LG AI Research has used Google Cloud's artificial intelligence (AI) infrastructure to train its latest generative AI model.
The Exaone 3.0 model is LG AI's latest generative AI family of models and was trained on Google's TPUs and GPUs via the Google Cloud AI hypercomputer.
Google's AI Hypercomputer combines the cloud giant's own-developed TPUs, GPUs, Google Cloud Storage, and the underlying Jupiter network to provide the infrastructure needed to train large-scale models.
The two companies have an ongoing collaboration, with LG AI Research previously using Google Cloud infrastructure in December 2021 to train Exaone 1.0, South Korea's first bilingual and multimodal generative AI model, and Exaone 2.0 last year.
By using Google's TPUs and GPUs, LG AI Research was able to significantly reduce the time needed for training Exaone 3.0 and improve efficiency compared to Exaone 2.0.
In total, inference processing time has been reduced by 56 percent, memory usage by 35 percent, and operating costs by 72 percent.
“Exaone 3.0 has been trained on 60 million pieces of specialized data sets, 350 million images, patents, software codes, mathematics, and chemistry. Google Cloud’s AI Hypercomputer has been instrumental in achieving LG AI Research’s goals of AI model lightweight technology and cost efficiency,” said Hwayoung Lee, lead of AI Biz Dev and DX Group at LG AI Research.
According to LG AI Research, Google's data governance and privacy controls were a motivation factor in selecting its Cloud platform.
Exaone 3.0 will be made available on the Google Cloud Marketplace.
"Our collaboration with LG AI Research exemplifies the power of partnership in pushing the boundaries of AI innovation. By providing LG with our cutting-edge AI infrastructure, we're not only accelerating their groundbreaking work but also shaping the future of industries worldwide," said Ki-sung Chi, country director of Google Cloud Korea.
LG AI Research was launched by South Korean conglomerate LG in December 2020 and is described as an 'AI think tank' dedicated to securing the latest AI source technologies and solving AI challenges.