Chinese semiconductor industry

Status
Not open for further replies.

PopularScience

Junior Member
Registered Member
I thought I read about this somewhere too, but looks like foofy posted it

I assume this is with SMIC improved 14nm process (aka 12nm). It's going to be a problem for Phytium that its new server chip is going to be worse than Loongson's 3D5000 CPU (I think that one was 1800 score on specint 2006)
But let's wait for this and D3000 to be released. I don't see how it can be competitive with Loongson's 6000 series.

Also, this tells me Huawei has good chunk of SMICs 7nm capacity since these CPU makers are stuck with SMIC's 14nm process
I read in baidu tieba
 

FairAndUnbiased

Brigadier
Registered Member
I thought I read about this somewhere too, but looks like foofy posted it

I assume this is with SMIC improved 14nm process (aka 12nm). It's going to be a problem for Phytium that its new server chip is going to be worse than Loongson's 3D5000 CPU (I think that one was 1800 score on specint 2006)
But let's wait for this and D3000 to be released. I don't see how it can be competitive with Loongson's 6000 series.

Also, this tells me Huawei has good chunk of SMICs 7nm capacity since these CPU makers are stuck with SMIC's 14nm process
Can't make that conclusion. Server chip dies tend to be big and if defect rates on 7 nm are even a little higher per area could render it too expensive.
 

Alb

New Member
Registered Member
That chart is not completely correct. NXT1960B is for 20nm node. I am certain because I was involved in its rollout for 20nm HVM.
And 20nm and 14/16nm should be split into two node families.

other than that, this is chart got it right that NXT1980 is tied to 10nm node and NXT2000 for 7nm node.

Winbond is engaged in “specialty application” DRAM. Lots of fine tuning specific to the use case is involved. This is why they are able to secure good margin and maintain some “stickiness” to hold on to their customers. The niche market they play in do not require very large DRAM memory capacity so DDR3 would suffice…their DDR4 also uses older DRAM technology than the mainstream DRAM players. Winbond is also mainly on 25nm node and older DRAM process. And since the big boys focuses more on lead ing edge, High bit density, high memory capacity space, Winbond has little competition in this niche.


If I recalled correctly, Samsung already stopped making DDR3. Winbond counts Samsung as a customer for their DDR3 and NOR needs.

JHICC's “off-the-shelf” DRAM are geared toward PC/mobile use. Lot more development is necessary for JHICC to play in the same niche market as Winbond. Not saying they won't be able to play in this arena, just not as-is. And since JHICC doesn't really have a real R&D team, I'm not holding my breath on JHICC jumping in to this very specialize and small niche market.
What JHICC have achieved is remarkable. They were the 1st Chinese semi company to be put on the infamous entity list and they managed to survive and expand production. They demonstrated that it is possible to de-americanize a 25nm dram production line. Hopefully they will be an example for all the Chinese companies that have been blacklisted.
 

tonyget

Senior Member
Registered Member
JHICC’s “off-the-shelf” DRAM are geared toward PC/mobile use. Lot more development is necessary for JHICC to play in the same niche market as Winbond. Not saying they won’t be able to play in this arena, just not as-is. And since JHICC doesn’t really have a real R&D team, I’m not holding my breath on JHICC jumping in to this very specialize and small niche market.

If JHICC doesn't want to do R&D,they can just position themself as a fab for those fabless memory companies,there is quite a few of those companies in China.
 

tphuang

Lieutenant General
Staff member
Super Moderator
VIP Professional
Registered Member
In recent China Mobile Cloud 2023, China Mobile announced its COCA (compute on chip architecture) for high performance computation
Please, Log in or Register to view URLs content!

GPU单元方面,移动云希望与合作伙伴共建自主GPU联合生态,统一算子接口标准、提供统一北向接口,兼容各GPU的深度学习框架。通过AI抽象降低客户迁移成本;通过AI加速提高训练+推理效率;通过AI池化提升GPU资源利用率。
on GPU front, China mobile wishes to form independent GPU ecosystem alliance with its partners. Who are it's GPU partners?

Well, Biren tech was there
Please, Log in or Register to view URLs content!
壁仞科技此次展示了其已量产落地的壁砺™系列通用GPU产品以及搭载的芯片。依托原创的芯片架构,该系列通用GPU具有高算力、高能效比、高通用性等优势,配合自主研发的BIRENSUPA软件开发平台,壁砺™系列能够为包括移动云在内的诸多合作伙伴,在大模型与AIGC等诸多人工智能及通用计算场景中提供强大的算力支持。
Again, announced their GPU is now mass produced and says it has high performance and high general performance performance and can be used for AIGC and big model

so was Moore Threads
Please, Log in or Register to view URLs content!
此前,摩尔线程已与中国移动云能力中心签署了战略合作备忘录,围绕云计算、云应用和加速计算等领域开展广泛合作。未来,摩尔线程还将与联合体生态伙伴一道,围绕云服务行业验证示范中心,联合产业链上下游企业、高校、科研院所等相关创新单位,开展适配验证,标准研究、技术培训、解决方案孵化、创新成果推广和人才教育培训等,通过优势互补、资源共享、开拓创新,共同推动信息技术应用创新产业项目落地,推动产业链协同发展
So Moore Threads already signed MOU with China Mobile cloud a while back in cloud computing, acceleration and such. Will they will work with partners to test and validate in more areas.

I think we maybe need a couple of years for all these domestic GPU makers products to be fully integrated with all these major cloud computing AI platforms
 

tokenanalyst

Brigadier
Registered Member

Zhongwei Company (AMEC): 23Q1 MOCVD has a significant volume of platform transformation and the future can be expected​



On April 29, the company announced the first quarter report of 2023. In 23Q1, the company achieved revenue of 1.223 billion yuan, +28.86% year-on-year; realized net profit attributable to the parent company of 275 million yuan, +134.98% year-on-year; The profit was 228 million yuan, +22.22% year-on-year.

The volume of MOCVD has increased significantly, and revenue in 23Q1 has grown steadily. In 23Q1, the company achieved revenue of 1.223 billion yuan, +28.86% year-on-year. Main departments: 1) Etching equipment: It has been recognized by more customers at home and abroad, and its market share has continued to increase; 23Q1 achieved revenue of 814 million yuan, +13.94% year-on-year. In terms of CCP etching equipment, on the world's most advanced 5nm chip production line and the next-generation more advanced production line, it has achieved multiple batch sales, and more than 200 reaction tables have been qualified in the production line. In terms of ICP etching equipment, more etching applications are continuously approved, the market is rapidly expanding and bulk orders from leading customers are continuously received. 2) MOCVD equipment: In the industrialization of the new generation of Mini-LED, it continues to maintain an absolute advantage in the blue-green LED production line. In 23Q1, it achieved revenue of 167 million yuan, +300.48% year-on-year.
 

tonyget

Senior Member
Registered Member
I thought I read about this somewhere too, but looks like foofy posted it

I assume this is with SMIC improved 14nm process (aka 12nm). It's going to be a problem for Phytium that its new server chip is going to be worse than Loongson's 3D5000 CPU (I think that one was 1800 score on specint 2006)
But let's wait for this and D3000 to be released. I don't see how it can be competitive with Loongson's 6000 series.

Also, this tells me Huawei has good chunk of SMICs 7nm capacity since these CPU makers are stuck with SMIC's 14nm process

SMIC's 14nm capacity is limited,too many customers are waiting on the line,Phytium would have to wait a while for a slot. Could this "domestic 14nm" refers to HuaHong?

This 2019 news said HuaHong's 14nm yield was 25%,I don't what's the status now
Please, Log in or Register to view URLs content!
 

tphuang

Lieutenant General
Staff member
Super Moderator
VIP Professional
Registered Member
SMIC's 14nm capacity is limited,too many customers are waiting on the line,Phytium would have to wait a while for a slot. Could this "domestic 14nm" refers to HuaHong?

This 2019 news said HuaHong's 14nm yield was 25%,I don't what's the status now
Please, Log in or Register to view URLs content!
I think SMIC is it. I mean they probably have 12 to 15k wpm of 14nm production at pretty high yield.

if we calculate 300 good CPUs per wafer, then 10k wafer alone could produce 3 million CPUs a month.

7nm I would imagine goes to Huawei and some GPU makers, basically anyone wanting to use 7nm process but can't access TSMC/Samsung 7nm process. Maybe Biren took 8 months and also taped out their GPUs with SMIC, who knows
 
D

Deleted member 24525

Guest
SMIC's 14nm capacity is limited,too many customers are waiting on the line,Phytium would have to wait a while for a slot. Could this "domestic 14nm" refers to HuaHong?

This 2019 news said HuaHong's 14nm yield was 25%,I don't what's the status now
Please, Log in or Register to view URLs content!
Phytium is owned by the cetc, it is the first in line in any bidding. Their chips have already been used to make at least one exascale supercomputer with smic's 14nm node so I think capacity probably isn't a super pressing issue.
keg.cs.tsinghua.edu.cn
Please, Log in or Register to view URLs content!
› ...PDF
BaGuaLu: Targeting Brain Scale Pretrained Models with over 37 Million ...
 
Status
Not open for further replies.
Top