Intel releases Aurora supercomputing system specifications: 21248 CPUs and 63744 GPUs
In mid-2019, Intel won the bid for the Argonne National Laboratory’s supercomputer project under the U.S. Department of Energy, planning to construct the Aurora supercomputer system based on Sapphire Rapids and Ponte Vecchio, achieving exascale (ExaFLOP) computational capabilities. However, owing to Intel’s development issues causing months of delay, the Aurora supercomputer system has been significantly postponed, prompting the U.S. Department of Energy to purchase the Polaris supercomputer system, powered by AMD and NVIDIA chips, for immediate use.
Supported by Intel’s Xeon Max series CPUs and Max series GPU chips, the Aurora project was initially conceived in 2015, with an original objective of 1 ExaFLOP, which was later raised to 2 ExaFLOPs. Theoretically, this could surpass the current leader in the supercomputing Top500 list, the Frontier at the Oak Ridge National Laboratory in the United States. Recently, Intel announced that Aurora would be fully operational later this year, having already delivered more than 10,000 blade rack servers, and shared some information regarding this supercomputer.
The Aurora supercomputer system includes 21,248 Xeon CPUs based on Sapphire Rapids-SP with HBM and 63,744 GPUs based on Ponte Vecchio. It also includes 10.9PB of DDR5 memory with a peak bandwidth of 5.95 PB/s, 1.36PB of HBM connected to the CPU with a peak bandwidth of 30.5 PB/s, 8.16PB of HBM connected to the GPU with a peak bandwidth of 208.9 PB/s, and 230PB of storage capacity with a peak bandwidth of 31 TB/s. The entire system is made up of 166 racks, each comprising 64 blade rack servers, forming 10,624 nodes.
Owing to the project’s delay, not all current CPUs are Xeon Max chips, representing approximately 75%. This prevents the system from performing at full capacity, leading to Aurora’s inability to submit to the supercomputing Top500 list in a timely manner. However, when Aurora goes fully online at the end of this year, the supercomputer system is likely to achieve its anticipated performance.