done
0.590756
Relative Brier Score
307
Forecasts
53
Upvotes
Forecasting Calendar
Past Week | Past Month | Past Year | This Season | All Time | |
---|---|---|---|---|---|
Forecasts | 6 | 49 | 338 | 307 | 1300 |
Comments | 2 | 5 | 36 | 34 | 614 |
Questions Forecasted | 6 | 32 | 64 | 52 | 191 |
Upvotes on Comments By This User | 1 | 10 | 54 | 53 | 263 |
Definitions |
Very small chance of a tie...
Some instability related to economy and potential leadership change, but much larger repression apparatus
"...gaining a detailed understanding of what is happening in China’s semiconductor industry is becoming increasingly difficult"
From the excellent article by Paul Triolo. The Evolution of China’s Semiconductor Industry under U.S. Export Controls
As I have noted previously, Chinese front-end manufacturers approach to the advanced lithography problem takes several paths. Manufacturers will continue to use advanced deep ultraviolet (DUV) tools from ASML for as long as servicing and spare parts allow. This has been further complicated by the new U.S. export controls in October 2023 and late 2024 restricting tool sales. It remains unclear how the Dutch government, working with ASML, will implement any new controls on ASML’s existing DUV installed base in China, particular at SMIC and associated facilities, including any involving Huawei affiliated personnel and R&D efforts. This raises many issues, such as overlay thresholds and how ASML handles software updates to its existing tools at now-restricted facilities in China, including SMIC’s.
This confirms the argument that there may be proof that chinese DUV exists in the products that are made, but no outside verified report that will serve for resolution (Also, see the case of the korean accused of michochip espionage in China recently: https://www.ft.com/content/f6fc898d-9f3b-4aa5-962e-621ba450ddbe )
At the same time, it is increasingly clear that SMIC will not be able to get to something that can be called a 3 nanometer process using its existing DUV tools. Most industry observers agree that 5 nanometer is the limit here. SMIC’S so-called N+3 process could allow the firm to reach close to the density level of the TSMC N6 or Samsung 5 nanometer processes. In a best-case scenario, another iteration of process technology could mean achieving a node that is just marginally less capable than TSMC’s N5 process.
Much has been written over the past year on Huawei’s efforts to pursue with SMIC and SiCarrier, techniques including quadruple patterning (SAQP)50 for using DUV down to 3-nanometer-class nodes. Huawei and SMIC have patented SAQP and other SAxP methods, as has SiCarrier. But industry experts are skeptical that it will be possible to use DUV based SAQP to get to 3-nanometer-class nodes. One well connected industry observer noted in November that “in discussions with a company that supplies materials to China’s semiconductor industry, he explained that doing SAQP the entire way is really difficult and SMIC still has not fully mastered this process.”Moving from some ability to do 7-nanometer-class node production down to 5-nanometer-class-node production is very difficult. Industry experts believe that getting to 3 nanometer-class node production via this pathway is not likely, and that efforts to move to EUV could prove successful before this could be accomplished.
If this argument is solid, someone has to explain how now Xiaomi is producing 3 nm chips too. https://www.thedailyscrumnews.com/chinas-semiconductor-leap-xiaomi-unveils-3nm-chip-without-euv-technology/
Because a detailed understanding IS increasingly difficult, this article by some respected defense analysts, is however rid with wishful thinking:
1. EDA tools for chip design have proliferated and such work can easily be outsourced to India for a fraction of the cost of a few years back.
2. Chinese have advanced rapidly in photoresist manufacturing, which was never a real chokepoint as the japanese are forced to sell to them, https://global.chinadaily.com.cn/a/202410/18/WS6711b8f9a310f1265a1c8419.html
3. As has been discussed previously, export controls of the most advanced lithography machines, and lack of an indigenous top of the line machine, does not impede many forms of contraband, adaptation, secret manufacturing, at volume .
4. Good luck with this in the Trump unilateral tariff era: "The United States needs a new multilateral export regime powered by consensus among key partners that make semiconductor manufacturing tools, many of whom would prefer to continue selling to China at some level. "
Postdata: In my first forecast I linked to the "peak lithography" argument made by some. ASML appears to have addressed this recently, but I have no access to the paywalled article:
passage of time
The intent behind this question is more interesting than the question itself: measuring the relative distance between chinese and US orgs on a topic that, by design, was built cooperatively mostly between researchers in the west and in China over the past few decades. That lately chinese orgs have produced similar products should not surprise us, nor that they are well on their way to producing better ones. It is a trajectory similar to car design and production, which many in the west have not assimilated current chinese superiority.
I write the above because a few months ago OpenAI released o1 with much fanfare, and this week Deepseek release a comparable model. Not only that, ongoing open research is increasingly more abundant and interesting from chinese orgs than comparable ones in the west, for example (cherry-picking, ofc), Alibaba's Marco Marco-o1: Towards Open Reasoning Models for Open-Ended Solutions Bytedance's InfiMM-WebMath-40B: Advancing Multimodal Pre-Training for Enhanced Mathematical Reasoning Deepseek's DeepSeek-Prover-V1.5: Harnessing Proof Assistant Feedback for Reinforcement Learning and Monte-Carlo Tree Search InternLM's InternLM2.5-StepProver: Advancing Automated Theorem Proving via Expert Iteration on Large-Scale LEAN Problems. These samples are mostly in the enhanced reasoning area, for which labs in the west have published little.