Chinese semiconductor thread II

tphuang

General
Staff member
Super Moderator
VIP Professional
Registered Member
don't forget Cambricon and Moore Thread.. MT even released entire line up of new Ai/GPU chips just last week.
MT AI really isn't all that impressive

Yeah I thought so. I don’t think private Chinese companies are ready to stay using guéris chips en masse. They see Huawei as a potential competitor plus they still prefer to use Nvidia chips/stack if given the chance, for obvious reasons .
if you don't know anything. You should stop posting on this thread. Seriously, this is not a crap thread for you to just post nonsense day after day. You have been warned.

That's pretty cool news but I wonder whether using Huawei's chips will be satisfactory for the companies buying. Nvidia is still desirable, but with the current environment I would expect more companies to be looking towards Huawei's stuff.
Top AI research team in China thus far do not like using CANN. This is pretty well known. However, the vast % of demand these days is in inference and there is no problem using Ascend-950PR for that. Also, they are pretty obviously Cambricon's top customer so they are likely to work intimately with them in developing the software ecosystem needed for training and fine-turning and such.

Not all companies view Huawei as a competitor. For example, Huawei has made very little progress in LLMs, at least in the ToC segment. Most of the AI companies and labs you can name are ahead of them.
Currently, this claim is only being made by a WeChat public account, and there are no other sources to corroborate it, so it still not clear
I don't see any reason to doubt that some type of large orders have been placed. The rumors behind ByteDance buying Ascend this year and next year have been going on for a while now. It's only Alibaba, Tencent and Baidu that have actively avoided Ascend. Ali & Baidu have their own AI chips.
 

tphuang

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

This is from Chengdu CEC/CSMT on their ADC chip being critical for China's advanced laser communication in satellites for LEO constellations.

HWD9361:具备每秒万次频率跳变能力,结合低本振泄漏与高增益控制精度,可有效应对太空复杂电磁干扰,保障星间、星地链路稳定传输,特别适合低轨星座的密集组网与抗干扰场景;

· HWD9009:面向C/Ku波段,带宽不低于100MHz,支持多通道扩展,满足低轨宽带星座高速数据回传与多用户并发接入的需求;

· HWD9040:覆盖Ka波段(26GHz–40GHz),实现高频段射频直采集成,适配Ka波段高速星地链路及高频段通信载荷,应对低轨卫星“高频高速、大容量”传输挑战。

此外,在卫星传感器监测领域,成都华微的HWD1263高精度ADC集成了PGA、电压基准与故障监视器,以低噪声、低漂移的特性,为应变计、热电偶、RTD等传感器提供高精度单芯片测量方案,保障信号采集的精密与稳定。
not clear the actual performance on these.
 

tphuang

General
Staff member
Super Moderator
VIP Professional
Registered Member
On MetaX plans for next year.

Please, Log in or Register to view URLs content!
12月25日举行的第七届人民网内容科技论坛上,沐曦股份研发副总裁黄向军展示了国产高性能GPU的最新进展。他在发言中表示,沐曦致力于打造世界一流的GPU芯片及计算平台,努力成为数字经济的算力基石,并已在关键技术上取得系列突破。

黄向军介绍,经过五年左右技术积累,沐曦完成了到产品矩阵落地的跨越式发展,已打造自研的完整产品线,明确三大产品系列:曦思(N系列)主要为推理产品;曦云(C系列)主打训推一体产品;此外还有曦彩(G系列),做图形渲染的产品线。目前,曦思与曦云系列产品已进入规模量产阶段。

谈及技术路线选择,黄向军强调了“通用性”这一核心理念。他解释,由于芯片研发长周期与软件快速迭代之间存在天然不匹配,且高性能GPU主要应用于多任务、多负载的数据中心场景,因此必须确保芯片具备足够的通用性以适应未来多样化的计算需求
Basically, C series for training (it unveiled C600 this year) and is working on C700 development. N series for inferences and G series for GPUs. Currently, its top seller is C500 product. C600 is fully domestic, including packaging.
 
Top