Chinese semiconductor thread II

[henrik]

Huawei's deputy chairman says they have delivered "over 300" CloudMatrix 384's this year. Still, that is a very small number considering that Nvidia was confirmed to have delivered 1,500 GB200 NVL72's by April of this year already, and the GB200 NVL72 is far more powerful than the CloudMatrix 384 once you consider the better HBM, NVLink, and interconnect performance. And even that delivery is believed to be possible only because of reused dies from TSMC that have since been cut off. Morgan Stanley says only 23,000 Ascend 910Cs will be made this year at SMIC at a 5% yield. So it's ridiculously expensive and only enough for 59 CloudMatrix 384's even if all the Ascend 910Cs are used for the CloudMatrix 384.

And even if they can fix their Ascend 910C problem, they've got the HBM problem. And even if they fix the Ascend 910C and HBM problem, they've got the toughest problem of all which is that Mindspore's ecosystem is underdeveloped and immature compared to CUDA. So they can announce all the ambitious plans they want but the question is, what have they achieved so far? The reality is that China is being starved of AI training power.

China should restrict rare earths export, in exchange for AI training power.

 

[huemens]

GB200 NVL72 is far more powerful than the CloudMatrix 384 once you consider the better HBM, NVLink, and interconnect performance.

It is NOT.
CM384 outperforms GB200 NVL72 in every metric except for power efficiency. China has cheaper and more abundant power than those who are running NVL72 systems, so it's not really that big of an issue. Compared to NVL72, CM384 has 1.7x more compute, 3.6x more HBM, 2.1x more HBM Bandwidth, 2.1x more Scale Up Bandwidth.

 

[tphuang]

Huawei's deputy chairman says they have delivered "over 300" CloudMatrix 384's this year. Still, that is a very small number considering that Nvidia was confirmed to have delivered 1,500 GB200 NVL72's by April of this year already, and the GB200 NVL72 is far more powerful than the CloudMatrix 384 once you consider the better HBM, NVLink, and interconnect performance. And even that delivery is believed to be possible only because of reused dies from TSMC that have since been cut off. Morgan Stanley says only 23,000 Ascend 910Cs will be made this year at SMIC at a 5% yield. So it's ridiculously expensive and only enough for 59 CloudMatrix 384's even if all the Ascend 910Cs are used for the CloudMatrix 384.

And even if they can fix their Ascend 910C problem, they've got the HBM problem. And even if they fix the Ascend 910C and HBM problem, they've got the toughest problem of all which is that Mindspore's ecosystem is underdeveloped and immature compared to CUDA. So they can announce all the ambitious plans they want but the question is, what have they achieved so far? The reality is that China is being starved of AI training power.

300+ Cloud matrix 384 -> probably around 120,000 NPUs. That's in just 4 months. And at end of June, they said they already have built a 160,000 card cluster in one of their cloud data centers using CM-384 for themselves. They had 3 over ones using CM-384. So they are at least 300k for a very short time. They've likely added a lot more since then.

Have you honestly done the maths before writing crap like this?

 

[sunnymaxi]

Huawei's deputy chairman says they have delivered "over 300" CloudMatrix 384's this year. Still, that is a very small number considering that Nvidia was confirmed to have delivered 1,500 GB200 NVL72's by April of this year already, and the GB200 NVL72 is far more powerful than the CloudMatrix 384 once you consider the better HBM, NVLink, and interconnect performance. And even that delivery is believed to be possible only because of reused dies from TSMC that have since been cut off. Morgan Stanley says only 23,000 Ascend 910Cs will be made this year at SMIC at a 5% yield. So it's ridiculously expensive and only enough for 59 CloudMatrix 384's even if all the Ascend 910Cs are used for the CloudMatrix 384.

And even if they can fix their Ascend 910C problem, they've got the HBM problem. And even if they fix the Ascend 910C and HBM problem, they've got the toughest problem of all which is that Mindspore's ecosystem is underdeveloped and immature compared to CUDA. So they can announce all the ambitious plans they want but the question is, what have they achieved so far? The reality is that China is being starved of AI training power.

COPE

each and every information you posted is false and from where Morgan Stanley take this 5% yield number ?? if you believe SMIC has only 5% yield. then you seriously have some problem.. Goldman sachs/Morgan both are dogshit sources.

HW announces upgrading HW Cloud to CloudMatrix384 across 4 major DCs in Wushu, GuiAn, Ulanqab & Horinger. Each CM384 node has 384 Ascend-910C NPU + 192 Kunpeng CPU -> 300PLFOPS..

Largest Ascend cloud cluster has 432 CM-384 and > 160k AI chips (~129EFLOPS). HW Cloud now has 4 clusters upgraded to CM384.

look how these jokers still insist 200k is the limit.. LOOOOOOOOOOOL
-----------------------------------------------------------------------------------------------

there will be three Chinese companies by next year to capable to produce HBM.

CXMT HBM3 samples currently is in testing and likely to reached its full potential by next year & HBM3E by 2027. 3E could be either D1a or D1b.

Huawei's own HBM2E will produce in next Q1,2026 and Ascend 950PR.. co-developed with Swaysure..

 

[def333]

SMEE SSA800 = ASML 1980i
SMEE SSA800A = ASML 2050i

 

[HighGround]

I've been catching up on all these pages and going backwards, but I'm clearly missing something.

What has happened to Ascend 920? I feel like an idiot, because I can't find anything whether it's been canceled or renamed...?

 

[huemens]

I've been catching up on all these pages and going backwards, but I'm clearly missing something.

What has happened to Ascend 920? I feel like an idiot, because I can't find anything whether it's been canceled or renamed...?

Did Huawei ever announce an Ascend 920? I don't think so. Ascend 910D, Ascend 920, etc looks like just speculative names from the community for whatever card Huawei will release after 910C. Ascend 950 looks like a significant architectural change from current Ascend 910 series (eg: it has new fp formats, drops some old ones, uses custom HBM), which may be the reason they didn't call it 920.

 

[LanceD23]

Yuelangshen is 50% owned by Sicarrier . I am hearing they are using solid state laser different than Rslaser. Sicarriers captured alot of orders for nonlitho tools. How does a new comer beat more established domestic players like naura, smee..smic domestic tool line for 7nm even 5nm is based on sicarrier and yulangshen.

 

[tokenanalyst]

Yuelangshen is 50% owned by Sicarrier . I am hearing they are using solid state laser different than Rslaser. Sicarriers captured alot of orders for nonlitho tools. How does a new comer beat more established domestic players like Laura, smee..

There are not currently ArF solid state laser for lithography with enough power, CAS is working on it thought.

 

[snake070]

What is the source of this image?

A book

Please, Log in or Register to view URLs content!