Chinese semiconductor industry

Status
Not open for further replies.

tphuang

Lieutenant General
Staff member
Super Moderator
VIP Professional
Registered Member
I have deleted pandaChen posts and replies to that and issued him a warning. He is clearly trolling here. Future infractions will result in suspension. Please do not engage with him on this thread. Your post will also be deleted.
 

tonyget

Senior Member
Registered Member
View attachment 120626
well, depending on how much you believe this stuff, but this is Huawei Ascend-910B apparently being used to train Zhipu AI in comparison with A100.

In both cases, I'm assuming 8 card and using PyTorch. This seems to match what Huawei itself claimed before. i think MindSpore + Ascend-910 would yield better numbers.

Basically, it looks like they are about the same for ChatGLM-6B and 910B was a little worse for GLM-130B

Remember A100's capability is far below H100,if Ascend-910B is slightly below A100,then Huawei has lots of catch up work to do
 

CMP

Senior Member
Registered Member
The easiest way to not engage with pandaChen is to ignore list him. Everyone should do so at the earliest opportunity.
 

SanWenYu

Captain
Registered Member
Aerospace Information Research Institute of CAS released an "all-in-one AI solution for satellite images interpretation". It is based on Huawei's Ascend chips and MindSpore software framework.

For an image that covers the area of 50km by 50km, this all-in-one machine can reduce time for interpretation from hours to minutes and improve classification accuracy by 20%.

Being all-in-one, small in size, power efficient and low cost, this machine can be deployed in space, air and on land for "edge scenarios".

Please, Log in or Register to view URLs content!

突破遥感基础模型关键技术 我国科研人员研发出遥感智能训推一体机​


随着我国空天信息和人工智能技术的持续发展,基础模型已成为遥感智能解译领域的有效解决方案。然而,遥感单幅影像幅宽大、基础模型参数规模大,导致推理速度与应用场景较为受限。近日,中国科学院空天信息创新研究院(空天院)科研团队基于自主研制的遥感智能解译基础模型,联合国内相关研发机构,共同研发出“空天·灵犀”遥感智能训推一体机。该机型突破了基础模型轻量化等关键技术,实现遥感基础模型的高时效灵活部署,可以提供数据、模型、平台、硬件一体化的解决方案。

过去,遥感影像分割任务中,在单张图形处理卡上,利用基础模型处理一景50公里×50公里幅宽的遥感影像,推理耗时一般为小时级,难以满足实际应用中数据快速处理的需求。“空天·灵犀”遥感智能训推一体机集成了遥感多模态数据、轻量化基础模型、智能解译软件系统,具有多任务高精度、模型高效训练推理、低成本灵活部署及自主创新四大核心特色。其中,在多任务高精度方面,模型支持包含地物要素提取、区域变化检测等9大类36子类下游任务,多任务平均精度较经典网络模型提升6%~12%。在模型高效训练推理方面,一体机内嵌模型微调技术,仅需更新不到5%的训练参数即可实现同等效果,在昇腾环境下推理速度可提升3.5倍。在灵活自主部署方面,相较于大参数量模型通常部署在高算力、高功耗的大型服务器,“空天·灵犀”可在星载、机载、车载等多个边缘场景低成本灵活部署。在自主创新方面,一体机适配了昇腾AI环境与昇思MindSpore框架,实现了软硬一体的自主创新。

“空天·灵犀”遥感智能训推一体机搭载的高精度、高时效轻量化基础模型,能够在自然资源、交通住建、农林牧渔、应急救灾及水利等国民经济多行业实现要素精细分类、实景三维重建、地物广域搜索等多任务应用,有效提高遥感数据服务能力。以复杂环境下地物要素分类为例,模型实现了数十类典型要素的精细化自动分类,部分可细化至子类,如水生耕地、旱耕地、工业建筑、低矮住宅、商业建筑、城市道路、乡村道路等,精度较传统模式提升20%以上,单幅影像处理耗时大幅降至分钟级。

未来,基于昇腾AI的“空天·灵犀”遥感智能训推一体机将不断迭代,进一步推动遥感智能解译技术的发展。
 

interestedseal

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

Lizhao unveiled their Multi-level Pulsing RF generator, which according to them is able to support HAR etching for advanced node. RF generator is probably the most important component for plasma process tools.

What I find most surprising is that EUV also requires RF generators! Trumpf’s CO2 lasers used for hitting tin droplets in EUV are all RF-excited, and Trumpf Huettinger supplies the RF generators for their lasers exclusively.
 

tonyget

Senior Member
Registered Member
Ascend 910 is already 4-year-old. The new Ascend 920 is targeting H800 performance according to some sources.

Nvidia will release B100 next year,it will be much more powerful than H100. Remember H800 is just H100 shrink version,and we don't even know when will Ascend 920 be released. So Huawei has a long way to go to catch up Nvidia
 
Status
Not open for further replies.
Top