The U.S. Department of Energy’s Argonne National Laboratory has paired Nvidia’s A100 GPU with AMD’s CPU for its new Polaris supercomputer Largely considered necessary Because Delay of Aurora supercomputer powered by Intel This ran into obstacles due to production difficulties associated with Intel’s Sapphire Rapids server chip.
Polaris will use four Nvidia A100s and two AMD CPUs for each node, with a total of 2,240 Nvidia GPUs distributed across 560 nodes. Polaris will provide up to 44 PetaFLOPS of FP64 performance, which means it is much weaker than the planned Aurora exascale supercomputer, which will provide up to 1 ExaFLOP of sustained performance when it arrives in 2022. In contrast, Polaris will achieve 1.4 performance “AI ExaFLOPS”, not measured by the standard FP64 workload used to quantify supercomputer performance. This means that Polaris is not a tens of billions of machines-we have not seen the planned Nvidia-powered supercomputer to meet that standard.
However, Polaris’s 44 PetaFLOPs performance is eligible for Top ten in the top 500 A list of the fastest supercomputers in the world. Polaris is in the final stages of installation and will prepare for its first research work in early 2022 and expand its availability to the research community in the second quarter of 2022.
According to the press release, Polaris will be used to help develop code for the upcoming Aurora:
“The system will accelerate transformative scientific exploration, such as advancing cancer treatment, exploring clean energy, and promoting particle collision research to discover new methods of physics. It will enable researchers to update their scientific workloads for Argonne’s upcoming exascale system Aurora, thus bringing ALCF into the era of exascale artificial intelligence. “
As we saw above (via Dylan Patel of Semi-analytical), Polaris is expected to consume approximately 2 megawatts of power during peak operation, which pales in comparison with Aurora’s 60 megawatts. Polaris will also be equipped with 560 nodes instead of the 9,000 nodes planned by Aurora. The system will also use Cray’s Slingshot network structure, which means that the system is produced by Hewlett Packard Enterprise (HPE). DoE has not specified its specific AMD Milan CPU for Polaris.
The delayed Aurora exascale system is another blow to Intel’s recent HPC work. In fact, Aurora originally planned to be a much smaller 180-PetaFLOP machine, which will make its debut in 2018, but Intel’s Xeon Phi processor was delayed and eventually cancelled. This led to the complete redesign of the Aurora system as an Exascale system powered by Intel’s Ponte Vecchio GPU and Sapphire Rapids GPU.In fact, Intel recently revealed that it developed a key aspect of it Old Bridge GPU, Xe Tile, especially in response to the Aurora requirements of the US Department of Energy. This work took a year of development work.
The U.S. Department of Energy has Confirm that the second revision of Aurora has now been postponed Starting from the initial 2021 release window, in view of Intel’s 7nm process node (Now renamed to Intel 4) For Sapphire Rapids processor.Therefore, Aurora now plans to deploy in the time frame from 2022 to 2023, giving up the title of the world’s first exascale supercomputer to AMD driven frontier.