<img height="1" width="1" style="display:none" src="https://www.facebook.com/tr?id=145304570664993&amp;ev=PageView&amp;noscript=1">
ETRI Graphcore partnership

Jul 06, 2022

ETRI, Graphcore partner on high-efficiency software for large models

Written By:

Minwoo Kang

Join the IPU conversation

Join our Graphcore community for free. Get help and share knowledge, find tutorials and tools that will help you grow.

Join on Slack

Graphcore and Korea’s Electronics and Telecommunications Research Institute (ETRI) have entered a multi-year partnership to develop new software approaches for high-efficiency AI compute.

Running from 2022 through 2025 and funded by the Korean government, the partnership will combine the world-leading capabilities of ETRI—Korea’s largest public research institute by R&D expenditure and license income—with Graphcore’s proven leadership in developing and commercialising efficient, high-performance compute systems for machine intelligence.

Meeting the requirements of ever-growing models

State-of-the art models for tasks such as natural language processing (NLP) have increased exponentially in size over recent years, growing from hundreds of millions of parameters to hundreds of billions of parameters, and the trend is set to continue into the trillions.

AICAS_modelgrowthTraining such models in a cost and time-efficient manner will become increasingly challenging—especially since the rate of increase in model size is no longer being met by improvements in transistor density, clock speed, and arithmetic precision.

The partnership between Graphcore and ETRI targets this problem by exploring new software approaches to improving AI computing performance, efficiency, and accessibility, and reducing the burden of ownership and dependency on single systems and technologies.

Ownership, independence, and convenience

Three main areas of consideration have been defined under the partnership:

Ownership—unlocking new performance and efficiency improvements to ease the burden associated with maintaining compute resources for training ever-larger models.

Independence—fostering an environment in which developers can choose between a broader range of competitive and efficient equipment structures and accelerator technologies.

Convenience—reducing the cost burden and complexity of developing parallelised models by identifying easier ways to apply system optimisation technology to new models.

ETRI partnership fig_1500Under the partnership, ETRI will be primarily responsible for the development of software for improving performance, efficiency, convenience, and independence. Graphcore will support ETRI through technology verification and guidance related to potential commercialisation.

“The AI revolution should be open to everyone, and not subject to prohibitive cost or time commitments or constrained by technological homogeneity,” said Kim Myung-joon, ETRI’s president. “We are particularly excited to be working with Graphcore to lower the barriers to entry and increase the range of options for developers working with the latest state-of-the-art AI models.”

“In ETRI, we have a partner who shares our commitment to identifying and developing technologies to efficiently meet the rapidly changing needs of AI,” said Fabrice Moizan, Graphcore’s SVP of Sales. “We believe our experience developing high-efficiency, massively parallel compute systems for machine intelligence will help ETRI find new ways to alleviate the technological constraints and burden of ownership associated with widely used AI compute systems."

Graphcore’s IPU technology is delivering real-world performance benefits at scale for a wide range of AI applications—from GPT and BERT for natural language processing to EfficientNet and ResNet for computer vision, to graph neural networks and many more.

Following on from this year’s launch of the latest Bow IPU processor and the Bow Pod lineup of AI computer systems, Graphcore is already developing the next generation of IPU technology that will power the world’s first ultra-intelligence AI computer capable of surpassing the parametric capacity of the brain—the Good Computer. Further updates on the Good Computer will be provided over the coming quarters, and Graphcore is keen to engage with companies and AI innovators who can help develop the next breakthroughs in AI that this ultra-intelligence machine will make possible.