planar ring. Linear beam dynamics are studied in detail, including bunch compression, modulation cancellation, and vertical-longitudinal decoupling. An instance of a linear lattice layout is given, and kW-level quasi-continuous-wave 13.5-nm EUV radiation can be achieved in a GLSF SSMB storage ring with 1-MW modulation laser power. We believe that the GLSF scheme can also be used in conventional storage rings for various purposes concerning beam manipulation.- Status