Habana Blog
News & Discussion

Habana Showcases Gaudi2 Performance on Large Language and Generative AI Models at ISC

We’re excited to participate in this year’s ISC High Performance Compute 2023 event in Hamburg Germany. This year our team will demonstrate the capabilities of our Habana Gaudi2® processors, which deliver high-performance, high-efficiency deep learning training and inference. Gaudi2’s speed, efficiency and ease of use are particularly useful to customers wishing to take advantage of the growing innovations in large language and generative AI models.

Attendees can join us at Intel’s booth #A105 to gain useful insights from Intel and Habana experts and architects on how to optimize deep learning performance and lower total cost of ownership. We’ll show off inference of large-scale workloads on two state-of-the-art large language models – the 12 billion parameter Dolly 2.0 and 176 billion parameter BLOOM models. We’ll also show Gaudi®2’s near-linear scaling of training capacity on the Stable Diffusion model, from 1 to 64 processors,.

BLOOM (Big Science Language Open-science Open-access Multilingual) is a 176-billion-parameter model trained to perform a variety language models such as question/answer, content generation and text sequence completion. It can handle 46 different languages and 13 programming languages and is slightly larger than GPT-3. Designed and trained as part of the BigScience initiative, BLOOM is an open-science project that involves researchers and engineers from all over the world. Hugging Face, a leading innovator in state-of-the-art AI software, provided this assessment of Gaudi2 on the BLOOM model in March.

Dolly 2.0 is a new 12B parameter model released in April that enables chat, multi-question/answer and other language functions. This model was crowdsourced from 13,000 behavioral demonstrations done by 5,000 Databricks employees, who spent several months creating a new data set for commercial use. The initial data set and model details are available under the Creative Commons License, allowing full access for any purpose.

Habana’s Jarek Dukat will provide a deep dive on scaling networking capacity with Gaudi on the Voyager Supercomputer in the San Diego Supercomputer Center’s workshop, “Experience with Building and Providing Novel AI-Focused Hardware for the Science and Engineering Community.” The workshop will be conducted on Thursday, May 25 from 9 to 1pm, in Hall Y9 on the second floor.