Huawei Cloud x iFLYTEK
Why HUAWEI CLOUD?
iFLYTEK's decision to partner with Huawei Cloud is grounded in considerations of technology complementarity, security assurance, localization strategy, and policy support. This collaboration aims to assist iFLYTEK in establishing a more resilient and sustainable supply chain amidst external challenges, such as sanctions. This endeavor aims to ensure the company's development and competitiveness.
After iFLYTEK was added to a trade restriction list in 2019, on October 7, 2022, sanctions were imposed on 28 leading Chinese enterprises and institutions in the field of AI and high-performance chips, including iFLYTEK. This means that it is challenging for iFLYTEK to access relevant technology and resources from certain sources. In the process of developing large models, iFLYTEK requires significant computing power and frameworks capable of supporting parallel training with 100 billion parameters. Typically, thousands of AI acceleration cards with high computing power are necessary to run continuously. These restrictions limit iFLYTEK's collaboration with certain international chip suppliers, potentially impacting its big model research and development. This indicates the necessity of transitioning to a supply chain system that isn't reliant on specific sources, with a focus on domestic suppliers.
Solutions
In this solution, an Artificial Intelligence (AI) platform is built using Elastic Compute Service (ECS), a flexible computing service provided by Huawei Cloud for easy expansion and management of computing resources. The platform utilizes API calls to connect to a public resource pool constructed by Ascend. The public resource pool allows multiple users to share these resources, thereby optimizing hardware utilization.
On this AI platform, two main tasks can be performed: model training and model inference.
Model Training: Leveraging the Ascend on the platform, deep learning algorithms can be executed to train large-scale datasets. This includes adjusting model parameters, optimizing loss functions, and updating model weights through backpropagation algorithms to progressively enhance performance.
Model Inference: Once model training is completed and meets the desired performance level, it can be deployed on the platform for inference tasks. During this stage, the model receives input data and generates corresponding outputs. By harnessing the high-performance computing capabilities of Ascend, fast and efficient inference processes can be achieved for real-time applications or batch processing tasks.
The entire process is facilitated through API calls, allowing users to access and manage platform functionalities through simple interface calls, thereby automating and enhancing flexibility in the overall workflow. Such an architecture not only provides powerful computing capabilities and resource sharing but also offers a convenient development and deployment environment for users, accelerating the research and deployment of AI applications.
Customer Benefits
When iFLYTEK collaborates with Huawei Cloud, the primary objective is to overcome the constraints of sanctions, attain technological independence, and establish a secure and stable supply chain.