Chinese semiconductor thread II

[jli88]

The third phase of the National Integrated Circuit Industry Investment Fund (Big Fund III) invests in Nantong Crystal.​

The third phase of China’s National Integrated Circuit Industry Investment Fund (Big Fund III), known as Guotou Jixin, has invested 100 million yuan in Nantong Crystal Co., Ltd., acquiring a 25% stake and boosting its registered capital to 400 million RMB. This marks Nantong Crystal’s first completed investment by SDIC Jixin this year, following earlier plans for a potential investment in Tuojing Key Technology (still pending).

Nantong Crystal specializes in high-performance synthetic quartz materials used in lasers, semiconductors, and precision optics, particularly as core components in photomask substrates critical to advanced semiconductor manufacturing below 5nm. High-purity quartz glass is essential due to its thermal stability and accuracy in transferring chip patterns.

With the continued expansion of domestic wafer fab capacity, the market demand for photomasks is expected to grow steadily, creating vast opportunities for domestic substitution. Nantong Crystal's products precisely meet this demand. Its synthetic quartz materials for semiconductor photomasks feature low metal impurity content (high purity), high UV transmittance, high optical uniformity, low hydroxyl content, low stress birefringence, and excellent radiation resistance. It can supply quartz materials in batches to meet i-line/g-line/KrF/ArF grade photomask requirements, including ingots, blanks, and rough-polished wafers. Through controlled production processes, it ensures excellent dimensional accuracy, stress, and optical uniformity.​

Currently, the global photomask substrate market is dominated by Corning, Heraeus, and Shin-Etsu, with China’s localization rate of high-end substrates remaining under 3%. Nantong Crystal aims to break technological barriers through synthetic quartz vapor deposition technology, promoting self-reliance in this vital industry.

The company is controlled by Zhongtian Technology, a leading Chinese manufacturer of submarine optical cables with a 30–40% domestic market share. Zhongtian has expanded into optical communication modules, successfully developing and delivering industrial-grade 400G/800G silicon photonics modules.

In Q3 2025, Zhongtian Technology reported RMB 37.974 billion in operating revenue (up 10.65% YoY) and RMB 2.338 billion net profit (up 1.19% YoY), reflecting strong performance across its core businesses.

This investment underscores China’s strategic push to advance domestic capabilities in advanced semiconductor materials, particularly in photolithography, reducing dependence on foreign suppliers.​

Please, Log in or Register to view URLs content!

Chinese firms desperately need some M&A and scale in materials, specialty chemicals space. China needs its own Shin-Etsu, Corning, Toray, BASF, Zeiss etc.

Having large competitive firms helps in getting greater synergies. Toray for example can have huge market share in low and middle grade carbon fiber, which are set processes, and funnel those profits to develop higher grade carbon fiber. Zeiss can do something similar for optics.

China needs something akin to Huawei in the chemical space. Huawei has setup Zhuhai Cornerstone but its still a fresh firm.

 

[tokenanalyst]

Chinese firms desperately need some M&A and scale in materials, specialty chemicals space. China needs its own Shin-Etsu, Corning, Toray, BASF, Zeiss etc.

Having large competitive firms helps in getting greater synergies. Toray for example can have huge market share in low and middle grade carbon fiber, which are set processes, and funnel those profits to develop higher grade carbon fiber. Zeiss can do something similar for optics.

China needs something akin to Huawei in the chemical space. Huawei has setup Zhuhai Cornerstone but its still a fresh firm.

That area is advancing really fast, from EUV ULE Glass, DUV optics, ultra high purity semiconductor grade chemicals, high end ceramics, Carbon Fiber and of course photoresist.

 

[jli88]

That area is advancing really fast, from EUV ULE Glass, DUV optics, ultra high purity semiconductor grade chemicals, high end ceramics, Carbon Fiber and of course photoresist.

What I mean is having large well reknowned companies so that they take advantage of:

1. Scale
2. Synergies
3. Cross-subsidization of R&D
4. Better brand positioning and name recognition
5. One-stop solution in various fields

Zeiss, and Toray are perfect examples.

 

[tokenanalyst]

What I mean is having large well reknowned companies so that they take advantage of:

1. Scale
2. Synergies
3. Cross-subsidization of R&D
4. Better brand positioning and name recognition
5. One-stop solution in various fields

Zeiss, and Toray are perfect examples.

China semiconductor material companies are gigantic and more entrenched in the global semiconductor supply chain than people think, the issue is that talking about precursors like trimethylindium is not as sexy as talking about EUV machines.

Either way, just 5 years ago China share of domestic tools didn't reached 2%, now is almost 50% with a few of them now entering the top ten suppliers. The trend is that China high end material companies are going the same trend, high end ceramic and metal alloy companies are growing really fast and with China building their lithography supply chain I expect lithography materials, optics, chemicals, alloys and so on to grow extremely fast the coming years.
And more given that Think Tank clowns are pushing mentally challenge US politicians for more export controls.

 

[tokenanalyst]

MoEntwine: Unleashing the Potential of Wafer-scale Chips for Large-scale Expert Parallel Inference​

​

Abstract​

As large language models (LLMs) continue to scale up, mixture-of-experts (MoE) has become a common technology in SOTA models. MoE models rely on expert parallelism (EP) to alleviate memory bottleneck, which introduces all-to-all communication to dispatch and combine tokens across devices. However, in widely-adopted GPU clusters, high-overhead cross-node communication makes all-to-all expensive, hindering the adoption of EP. Recently, wafer-scale chips (WSCs) have emerged as a platform integrating numerous devices on a wafer-sized interposer. WSCs provide a unified high-performance network connecting all devices, presenting a promising potential for hosting MoE models. Yet, their network is restricted to a mesh topology, causing imbalanced communication pressure and performance loss. Moreover, the lack of on-wafer disk leads to high-overhead expert migration on the critical path.

To fully unleash this potential, we first propose Entwined Ring Mapping (ER-Mapping), which co-designs the mapping of attention and MoE layers to balance communication pressure and achieve better performance. We find that under ER-Mapping, the distribution of cold and hot links in the attention and MoE layers is complementary. Therefore, to hide the migration overhead, we propose the Non-invasive Balancer (NI-Balancer), which splits a complete expert migration into multiple steps and alternately utilizes the cold links of both layers. Evaluation shows ER-Mapping achieves communication reduction up to 62%. NI-Balancer further delivers 54% and 22% improvements in MoE computation and communication, respectively. Compared with the SOTA NVL72 supernode, the WSC platform delivers an average 39% higher per-device MoE performance owing to its scalability to larger EP.​

Please, Log in or Register to view URLs content!

 

[Michael90]

Chinese firms desperately need some M&A and scale in materials, specialty chemicals space. China needs its own Shin-Etsu, Corning, Toray, BASF, Zeiss etc.

Having large competitive firms helps in getting greater synergies. Toray for example can have huge market share in low and middle grade carbon fiber, which are set processes, and funnel those profits to develop higher grade carbon fiber. Zeiss can do something similar for optics.

China needs something akin to Huawei in the chemical space. Huawei has setup Zhuhai Cornerstone but its still a fresh firm.

Yeah, I noticed China's market/industry is more fragmented with many players. This is good for competition but at some point consolidation is needed , since some critical industries need bigger players for greater synergies and scale. A couple of big players is preferable (instead of just 1 or 2) to having dozens of small/medium players doing the same thing. So hopefully the industry will consolidate with time so they can be more competitive against the established giants from the West/Korea/Japan, else it will be difficult for small/medium Chinese companies to compete with this giants. However i think it will happen this coming years, since many of those players are still fairly new companies , so the market will take some time to develop and for M&A to happen. The market is still big and growing, since there is a huge gap to fill, so every player has a big market to fill, so the need for consolidation is still small, but as the domestic market matures and saturates, they will naturally lead to consolidation just like it happened in the smartphone market.

 

[Hyper]

MoEntwine: Unleashing the Potential of Wafer-scale Chips for Large-scale Expert Parallel Inference​

​

Abstract​

As large language models (LLMs) continue to scale up, mixture-of-experts (MoE) has become a common technology in SOTA models. MoE models rely on expert parallelism (EP) to alleviate memory bottleneck, which introduces all-to-all communication to dispatch and combine tokens across devices. However, in widely-adopted GPU clusters, high-overhead cross-node communication makes all-to-all expensive, hindering the adoption of EP. Recently, wafer-scale chips (WSCs) have emerged as a platform integrating numerous devices on a wafer-sized interposer. WSCs provide a unified high-performance network connecting all devices, presenting a promising potential for hosting MoE models. Yet, their network is restricted to a mesh topology, causing imbalanced communication pressure and performance loss. Moreover, the lack of on-wafer disk leads to high-overhead expert migration on the critical path.

To fully unleash this potential, we first propose Entwined Ring Mapping (ER-Mapping), which co-designs the mapping of attention and MoE layers to balance communication pressure and achieve better performance. We find that under ER-Mapping, the distribution of cold and hot links in the attention and MoE layers is complementary. Therefore, to hide the migration overhead, we propose the Non-invasive Balancer (NI-Balancer), which splits a complete expert migration into multiple steps and alternately utilizes the cold links of both layers. Evaluation shows ER-Mapping achieves communication reduction up to 62%. NI-Balancer further delivers 54% and 22% improvements in MoE computation and communication, respectively. Compared with the SOTA NVL72 supernode, the WSC platform delivers an average 39% higher per-device MoE performance owing to its scalability to larger EP.​

View attachment 163922

Please, Log in or Register to view URLs content!

NVL72 is not SOTA. Google OCS is SOTA. Individual Nvidia GPUs are SOTA for AI and AMD GPUs for HPC.