Nvidia’s next-generation GPU is adopted by the Kestrel supercomputer
The National Renewable Energy Laboratory (NREL) announced that it will build a new generation of a supercomputer called Kestrel. The new system will be built by Hewlett-Packard Enterprise (HPE) and supported by Intel Xeon Scalable Sapphire Rapids processors and Nvidia’s next-generation GPU for HPC.
Kestrel supercomputing adopts HPE’s Cray Ex architecture, which can provide about 44 FP64 PetaFLOPS performance. The system will be built in the fall of 2022 and will be completed in early 2023, mainly for energy efficiency and renewable energy research. It is not surprising that these scientific studies require stronger performance, but the combination of CPU and GPU is different from many new supercomputers currently under construction. Most of the CPUs tend to choose AMD’s EPYC series processors, while GPU is the first appearance of Nvidia A100Next Tensor Core compute GPU.
Nvidia’s GPU is likely to be a product based on the Hopper architecture and will use TSMC’s 5nm process and CoWoS advanced packaging. According to existing information, the Intel Sapphire Rapids processor and NVIDIA’s new computing card will work together via the PCIe bus. Since Intel processors do not support NVLink. Nvidia’s computing card should not support the CXL protocol. How to effectively share the memory pool is a problem worthy of attention.
The system will also use HPE Slingshot, an Ethernet architecture that achieves higher transmission speed and flow control for larger data-intensive and AI workloads. By using HPE’s Cray ClusterStor E1000 to connect more than 75 PB of parallel file systems, high-speed connections are made on the storage to meet the processing of complex workloads centered on data streams and to achieve interactive data analysis and visualization.