Chinese semiconductor thread II
- Thread starter vincent
- Start date
Alot of people are saying SMIC is doing 7nm production currently with the new scanner.Yulangsheng is essentially SiCarrier (plus CASAC and some other government backers), which in turn is just Huawei but legally distinct. The machine installed would be the same ones they announced at SemiCon China this year. Given that a substantial number of engineers at SiCarrier are ex-ASML there's a good chance that it will be similar in design. How well it performs? No idea, but definitely better than the SSA800.
Also, just because I have some insider knowledge doesn't mean I know everything, or that everything I have been shown is correctI'm only this certain about SSA800 because what I have seen matches up with publically available information.
Yeah right. Grammar.
Whats wrong with that grammar that your brain couldn't comprehend ? You really dont understand or you pretending. If you don't understand what I am saying then why bother to answer my post.
Boy you have fragile mentality. He claims he has insider then its perfectly legitimate to ask him for more information.
Considering that the computing power of Ascend 910C(2025, FP16 800TFLOPS) is close to that of A100 SXM(2020, FP16 Tensor Core 624 TFLOPS), and the computing power of Ascend 970(2028, FP8 4000 TFLOPS) shall be close to that of H100 SXM(2023, FP8 Tensor Core 4000 TFLOPS), so the computing power gap between the most advanced AI chips of US and China is around 5 years? Does this indicate that the EUV would not be possibly put into chip production in China at least until 2028?(Considering that part of the performance gap between Huawei chips and NVIDIA chips is due to outdated manufacturing processes.)not sure if this has been posted:
You are quoting Dense numbers for Huawei and Sparse numbers for Nvidia. That's what Nvidia uses in marketing materials with a tiny asterisk and footnote. Nvdia H100 and H200 has FP8(dense) of only 1.9 PFLOPS. Basically you can divide the numbers that appear in Nvidia marketing materiels by 2 to get the actual number to compare with other brands. Previously Nvidia used to publish both numbers, but nowadays they only post the *with Sparsity* number to make it look much better.
Last edited:
I think computing power should also factor in the price per GPU card. Huawei's individual cards may have lower performance but should be cheaper. Building servers with a large number of chips might not necessarily be inferior to NVIDIA's solutions.
you can hook up with your insider for EUV info?Yulangsheng is essentially SiCarrier (plus CASAC and some other government backers), which in turn is just Huawei but legally distinct. The machine installed would be the same ones they announced at SemiCon China this year. Given that a substantial number of engineers at SiCarrier are ex-ASML there's a good chance that it will be similar in design. How well it performs? No idea, but definitely better than the SSA800.
Also, just because I have some insider knowledge doesn't mean I know everything, or that everything I have been shown is correct
A while back rumors of EUV coming on line 3rd quarter of 2025.
ACM Shanghai launches Ultra ECDP electrochemical stripping equipment for compound semiconductor gold etching process.
ACM Semiconductor Equipment (Shanghai) Co., Ltd. a leading provider of wafer process solutions for semiconductor front-end and advanced wafer-level packaging applications, today announced the launch of its Ultra ECDP electrochemical stripping tool, the first specifically designed for wide-bandgap compound semiconductor manufacturing. Designed for electrochemical wafer-scale gold (Au) etching outside of wafer pattern areas, this new tool delivers higher uniformity, reduced undercutting, and enhanced gold line appearance.
Ultra ECDP equipment offers specialized processes including gold bump removal, thin-film gold etching, and deep-hole gold stripping, and is equipped with an integrated pre-wet and clean chamber. Featuring precise chemical circulation and advanced multi-anode electrochemical stripping technology, the system achieves minimized lateral etching, excellent surface finish, and superior uniformity across all pattern features.
“The compound semiconductor market continues to grow, driven by strong demand from sectors such as electric vehicles, 5G/6G communications, RF, and artificial intelligence applications,” said Wang Jian, General Manager of ACM Shanghai . “Gold is becoming a preferred material for these devices due to its high conductivity, corrosion resistance, and ductility, but this also presents challenges with etching and plating. Our new Ultra ECDP tool overcomes these obstacles, providing a reliable, production-ready solution that helps our customers achieve high-performance outcomes. This is another example of how we are addressing our customers’ challenges through innovation.”
The Ultra ECDP is designed to meet the evolving requirements of compound semiconductor manufacturing, adapting to the unique physical characteristics of various substrates, including silicon carbide (SiC), gallium arsenide (GaAs), and lithium phosphate (Li₃PO₄), such as weight, stress, and thickness. The modular design of the Ultra ECDP provides flexibility, integrating plating and stripping processes within a single platform and utilizing multi-anode technology to control the stripping process in different areas. Furthermore, the Ultra ECDP offers horizontal full-surface stripping to prevent cross-contamination during processing.
Well, that's a bummer. It's seems SMEE's reputation for incompetence wasn't unfair.
How about the stuff upstream of SMEE like RSLaser, Guowang, Qier, UPrecision, etc. Do you have any idea on the performance of their components? Is SiCarrier/Yulangsheng/The Artist Formerly Known As Huawei using their stuff or do they have their own vertical?
there is a lot to this that people can read and figure out.not sure if this has been posted:
If you look at the interconnect bandwidth and compare that to which Nvidia chips have that.
Same with memory bandwidth and compare that to HBM3/3E/4, gives you a sense of when China will get to certain node level for logic and DRAMs. EUV implication with this is quite profound.
It looks to me that they will have an Ascend-950 die in 2026Q1 that will be about 500 TFLOPS FP16 or 1 PFLOPS FP8 vs 400 TFLOPS per die for 910B die used on Ascend-910C. Ascnend-960 seems to be the 2 die version.
My guess is that Ascend-970 is the next generation die since the bandwidth is so much higher. or it is also possible that 960 could be that. From 2026Q1 to 2028Q4 is almost 3 years. Thinking that they will stick with same die over that period of time is quite uncharitable, so I think we are going to see a pretty big jump in transistor density from 2025H2 (which is when the original 950 dies are produced) to 2028H1 (which is when the 970 dies are produced)