Artificial Intelligence thread
- Thread starter 9dashline
- Start date
computing system-DeepLink to jointly build an AI software and hardware ecosystem
July 7, 2023, Shanghai, China ——Suiyuan Technology cooperated with Shanghai Artificial Intelligence Laboratory to jointly build an AI software and hardware ecosystem based on the artificial intelligence open computing system-DeepLink. The current cooperation between the two parties is mainly based on the mass-produced The second-generation training product Yunsui T20 and the second-generation reasoning product Yunsui i20.
In November 2022, Suiyuan Technology and Shanghai Artificial Intelligence Laboratory began to conduct in-depth cooperation on the adaptation of Yunsui T20, Yunsui i20 and deep learning framework, and gradually realized the access to T20 and i20, jointly create an artificial intelligence software and hardware ecosystem.
At present, T20 and i20 have implemented training and reasoning adaptation of multiple models based on DeepLink, supporting core fields such as CV and NLP, as well as open source LLaMA large models. The cooperation between the two parties lowers the threshold for hardware access to AI model training and reasoning, and provides strong computing power guarantee for the development of large models.
Suiyuan Technology also participated in the DeepLink chip evaluation work. The DeepLink evaluation system conducts multi-dimensional tests on the tested chips and provides quarterly evaluation reports. The evaluation conclusions provide optimization references for various accelerator cards and promote the standardization of AI chip technical specifications.
DeepLink serves as a "bridge" that connects hardware chips and deep learning software frameworks. Manufacturers can access a wide range of algorithm ecosystems based on DeepLink's one-time adaptation, and fundamentally realize the decoupling of software and hardware. At present, DeepLink has provided more than 200 standard operator interfaces, supported more than 5,000 operator test cases, and supported both large language models and visual models. Following the DeepLink standard can achieve efficient adaptation of mainstream frameworks and AI chips, greatly reducing the threshold for using computing power, saving adaptation time, and making it easier to enable computing power to models. At the same time, through compilation optimization, the overall training efficiency is systematically improved.
Yunsui T20 is a second-generation artificial intelligence training accelerator card for data centers. It has the characteristics of wide model coverage, strong performance, and open software ecology, and can support a variety of artificial intelligence training scenarios. At the same time, it has flexible scalability and provides industry-leading artificial intelligence computing power cluster solutions.
Yunsui T20
Yunsui i20 is the second-generation artificial intelligence reasoning accelerator card for data centers. It has the characteristics of high performance, high energy efficiency, wide model coverage, easy deployment and easy operation and maintenance, and has been widely used in computer vision, speech recognition and synthesis, and natural language processing. , search and recommendation and other reasoning scenarios.
Yunsui i20
The relevant person in charge of Suiyuan Technology said: "DeepLink can effectively decouple the upstream chip and downstream framework two layers during the implementation of the adaptation project, and reuse them in different training framework adaptations, thereby reducing the chip and framework. Adapting the cost while ensuring the correctness of the operator. Suiyuan Technology will cooperate with the Shanghai Artificial Intelligence Laboratory to deeply integrate computing power and algorithms to build a more optimized computing solution integrating software and hardware for the industry. Empowerment, build a collaborative ecosystem of software and hardware, and promote the transformation and upgrading of China's digital industry."
A list of news from WAIC 2023
some notable ones
AI cloud cluster coming online in Ulanqab and Gui'an
A lot of news and a lot of huawei technology from what I can see
some notable ones
Baidu PaddlePaddle has 7.5 million developers. Significant improvement to Erniebot since being upgraded to version 3.5. Increased training and inference speed as well as model result
Huawei announce their Atlas-900 AI cluster, impressive stuff
1.8 million Ascend developers, 30 hardware partner, 1200 independent software vendors and 2500 AI business solution
AI cloud cluster coming online in Ulanqab and Gui'an
iFlytek announce their partnership with Ascend in their spark model to develop China general intelligence platform
baidu, huawei, ali & 360 group all joined China's large model team
China mobile announces its nine days larg model (I believe this uses Ascend also)
Huawei's Pengcheng-2 lab has been used to develop NLP with 200 billion parameters
A lot of news and a lot of huawei technology from what I can see
China to create and implement national standard for large language models in move to regulate AI, while using its power to transform industries
- The China Electronic Standardisation Institute, under the Ministry of Industry and Information Technology, will enact a local standard for LLMs
- Baidu, Huawei, 360 Security and Alibaba have been enlisted by the institute to lead a special task force that will draw up the new LLM standard
The China Electronic Standardisation Institute, which is under the Ministry of Industry and Information Technology (MIIT), is currently in the process of enacting a local standard for LLMs to support the growing number of fresh AI development initiatives now under way across the mainland, the agency announced on Friday at the (WAIC) in .
HUAWEI CLOUD Pangu Meteorological Large-scale Model Published in "Nature" Magazine: 10,000 Times Faster than Traditional Numerical Forecasting, Can Provide Global Weather Forecast in Seconds
According to reports, the Huawei Cloud Pangu Meteorological Large Model is the first AI model with an accuracy exceeding that of traditional numerical forecasting methods, and its speed is more than 10,000 times faster than traditional numerical forecasting. At present, the Pangea Meteorological Large Model can provide global meteorological forecasts at the second level. Its meteorological forecast results include geopotential, humidity, wind speed, temperature, sea level pressure, etc., which can be directly applied to multiple meteorological research subdivision scenarios. The European Medium-range Forecast Center and The Central Meteorological Observatory and others have found the superiority of Pangu prediction in actual measurements.
In May of this year, the direction of typhoon "Mawar" received widespread attention. The Central Meteorological Observatory stated that the Huawei Cloud Pangu large model performed well in the path forecast of "Mawa", and it predicted that it would turn to the east waters of Taiwan Island five days in advance. At the just-concluded 19th World Meteorological Congress, the European Center for Medium-Range Forecasting also pointed out that the Huawei Cloud Pangu meteorological model has an undeniable ability in accuracy, and the purely data-driven AI weather forecast model has demonstrated compatibility with numerical models. Comparable forecasting power.
The research and development team of the Huawei Cloud Pangu large model found that there are two main reasons for the lack of accuracy of the AI weather forecast model: First, the original AI weather forecast model is based on a 2D neural network, which cannot handle uneven 3D weather data well. ; Second, the AI method lacks mathematical and physical mechanism constraints, so iterative errors will continue to accumulate during the iterative process. To this end, the team creatively proposed a three-dimensional neural network (3D Earth-Specific Transformer) adapted to the earth coordinate system to process complex uneven 3D meteorological data, and used a hierarchical time-domain aggregation strategy to reduce the number of forecast iterations, thereby reducing iterations error. By training a deep neural network on 43 years of global weather data, the Pangea Meteorological Large Model surpasses traditional numerical prediction methods in terms of accuracy and speed.
HUAWEI CLOUD said that in the future, it will cooperate with global meteorological agencies to continue to explore and give full play to the application potential of AI in the meteorological field, and provide support for various industries such as agriculture, forestry, animal husbandry, fishery, aviation and navigation.
HW showed impressive AI talent training program
Ascend AI has 1.8m devs
160k+ trained from HW's partnership w/ 300+ schools & 3000+ classes
HW looking to increase this to >1000 schools & 3.5m devs over next 3 yrs +Provide money & computing resources for schools
They invest money not just in R&D but also first rate training for different businesses.
HW Cloud has 4.6m devs now.
Nvidia/Cuda's advantage over competition is not the hardware but rather being the 1st mover & having ecosystem & developer pool. They became the only player in town
The tech sanctions have forced China & especially HW to develop its own AI ecosystem & developer pool
After all, you cannot win in AI without developers
Director of the Pengcheng Lab revealed that in Oct 2020 they built the "Pengcheng Cloud Brain-2" cluster with 4000 Huawei Ascend and Kunpeng processors for 1000P FLOPS of computation power. It is about same as 4000 Nvidia A100 GPUs combined, according to him.
By Jun 2022 when the construction completed, more processors from other suppliers have been added to increase the total power to 2300P.
The next version under construction, "Pengcheng Cloud Brain-3", will provide 16000P FLOPS of computation power. He expects that it will be ready by end of 2024 or start of 2025.
By Jun 2022 when the construction completed, more processors from other suppliers have been added to increase the total power to 2300P.
The next version under construction, "Pengcheng Cloud Brain-3", will provide 16000P FLOPS of computation power. He expects that it will be ready by end of 2024 or start of 2025.