The DCU series of propeller and haiguang artificial intelligence accelerator card have completed mutual certification to help the artificial intelligence application innovation of domestic AI accelerator card

Time:2022-5-24

Recently, the installation and deployment test, basic function test and stability compatibility test have been carried out between Baidu PaddlePaddle deep learning framework and haiguang artificial intelligence acceleration card DCU series. The joint test results show that Baidu PaddlePaddle deep learning framework can be successfully installed in haiguang DCU series and haiguang 3000, 5000 and 7000 Series CPU environments, and can operate reliably, stably and with high performance to meet the key application needs of users.

 

The DCU series of propeller and haiguang artificial intelligence accelerator card have completed mutual certification to help the artificial intelligence application innovation of domestic AI accelerator card

Figure 1: certification certificate of ecological compatibility between propeller and haiguang DCU

 

Among them, DCU (deep computing unit depth calculator) is an acceleration card launched by hygon, which is specially used for AI Artificial Intelligence and deep learning. At present, the ROCM version of the propeller frame is based on the haiguang CPU (x86) and DCU, and supports the training and reasoning of single machine single card / single machine multi card of the following model.

 

Installation instructions and test environment instructions for ROCM version of propeller frame:

The adaptation and testing work is carried out by the propeller team and the haiguang team on the basis of haiguang 7000 Series CPU and Haiguang dcu-z100 deep computing processor chip in centos7 The correlation test is carried out under the operating system. At present, the ROCM version of the propeller frame supports Python training and native prediction based on haiguang CPU and DCU. The ROCM version of the propeller frame currently supported is 4.0.1 and 2.1.0. Two installation methods are provided:

  • Install through precompiled wheel package
  • Install through source code compilation

 

The DCU series of propeller and haiguang artificial intelligence accelerator card have completed mutual certification to help the artificial intelligence application innovation of domestic AI accelerator card

Table 1: software environment

 

The DCU series of propeller and haiguang artificial intelligence accelerator card have completed mutual certification to help the artificial intelligence application innovation of domestic AI accelerator card

Table 2: hardware environment

 

ROCM model support:

At present, 80 + models have been officially verified on haiguang DCU chip, including models of image classification (paddleclas), target detection (paddedetection), image segmentation (paddeseg), character recognition (paddeocr), generation countermeasure network (paddegan), natural language processing (paddlenlp), recommendation (padderec) and speech (Parakeet). The following table shows some supported target detection models. For complete support information, please refer to:

Propeller official website > documentation > tutorial > hardware support > haiguang DCU chip running propeller > propeller frame ROCM version support model.

(document link:

https://www.paddlepaddle.org…. )

 

Training and prediction of ROCM version of propeller frame:

Training:Training with haiguang CPU / DCU is the same as training with Intel CPU / NVIDIA GPU. The current ROCM version of the propeller frame is fully compatible with the API of the CUDA version of the propeller frame. You can directly use the original GPU training commands and parameters.

 

forecast:The prediction using haiguang CPU / DCU is the same as that using Intel CPU / NVIDIA GPU. It supports the paddle inference, which is suitable for high-performance server-side and cloud reasoning. The current ROCM version of the propeller frame is fully compatible with the C + + / Python API of the CUDA version of the propeller frame. You can directly use the original GPU prediction commands and parameters.

 

For complete training and prediction examples, please refer to the operating propeller document of haiguang DCU chip on the official website.

 

The DCU series of propeller and haiguang artificial intelligence accelerator card have completed mutual certification to help the artificial intelligence application innovation of domestic AI accelerator card

 

The artificial intelligence acceleration card DCU series independently developed by haiguang Information Technology Co., Ltd. (hereinafter referred to as “haiguang company”) can fully cover the scenes supporting in-depth learning and training, easily deal with complex neural network training, and is suitable for providing strong computing power for Artificial Intelligence Computing. Products have been widely used in telecommunications, finance, education, scientific research, artificial intelligence and other important fields.

 

The DCU series of propeller and haiguang artificial intelligence accelerator card have completed mutual certification to help the artificial intelligence application innovation of domestic AI accelerator card

 

Beijing Baidu Netcom Technology Co., Ltd. (hereinafter referred to as “Baidu”) independently developed the deep learning platform flying paddle. Based on Baidu’s years of deep learning technology research and business application, it integrates the core training and reasoning framework of deep learning, basic model library, end-to-end development kit and rich tool components. It is China’s first independently developed, feature rich, open-source and open-source industrial level deep learning platform.

 

This time, the mutual authentication between the Baidu PaddlePaddle in-depth learning framework and the haiguang artificial intelligence acceleration card DCU series will further enhance the product competitiveness of both parties in the field of domestic software and hardware, continuously improve the product performance, improve the support ability of the AI platform for customers, and help customers realize digital and intelligent upgrading and transformation.

Baidu AI developer community https://ai.baidu.com/forum , provide a platform for developers all over the country to communicate, share, answer questions and solve doubts, so that developers can no longer “fight alone” on the road of R & D, and find better technical solutions through continuous communication and discussion. If you want to try all kinds of AI technologies and explore application scenarios, join Baidu Ai Community quickly. All your visions of AI can be realized here!

Scan the QR code below and add the little assistant wechat “JD card, small customized peripheral, mysterious gift box, suitcase” and other more benefits. You can get them~

The DCU series of propeller and haiguang artificial intelligence accelerator card have completed mutual certification to help the artificial intelligence application innovation of domestic AI accelerator card