Cray is enhancing its CS-Storm series GPU-accelerated systems with the addition of a new four-GPU version - the CS-Storm 500NX 4-GPU server, a 1U server with two Intel Xeon CPUs and four NVIDIA Volta GPUs, designed for customers whose AI models and HPC applications require lower GPU-to-CPU ratios for optimal performance. Including support for NVIDIA Volta GPUs, the new system is well suited for applications ranging from deep learning neural network training and inference to high-performance computing.
Implementing machine and deep learning in many organisations is a journey - from investigation to proof of concept to production applications - that data science and IT teams undertake. Different AI use cases require unique combinations of machine intelligence tools, model designs and compute infrastructure. However, no single system can address the entire spectrum of uses and models. Factors like I/O throughput, GPU-to-CPU ratio and GPU memory can have a direct impact on performance, and, ultimately, the success of the AI application.
"As companies approach AI projects, choices in system size and configuration play a crucial role", stated Fred Kohout, Cray's senior vice president of products and chief marketing officer. "Our customers look to Cray Accel AI offerings to leverage our supercomputing expertise, technologies and best practices. Whether an organisation wants a starter system for model development and testing, or a complete system for data preparation, model development, training, validation and inference, Cray Accel AI configurations provide customers a complete supercomputer system."
"We are seeing a wide range of customer use cases for GPU accelerated computing but with different configuration requirements, even within a category like deep learning neural network training", Fred Kohout stated. "Adding a smaller form factor allows our customers to choose the right node configuration for their application needs."
Cray CS-Storm series GPU-accelerated supercomputers with NVIDIA Volta 32 GB GPUs are available in three configurations to address a variety of data centre needs. For applications like distributed deep learning training where multiple GPUs are used and GPU-to-GPU communication performance is paramount, the CS-Storm 500NX 4-GPU and 8-GPU nodes include support for NVIDIA NVLink SXM2 GPUs. For applications where independent parallel GPU processing is required, the Cray CS-Storm 500GT system supports up to 10 NVIDIA Tesla PCIe V100 GPUs or FPGAs for HPC applications. All CS-Storm systems now include support for 32 GB V100 GPUs, making larger and deeper neural network models easier to use. The new 4-GPU CS-Storm system will be available in April of 2018.
Cray Accel AI fast-start offerings are available in three configurations: a stand-alone starter system for initial AI exploration, a cluster starter kit - designed for integration into existing data centre cluster environments - for initial proof-of-concept (PoC) applications, and a complete, production-level Cray cluster supercomputer for training and inference. With the addition of the CS-Storm 500NX 4-GPU node, the updated Cray Accel AI cluster starter kit makes it easier for organisations to deploy a PoC into existing data centre cluster infrastructure. All Cray Accel AI configurations come with a complete AI software environment from Bright Computing that includes the most popular AI frameworks and tools. Cray Accel AI fast start offerings make it easy for organisations just getting started with machine and deep learning to quickly move from pilot to production.