PaddlePaddle, the popular deep learning platform developed by Baidu, now fully supports the Graphcore IPU for both training and inference.
The open-source code library that enables PaddlePaddle developers to take advantage of Graphcore’s advanced AI compute was announced at Wave Summit+ 2021.
The partnership means that Graphcore can now reach more than 4 million members of PaddlePaddle’s global user-base, while expanding our appeal to customers in China, where the platform is most widely used.
The relationship includes the ongoing addition of new model libraries for the IPU on PaddlePaddle, such as BERT and ERNIE.
PaddlePaddle currently provides services to 157,000 businesses, with more than 476,000 models created using PaddlePaddle, making it a significant contributor to the growth and success of China's AI community.
Graphcore is one of the first members of PaddlePaddle’s new hardware ecosystem, helping the platform’s users further accelerate AI innovation by accessing the improved model performance and speed offered by the IPU.
We now support large-scale model training, as well as high-performance inference, via the PaddlePaddle inference library.
As part of the verification process, teams from both companies used an IPU-POD16 to run BERT-Base, demonstrating the ability of PaddlePaddle and the IPU to deliver excellent performance on one of today’s most widely used AI models.
A technical deep dive on running BERT for Paddle is available on our blog.
Welcoming the announcement of PaddlePaddle’s support for the IPU, Qiao Zhao, Head of PaddlePaddle Product at Baidu said: “Graphcore has been a good friend and important partner of Baidu PaddlePaddle for some time. We are thrilled to announce that PaddlePaddle has achieved full support for training and inference on Graphcore IPU, which is a major upgrade to the PaddlePaddle hardware ecosystem.
“Models in the PaddlePaddle model library, such as BERT, have achieved excellent training and inference performance results on IPU. At the same time, the availability of Graphcore's deep learning engine PopART also enables easy adaptation of PaddlePaddle on the IPU.
“We will continue to work with Graphcore and port more models from the PaddlePaddle model library onto IPU, and create a world-class AI solution based on PaddlePaddle and IPU, empowering AI developers and companies.”
Graphcore systems have seen tremendous uptake in China, where customers are keen to accelerate today’s AI workloads, while also pushing the boundaries of artificial intelligence by developing new models and techniques that exploit the unique, built-for-AI capabilities of the IPU.
Realising the full power of Graphcore’s hardware and making it available to the greatest number of users is reliant on forging close relationships with the leading AI frameworks and tools.
In PaddlePaddle, Graphcore found not only a powerful platform partner, but a company that shared our commitment to open source, a movement which gives the power to AI developers to innovate freely.
Just as large parts of the Poplar software stack – including our PopART runtime and PopLibs libraries - are openly available to developers, PaddlePaddle code is also available on both Graphcore and PaddlePaddle’s GitHub pages.