These accelerators bring to Genesis Cloud AI infrastructure proven training and inference performance that delivers roughly two times that of NVIDIA* A100 GPUs on large language and multimodal models: inference that's 1.42x on a 176 billion parameter1 and 2.89x on a 7 billion parameter BLOOMZ,2 2.44x for fine-tuning T5-3B,3 and 2.84x for inference on Stable Diffusion*.4 With robust performance, efficiency, and near-linear scaling,5 the Intel Gaudi 2 AI accelerator helps Genesis Cloud offer customers highly compelling performance per price, enabling more training and inference compute for less
Through its collaboration with instances based on Intel Gaudi 2 AI accelerators, Genesis Cloud helps customers accelerate their innovation with affordable and easier-to-use deep learning model training and deployment. These instances enable customers to run a wide array of language and vision-based AI applications as well as the newly emergent generative AI applications like large language models, increasingly used in numerous industries such as retail, construction, healthcare, security, and manufacturing.
Intel has made it easier to build and deploy new models or migrate existing models with a few lines of code to get started. Our Intel® Gaudi® software is designed to optimize performance, efficiency, and ease of use with Intel® Gaudi processors. Intel Gaudi software integrates PyTorch* and TensorFlow* frameworks and Kubernetes* orchestration, and provides a host of tools and how-to content to facilitate getting started. To support a vast set of customer AI applications, we've implemented the Optimum Library on the Hugging Face* hub to help developers enable more than 50,000 AI models.
Intel and Genesis Cloud share a commitment to energy efficiency to mitigate environmental impacts. Genesis Cloud infrastructure is powered by 100% renewable energy. An evaluation by server manufacturer Supermicro* shows two times the performance per watt of the server for the Intel Gaudi 2 AI accelerator employed in Genesis Cloud as compared to the comparable Supermicro server based on the NVIDIA A100 GPU.6 This translates to instances based on the Intel Gaudi 2 AI accelerator operating on half the power as consumed by competitor solutions to train or deploy a variety of models.
1 and 2 Fast Inference on Large Language Models: Latency
3 and 4 https://huggingface.co/blog/habana-gaudi-2-benchmark
6 Power performance measurements performed by Supermicro in their lab in April 2023. Results may vary. For test details see Claims Validation.