Oracle is collaborating with Nvidia to bring the GPU leader’s unified AI and HPC platform to the public cloud for accelerating analytics and machine learning workloads.
The move makes Oracle the first public cloud vendor to support Nvidia’s HGX-2 platform, the partners said this week. The Oracle cloud will support both bare-metal and virtual machine instances of the HGX-2, which is designed to “multi-precision computing,” double and single-precision floating point for accurate HPC, and reduced precision (FP16 and INT8) for AI. Nvidia’s NVSwitch interconnect with NVLink technology provides 300 GB/sec of GPU-to-GPU bandwidth to reduce the data bottlenecks encountered with HPC and deep learning workloads.
The partners said Wednesday (Oct. 10) their collaboration addresses growing demand for cloud-based GPU acceleration in data analytics as well as deep and machine learning applications. “Apart from enabling HPC and AI workloads, we’re targeting data science and analytics as a major area of investment,” Karan Batta, an Oracle product manager, noted in a blog post.
The expanded collaboration with Nvidia for cloud-based GPU acceleration underscores Oracle’s strategy of moving from a database vendor to a player in the cutthroat public cloud market. Much of the emphasis has been new platform and infrastructure services along with automated capabilities such as “self-driving” cloud database service unveiled in August. Batta noted that these and other services target data science teams, enabling them to work collaboratively on big data projects.
Oracle previously supported Nvidia’s Pascal GPU architecture cloud instances and, more recently, offered bare-metal instances of its Tesla V100 Tensor Core GPUs. Those chips were used to accelerate deep learning workloads.
Along with support for the HGX-2 platform, Oracle said its new cloud instances also would be backed by up to 48 cores of Intel Xeon processors (running at 3.5 GHz with Turbo Boost). The new instances will be available in early 2019, the cloud vendor said, noting it also had plans to offer a 16-way instance in the future.
The collaboration is the latest effort by Oracle to differentiate its cloud offerings from public cloud leader Amazon Web Services, Microsoft Azure and Google Cloud. The partnership also reflects Nvidia’s push into high-end data analytics via its just-announced RAPIDS platform, which Oracle also is supporting.
RAPIDS is a suite of software libraries for running data science and analytics pipelines entirely on GPUs. The open source libraries are intended to leverage GPU parallel processing and high-band memory via the Python programing language. Among the uses is accelerating machine learning training while improving model accuracy.
RAPIDS will be available this week on Oracle’s cloud infrastructure via Nvidia’s GPU cloud service.
Oracle also announced support for Nvidia’s GPU cloud container registry that allows cloud users to deploy container applications and frameworks on Oracle’s GPU-accelerated cloud instances. Those instances are available in multiple regions in the U.S. and Europe.