During its annual developer conference, Huawei unveiled two significant offerings: Huawei Cloud Pangu Models 3.0 and Ascend AI Cloud Services. The primary objective of Pangu Models 3.0 is to seamlessly integrate artificial intelligence into business operations, generating substantial value as a result. Ascend AI Cloud Services, on the other hand, are designed to expedite the training of AI models.
By leveraging Ascend AI’s cloud services, a single cluster can provide an impressive 2,000 petaFLOPS of computing power. Moreover, a cluster consisting of 1,000 cards can continuously train a model with billions of parameters for an uninterrupted period of 30 days.
The Huawei Cloud Pangu models cater specifically to the practical requirements of various industrial and commercial sectors. These models are built on a “5+N+X” three-layer architecture. The foundational L0 layer comprises five core models: NLP (Neuro Linguistic Programming), CV (Computer Vision), Multimodal, Prediction, and Scientific Computing. These models deliver fundamental capabilities that enable a multitude of industry-specific applications.
“The Pangu 3.0 models are available in different sizes: 10B parameters, 38B parameters, 71B parameters and 100B parameters to meet different customer needs and different latency and response time requirements,” explains Zhang Ping’an, Huawei Executive Director and CEO of Huawei Cloud. In addition, new features would be offered such as knowledge-based question and answer, text creation and code generation for the Pangu NLP model, and image generation and comprehension for the Pangu Multimodal model. All these functions are available to customers and partners and are consistent regardless of the model size (number of parameters).
Railway companies can use the Pangu Railway Model to accurately identify 67 types of freight cars and over 430 types of rail and freight car defects. “The AI software can quickly scan millions of images captured by the railway’s TFDS system and filter out 95 percent of the images that are error-free,” said Zhang Ping’an. This allows train inspectors to focus on the remaining images, helping them improve both efficiency and accuracy.
In meteorology, the Pangu Meteorology Model (or Pangu Weather) is the first AI model to outperform modern numerical weather forecasting (NWP) methods in terms of accuracy. The prediction speed is also several orders of magnitude faster. In the past, predicting the movement of a typhoon over ten days took four to five hours of simulation on a high-performance cluster of 3,000 servers. “The Pangu model is able to perform these calculations within ten seconds on a single server GPU – and with more accurate results,” explains Zhang Ping’an.
“In the field of drug research and development, it used to take an average of ten years and a billion US dollars to develop a new drug,” says Zhang Ping’an. “Now, the Pangu Drug Molecule Model can reduce lead discovery from several years to just a month and reduce drug discovery and development costs by up to 70 percent.” Using this model, a team led by Professor Liu Bing of Xi’an Jiaotong University recently discovered a new superantibiotic – the first new antibiotic in 40 years, with entirely new protein targets and in a category of its own.
“Previously, frequent GPU failures during training required frequent reboots,” comments Zhang Ping’an. “Stable Ascend AI cloud services help to reduce these time losses. The long-term stability of training with 1000 cards (30 days) reaches 90 percent, and interruptions are resolved within ten minutes.”
In addition, Huawei Cloud combines CodeArts with Pangu to create CodeArts Snap, an intelligent programming assistant for developers. “Trained with 76 billion lines of quality code and 13 million technical documents, CodeArts Snap makes generation, Q&A, and collaboration smarter,” said Zhang Ping’an.