Chinese semiconductor thread II

[antiterror13]

AFAIK it had incredibly high defect rates and they never got it work for more than testing runs

any link? so what SMEE model now is being tested ?

 

[Aperture05]

any link? so what SMEE model now is being tested ?

I don't have links but I've been shown conversations with people in the industry (plus circumstantial evidence, such as an identical node between Kirin 9010 and 9020a which indicates a lack of equipment change) that indicate that the SSA800i was never used in production. I'm aware that this isn't ideal and you're free to distrust what I say, but I stand by that the source I got it from is credible.

As to what's actually happening with SMEE? No idea. There's some talk in other forums that SiCarrier/Huawei poached some of the SMEE people but there's been pretty much no news.

 

[LanceD23]

AFAIK it had incredibly high defect rates and they never got it work for more than testing runs

Because you said so? NICE.

 

[Aperture05]

Because you said so? NICE.

Like I said, I was shown conversations with field engineers, and it doesn't really matter whether you believe be or not anyway. What is demonstrably true is that between August 2023 when the Kirin 9000s launched with N+2 and June 2025 when the X90 launched, there has been no change to the process. Had there been an equipment change to the production line we would have seen some variations, but Techinsights have found identical processes in both. SSA800 is not in high volume production.

 

[LanceD23]

Like I said, I was shown conversations with field engineers, and it doesn't really matter whether you believe be or not anyway. What is demonstrably true is that between August 2023 when the Kirin 9000s launched with N+2 and June 2025 when the X90 launched, there has been no change to the process. Had there been an equipment change to the production line we would have seen some variations, but Techinsights have found identical processes in both. SSA800 is not in high volume production.

If you have insider tell us more about yulangshen DUVi and that's what SMIC currently using.

 

[Aperture05]

If you have insider tell us more about yulangshen DUVi and that's what SMIC currently using.

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 I'm only this certain about SSA800 because what I have seen matches up with publically available information.

 

[valysre]

If you have insider tell us more about yulangshen DUVi and that's what SMIC currently using.

a) Grammar, please. Difficult (impossible, for me) to understand.
b) Even though the semantic meaning of the sentence is difficult to understand, the syntax of "If you have insider tell us more about..." is clear. Aperature05 has been polite and has not insisted on their statements being the absolute truth and has also been reasonable in providing evidence and arguments to support the statement that they make. Nothing to provoke the "well, if you're such a know-it-all, what about [blank]" attitude.

I'd like to point out that you've also made claims that you provide zero substantiation for (links, insider info, circumstantial evidence, or otherwise):

From what I heard, SMIC is doing 7nm production with the domestic scanner. Past the trial stage.

 

[gaussgun]

Please, Log in or Register to view URLs content!

Huawei's latest announced products

• SuperPoDs
  • Atlas 950 SuperPoD (with 8,192 Ascend NPUs)
  • Atlas 960 SuperPoD (with 15,488 Ascend NPUs).
  • TaiShan 950 SuperPoD - a general-purpose computing SuperPoD
• Superclusters comprised of multiple Huawei SuperPoDs
  • Atlas 950 SuperCluster (with over 500,000 Ascend NPUs)
  • Atlas 960 SuperCluster (with over one million Ascend NPUs)
• Interconnect
  • UnifiedBus interconnect protocol for SuperPoDs
  • UnifiedBus 2.0 - released as open standard for ecosystem adoption

 

[gaussgun]

Transcript of speech by Huawei's Eric Xu

Please, Log in or Register to view URLs content!

Key Announcements & Roadmap​

• Chip roadmap in the nearer term:
  • Ascend 950 series: versions for “prefill & recommendation” (PR) and “decode & training” (DT). Features include new data formats (FP8, MXFP8, MXFP4, plus “HiF8”), higher interconnect bandwidth (2 TB/s), better vector processing.
  • Ascend 960 (targeted Q4 2027): roughly double the performance/memory/interconnect vs. 950.
  • Ascend 970 (targeted Q4 2028): further step up in FP4/8 performance, interconnect, memory bandwidth.
• New products / infrastructure:
  • Atlas 950 SuperPoD: ~8,192 Ascend 950DT chips, delivering ~8 EFLOPS (FP8) / 16 EFLOPS (FP4). Available Q4 2026.
  • Atlas 960 SuperPoD: ~15,488 Ascend 960 chips; available Q4 2027.
  • TaiShan 950 SuperPoD for general-purpose computing (finance sector, legacy mainframe/mid-range computer replacement) with Kunpeng 950 processors. Available Q1 2026.
• SuperClusters:
  • Atlas 950 SuperCluster: ~500,000+ NPUs; to be available Q4 2026; supports both UnifiedBus-over-Ethernet (UBoE) and RoCE protocols.
  • Atlas 960 SuperCluster: over 1 million NPUs; to deliver ~2 ZFLOPS in FP8 / 4 ZFLOPS in FP4. Available Q4 2027.

 

[spaceship9876]

not sure if this has been posted: