Chinese semiconductor thread II

Wrought

Captain
Registered Member
Reuters is reporting that H200 sales have been approved by Beijing. Specific customers and volumes remain to be seen.

NEW YORK, March 17 (Reuters) - Nvidia
Please, Log in or Register to view URLs content!
has won Beijing's approval to sell its second-most powerful artificial intelligence chips to China and is also preparing a version of the Groq AI ‌chip that can be sold to the Chinese market, sources familiar with the matter said. The long-awaited regulatory approval paves the way for the U.S. chipmaker to resume sales of the H200 chips, which have emerged as a major flashpoint in U.S.-China relations, in a market that once generated 13% of Nvidia's total revenue.

Please, Log in or Register to view URLs content!
 

interestedseal

Junior Member
Registered Member
Power needs to increase an order of magnitude first.
Check out this one. World’s first 10kw MOPA pulsed fiber laser. Single pulse energy and frequency are both comparable to Trumpf’s CO2 laser. Power is almost sufficient, but we don’t know about other parameters like beam quality, spectral purity and stability.
Please, Log in or Register to view URLs content!
 

tphuang

General
Staff member
Super Moderator
VIP Professional
Registered Member
when was this photo taken?

If it says just 45k wpm for 12-inch, then this must have been from a while ago. So that says Fab 6 had already been producing 14nm for a while. Also if you look at it, Wuxi fab 7 was not fully ramped up by then. This must have been the status maybe 3 years ago?

If 7-5nm have been under research since then, it's not a surprise they are ready for production by now.
 

tphuang

General
Staff member
Super Moderator
VIP Professional
Registered Member
Please, Log in or Register to view URLs content!

Huawei showcases its all new AI data infrastructure product with FusionCube A1000 AI super all in one machine.
知识库:实时生成知识、高精度检索知识

企业数据在不断变化与更新,如果知识转化不及时,会导致模型的推理决策结果不精准。华为AI数据平台采用端到端的解析优化、表征优化与检索优化,提供一个可敏锐捕捉源端数据变化的AI数据管道,并通过多模无损解析、Token级编码等方式,将文本、图片、视频等资源转化成海量细颗粒度的知识,结合多维度检索与重排序,实现超95%知识检索精度。

KV Cache加速:容纳海量历史数据,加速推理响应

在长文本长序列、多用户多并发的推理过程中,大量的KV Cache数据存不下、系统响应迟缓,导致推理体验恶化。华为AI数据平台针对性提供了全局分层缓存能力,让记忆数据在On-chip memory、DRAM、外置SSD存储三层间按需流动,显著扩充KV Cache存储池,同时融合多种稀疏注意力算法,避免重复计算,大幅扩展上下文窗口,实现首Token时延降低90%。

记忆库:高效管理上下文记忆,让模型越用越聪明

当前,大模型普遍缺乏持久的记忆,会遗忘多步骤的中间结果和长期沉淀的历史经验,无法真正实现自我进化。华为AI数据平台提供精准的记忆萃取与召回能力,实现上下文状态保持、信息凝练。记得越多,推理就越精准,模型应用才能“越用越聪明”。

本次发布的AI数据平台可提供一体化和独立式两大部署模式。一体化部署以OceanStor A800为全栈一体机底座;独立式部署采用“数据引擎节点+OceanStor Dorado”架构,可在现有存储系统上新增数据引擎节点,保护历史投资,支撑业务平滑转型。
supports KV Cache acceleration, enterprise data knowledge base, memory banks.
 

ForcedTrend

Junior Member
Registered Member
Reuters is reporting that H200 sales have been approved by Beijing. Specific customers and volumes remain to be seen.



Please, Log in or Register to view URLs content!

NVIDIA READIES GROQ CHIP FOR CHINA​

Nvidia is also preparing a version of the Groq AI chip that can be sold to the Chinese market, Reuters reported earlier on Tuesday, citing two sources familiar with the matter.

It plans to tap Groq chips for what is known as inference, where AI systems answer questions, write code or carry out tasks for users. In the products ⁠Nvidia showed this week, the company plans to use its forthcoming Vera Rubin chips, which cannot be sold in China, in combination with the Groq chips.
While Nvidia dominates the market for training AI systems, it faces much more competition in the inference market. Several major Chinese firms, including AI heavyweights such as Baidu (9888.HK), opens new tab, already produce their own ⁠inference chips.

The chips being readied for China are not downgraded versions or made specifically for the Chinese market, one of the sources told Reuters. But the new variant can be adapted to work with other systems, the source said, adding that the Groq chip is expected to be available in May.
 

PopularScience

Senior Member
Registered Member
when was this photo taken?

If it says just 45k wpm for 12-inch, then this must have been from a while ago. So that says Fab 6 had already been producing 14nm for a while. Also if you look at it, Wuxi fab 7 was not fully ramped up by then. This must have been the status maybe 3 years ago?

If 7-5nm have been under research since then, it's not a surprise they are ready for production by now.
Not recent
 

ForcedTrend

Junior Member
Registered Member
Please, Log in or Register to view URLs content!

Hisilicon's self designed CIS CS5250V200 and AI ISP Hi3519DV500 chips have been used in 骁途 S7PRO MAX Camera priced at 9999 RMB.
interesting
DJI is currently using qualcomm chips in their cameras like the Qualcomm QCS 850 in the 5 pro i hope we get a comparaison and if dji will switch to these
 

Wahid145

Junior Member
Registered Member
Please, Log in or Register to view URLs content!

Hisilicon's self designed CIS CS5250V200 and AI ISP Hi3519DV500 chips have been used in 骁途 S7PRO MAX Camera priced at 9999 RMB.
Is this a Image sensor which Sony, OmniVision, SmartSens make as well? I didn't know Huawei got into CMOS sensors, if so they are becoming a fully fledged Samsung+++ of China
 
Top