At the recent AWS re:Invent 2020 (Amazon re:Invent 2020) conference, AWS CEO Andy Jassy announced EC2 instances with up to eight Habana® Gaudi® accelerators. For machine learning workloads, these new EC2 instances are up to 40% more cost-effective than current GPU-based EC2 instances. Gaudi accelerators are designed for training deep learning models for workloads such as natural language processing, object detection and machine learning training, classification, recommendation, and personalization. According to reports, Gaudi®-based EC2 instances are planned to be available in the first half of 2021.
△The picture shows the HL-205 Gaudi mezzanine card from Habana Labs. Gaudi-based EC2 instances provide excellent cost-effectiveness and high performance, while natively supporting common frameworks such as TensorFlow and PyTorch. (Image credit: Habana Labs)
During the keynote session, AWS CEO Andy Jassy highlighted the huge need for high-performance, more affordable AI jobs across industries. As the company plans to introduce new EC2 instances featuring Gaudi for deep learning training, AWS will further reduce the cost of training AI datasets and reduce the total cost of operations for customers looking to leverage AI to provide business analytics, improve efficiency, and enhance user experience .
The 8-card Gaudi solution can train a ResNet-50 model on TensorFlow at 12,000 images per second. Each Gaudi processor integrates 32GB of HBM2 memory and integrates the RoCE function for interconnecting processors within the server. Technology with the AWS Elastic Architecture Adapter (EFA) scales across servers, allowing AWS and its customers to seamlessly scale the use of multiple Gaudi-based systems for efficient and scalable distributed training.
Habana’s SynapseAI software suite is designed to provide high-performance deep learning training on Habana Gaudi processors. SynapseAI integrates with popular deep learning frameworks such as TensorFlow and PyTorch to optimize Gaudi. Developers will have access to open Gaudi software, reference models and documentation. The reference model will be publicly released on Habana’s GitHub repository and will include a variety of popular models such as image classification, object detection, natural language processing, and recommender systems. The SynapseAI software suite includes Habana’s graphics compiler and real-time Tensor Processor Core (TPC) kernel libraries, firmware and drivers, as well as development tools such as the TPC SDK and SynapseAI Profiler for custom kernel development.
“We are proud that AWS has selected the Habana Gaudi processor for its upcoming EC2 training instance,” said David Dahan, CEO of Habana Labs, an Intel company. “The Habana team looks forward to continuing to work with AWS to provide customers with continuous and advanced Sexual product roadmap.”
Why it matters: As the world’s leading cloud service provider, AWS is used by developers around the world to train their AI models. However, the increased sophistication of machine learning models has pushed up the time and cost of training, especially as more data is available and developers seek to improve their models. Gaudi-based EC2 instances are designed to meet these needs by delivering cost-effective and high performance while natively supporting common frameworks such as TensorFlow and PyTorch. With Habana’s SynapseAI software suite, developers will be able to easily build new training models or port existing training models from GPUs to the Gaudi accelerator.
Intel’s AI and XPU Vision: Intel acquired Habana in 2019 to advance its AI strategy and strengthen its portfolio of AI accelerators for cloud and data centers. This includes a range of products and technologies that support the most promising AI application scenarios in business, society and scientific research. This also reflects the company’s transformation to provide a comprehensive XPU architecture across CPU, GPU, FPGA, etc. to help customers and the entire ecosystem to unlock the potential of data.
Remi El-Ouazzane, Chief Strategy Officer of Intel’s Data Platform Group, said: “Our portfolio reflects the computing challenges where artificial intelligence is not a one-size-fits-all solution. Cloud service providers have widely adopted built-in artificial intelligence. Intelligent performance Intel® Xeon® processors for AI inference workloads. With Habana, we are now able to help them reduce the cost of training AI models at scale, providing a highly competitive alternative in a high-growth market opportunity choose.”
According to the data, Habana Labs, headquartered in Israel, is a manufacturer that provides programmable deep learning accelerators for data centers. On December 16, 2019, local time in the United States, Intel Corporation announced the acquisition of Habana Labs for $2 billion, aiming to enhance Intel’s artificial intelligence product portfolio and accelerate its development in the fast-growing emerging market of artificial intelligence chips.