Artificial Intelligence thread
- Thread starter 9dashline
- Start date
Is ChatGLM2 by Tsinghua really so good? I read it is a open-source bilingual model.
Huawei Ascend AI platform has over 1.8 million developers - This is a huge number. This would be probably on the same magnitude as Cuda and they are all just writing applications for the GPU and integration with software
中国工程院 using Huawei's Pengcheng lab to train a 200 Billion parameter NLP platform. Industry's first Chinese language NLP
20+ domestic large models including Pangu & 紫东.太初 developed using Ascend AI platform. It has been adapted to support ChatGLM、LLaMA、GPT-3、BLOOM & other open source large model
more details here, Huawei working with Zhipu, Cloudwalk, iFlytek & ModelBest reelease Ascend AI large model training & pushed integration solution to accelerate the application of large models in industries
23 Ascend Partners launched AI servers & intelligent edge product.
CAICT, Huawei & other units established "Large model industry working group" to jointly promote the implementation of China's LLMs.
30+ LLMs using Ascend now, more than a couple of months ago
Current size and scope of Ascend platform. A lot of developers, a lot of large models, research/lab partners & ecosystem partners
half of the large models in China developed using Ascend Platform.
They are building a 16 EFLOPS Pengcheng-3 project in Shenzhen and will work with universities to train AI talents. Remember 3 elements to AI:
computing power
data
Algo
That last part relies on AI talents
Here the million dollar question is of course: can Huawei currently manufacture/outsource Ascend chips?
I've not seen yet a clear confirmation of this.
Of course we have many hints that it can....but many hints do not make a confirmation.
Even US media, like WSJ, is quite silent on this topic. They usually are very quick to cry wolf as soon as CIA pass them info they want to spread...in many cases just fake news. But nevertheless we have nothing here.
The public released model and weights refer to the 6B version (6 billions parameter), instead the one topping is the big version, possibly the 130B version if, as it seems, is an improvement on the previous GLM-130B. Weights of this big version have not been released.
Last edited:
well, Huawei/Ascend guy said they will have a new Kunpeng & Ascend chip later this year. You can make your own guesses on where that's made. Maybe they will tell us, but most likely not
The specs for the GPU should be easy to derive.
PengCheng-3 lab is expected to be 16 EFLOPs (FP16 I think). Atlas-900 cluster can be as large as 16000 cards based on the unveiling from today. So each card should be doing about 1024 TFLOPS FP16. That's basically what Biren's BR100 is performing except that they are producing nerfed version of BR104 to stay under the radar. But fabbing it on 7nm should be doable.
The 130B model need server like hardware or a cloud service to run properly, the 6B models can run on consumer hardware and its seems to be pretty good.