Chinese semiconductor thread II

[tokenanalyst]

The learning curve for adding multiples of extra steps to stay with DUVi are not low either, and we can pretty much see which one is actually harder and yielding worse results from real world performance with different production lines. Just to re-emphasize this point I don’t think the last two years of process iterations we’ve seen coming out of the Huawei-SMIC partnership relative to the speed at which other fabs moved to 5 nm using EUV supports the idea that the effective performance difference between DUVi with multi-patterning and EUV are modest.

Well China is banned from buying the NXT2150i, they are achieving 5nm with basically with the I think NXT1980i or the NXT2000i that is an old scanner. Has SMIC and Huawei had access to ASML latest immersion scanners I can bet money they would be already doing 3nm with decent yields.

All the struggle that you see is because they are basically fighting hand tie.

 

[latenlazy]

Well China is banned from buying the NXT2150i, they are achieving 5nm with basically with the I think NXT1980i or the NXT2000i that is an old scanner. Has SMIC and Huawei had access to ASML latest immersion scanners I can bet money they would be already doing 3nm with decent yields.

All the struggle that you see is because they are basically fighting hand tie.

No one with access to EUV is doing an all NXT2150i line either for a reason. This isn’t just about Huawei and SMIC fighting with hands tied. I wouldn’t make that bet about 3 nm. I think that would be obscuring very real physical performance limits that will always be inherent to DUVi. EUV employment is much higher for 3 nm lines than 5 nm lines for a reason.

 

[jx191]

from April month.. this is Korean Media

1. CXMT (DRAM 및 HBM)
• 2024: ~100,000 WPM level
• 2025: ~200,000 WPM
• 2026: ~300,000 WPM target = Top 3 global rankings
• 2027: 350,000-400,000 WPM
A structure that quickly approaches the level of Samsung (500,000) and Hynix (400,000)
CXMT aims to improve constitution and economies of scale at the same time, with high value-added lines centered on DDR5 and HBM (High Bandwidth Memory) rather than traditional DDR4.
• Shanghai New Fab Construction: It is building a production facility in Shanghai that is two to three times larger than Hefei's headquarters, with the goal of operating it in 2027. Equipment imports will begin in the second half of 2026.
· Target to Expand Production Capacity: We have already expanded our DRAM production capacity from 105,000 units per month at the end of 2023 to 250,000 to 270,000 units per month. With the expansion of the Shanghai plant, our production capacity is expected to double or triple its current capacity.
• Securing HBM Exclusive Line: By 2026, we plan to allocate approximately **20%** of total production capacity to mass production of HBM3 modules. Target wafer input is known to be around 60,000 units per month.
• Hefei Phase 2 Expansion: It will secure 40,000 additional production capacity per month through the Phase 2 expansion of its existing Hefei plant, bringing its total production to over 300,000 units per month.
• Technology Node: To avoid regulation, we are continuing to introduce equipment, naming the 17nm process 18.5nm class, and we are rapidly transitioning our lines to high-performance products such as DDR5-8000 and LPDDR5X.

Please, Log in or Register to view URLs content!

From Korean media as well:

"However, industry assessments indicate that CXMT's HBM3 remains at the testing stage. Although mass production had initially been targeted for as early as the first half of this year, the schedule has continued to slip. Orders for the materials and components required for HBM3 are reportedly still at the sample production level."

A semiconductor industry source explained, "CXMT's technological progress has been rapid, but the HBM3 mass production schedule keeps being pushed back. Judging from development progress, mass production within this year appears unlikely."

No report is 100% accurate but this is a very different outlook to previous reports

 

[sunnymaxi]

Please, Log in or Register to view URLs content!

From Korean media as well:

"However, industry assessments indicate that CXMT's HBM3 remains at the testing stage. Although mass production had initially been targeted for as early as the first half of this year, the schedule has continued to slip. Orders for the materials and components required for HBM3 are reportedly still at the sample production level."

A semiconductor industry source explained, "CXMT's technological progress has been rapid, but the HBM3 mass production schedule keeps being pushed back. Judging from development progress, mass production within this year appears unlikely."

No report is 100% accurate but this is a very different outlook to previous reports

it will depend on Huawei's plan in second half of this year. they also have self-developed HBM (HiBL 1.0)..

Please, Log in or Register to view URLs content!

 

[tokenanalyst]

No one with access to EUV is doing an all NXT2150i line either for a reason. This isn’t just about Huawei and SMIC fighting with hands tied. I wouldn’t make that bet about 3 nm. I think that would be obscuring very real physical performance limits that will always be inherent to DUVi. EUV employment is much higher for 3 nm lines than 5 nm lines for a reason.

They already invested billions in EUV why would they invest billions in immersion again. The issue is, the investment in EUV started with the NXT1980i when companies thought immersion reached the limit and EUV was oversold. ASML has advanced immersion even more I think the NXT1250 is even more precise than their EUV scanners.

I think we would have a good case study if only EUV would had been banned in China but because most advanced immersion scanners are already banned is almost impossible to see the real limits of immersion, I think it would had match at least LowNA EUV, at least for while because of course ASML lowNA scanner productivity will increase with time.

 

[jx191]

it will depend on Huawei's plan in second half of this year. they also have self-developed HBM (HiBL 1.0)..

Please, Log in or Register to view URLs content!

Yes, I was concerned about how CXMT's delays could impact Huawei's own Atlas 950 rollout

Does the HiBL 1.0 and HiZQ 2.0 not require CXMT at all? I understand the HBM itself is Huawei's own, but surely they also need CXMT to help the mass production for lots of Ascend 950 DTs....

 

[tokenanalyst]

No one with access to EUV is doing an all NXT2150i line either for a reason. This isn’t just about Huawei and SMIC fighting with hands tied. I wouldn’t make that bet about 3 nm. I think that would be obscuring very real physical performance limits that will always be inherent to DUVi. EUV employment is much higher for 3 nm lines than 5 nm lines for a reason.

Don't get me wrong, I think immersion is going the way of dry ArF scanners, nobody is going to buy them, as EUV problems get solved and productivity increase. But there still gas in the tank in immersion lithography , if not ASML won't invest money developing scanners that don't sell.

 

[latenlazy]

They already invested billions in EUV why would they invest billions in immersion again. The issue is, the investment in EUV started with the NXT1980i when companies thought immersion reached the limit and EUV was oversold. ASML has advanced immersion even more I think the NXT1250 is even more precise than their EUV scanners.

I think we would have a good case study if only EUV would had been banned in China but because most advanced immersion scanners are already banned is almost impossible to see the real limits of immersion, I think it would had match at least LowNA EUV, at least for while because of course ASML lowNA scanner productivity will increase with time.

If EUV is not more performant relative to DUVi despite its higher costs fabs wouldn’t double down on a non performant investment for new process iterations. Fabs moving from 5 nm to 3 nm would be designing their iterative process shrinks with less, not more, EUV. Well run businesses don’t do sunk cost fallacies.

Don't get me wrong, I think immersion is going the way of dry ArF scanners, nobody is going to buy them, as EUV problems get solved and productivity increase. But there still gas in the tank in immersion lithography , if not ASML won't invest money developing scanners that don't sell.

My point is not that DUVi is defunct but that it is not sufficient. Not every step in an advanced node process requires the most unforgiving patterning tolerances, but that also doesn’t mean DUVi is a practical substitute for the critical patterning steps that do. It’s pretty clear what’s driving demand for new DUVi models, but the inference here is not that EUV is dispensable.

 

[sunnymaxi]

Yes, I was concerned about how CXMT's delays could impact Huawei's own Atlas 950 rollout

Does the HiBL 1.0 and HiZQ 2.0 not require CXMT at all? I understand the HBM itself is Huawei's own, but surely they also need CXMT to help the mass production for lots of Ascend 950 DTs....

yes. need CXMT to manufacture those HBM design. as i said CXMT's HBM mas production capability will reveal soon as they are planning for IPO. so they have to publish all the data. so far i also don't see Huawei changed its plan regarding 950 this year.

Edit - Cambricon too have big plans this year for Ai chip production. they too need HBM.

 

[jx191]

yes. need CXMT to manufacture those HBM design. as i said CXMT's HBM mas production capability will reveal soon as they are planning for IPO. so they have to publish all the data. so far i also don't see Huawei changed its plan regarding 950 this year.

Same, Ascend 950DT and Atlas 950 scheduled for Q4 this year and already revealed at MWC 2026 which is good. They will be very big for domestic AI compute.

I'm concerned that the HBM3 delay could completely ruin that timeline or massively reduce the supply of 950s so it's not as significant as it could be.

If CXMT is not producing HBM3 at reasonable scale, I think it could be a disaster for Huawei but I guess we will have to wait for the IPO. The IPO itself will be very big.