Shulou(Shulou.com)11/24 Report--

Thanks to CTOnews.com netizens Wu Yanzu and OC_Formula for the clue delivery! CTOnews.com, May 23, the Aurora supercomputer designed by Intel for Argonne National Laboratory has been delayed for a long time, but it has finally taken shape.

It is worth mentioning that Intel, HPE, and Argonne are working together to develop a large language model for scientific computing called AuroraGPT. It is based on a basic model of 1 trillion parameters, and its scale is obviously larger than that of ChatGPT based on GPT-3.

The system, powered by Intel's Xeon CPU Max and Xeon GPU Max series chips, has been upgraded to 2 Exaflop computing power compared to the original 1 Exaflop target, putting it on a par with AMD chip-based Frontier supercomputers (CTOnews.com Note: the latter is the most powerful supercomputer ever announced on earth).

Intel revealed that the Aurora supercomputer has a total of 10624 nodes, including 21248 Xeon CPU based on the Sapphire Rapids-SP series and 63744 GPU based on Ponte Vecchio design. The system provides a peak bandwidth of 2.12 PB/s and a peak binary bandwidth of 0.69 PB/s.

For memory, the Aurora supercomputer is equipped with 10.9 PB DDR5 DRAM, 1.36 PB CPU HBM, and 8.16 PB GPU HBM.

In other words, the DRAM of this system reaches the peak bandwidth of 5.95PB/s, CPU HBM provides the peak bandwidth of 30.5PB/s, and GPU HBM is the peak bandwidth of 208.9PB/s. In terms of storage, the system is equipped with 230PB DAOS capacity and runs at a peak bandwidth of 31 TB / s, with a total of 1024 nodes configured.

Thanks to the Intel data center GPU Max family of processors, Aurora achieves the fastest SimpleFOMP performance, surpassing the NVIDIA A100 and AMD Instinct MI250X accelerators.

In addition, Intel has accelerator-free solutions in fusion reactor prediction, Monte Carlo methods (maximization) and QMCPACK (Computing Quantum Mechanical Properties).

The Aurora supercomputer will be upgraded later this year and its peak performance will exceed 2 Exaflops. It is worth mentioning that the supercomputer will also run the latest Aurora gen AI model, which can provide 1 trillion parameters.

In addition to the Aurora supercomputer, Intel also announced its new data center GPU Max subsystem, which is designed with x8 UBB and has eight Ponte Vecchio GPU.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.