Prometheus with its 2,65 PFlops of theoretical performance has been built by Hewlett-Packard according to requirements and partial design provided by Cyfronet.
Prometheus consists of more than 2,200 servers based on HP Apollo8000 platform, combined with super-fast InfiniBand network with 56 Gbit/s capacity. Its energy saving and high-performance Intel Haswell latest-generation processors offer 53,748 cores. These are accompanied by 283.5 TB DDR4 RAM in total, and by two storage file systems of 10 PB total capacity, and 180 GB/s access speed. Prometheus has been also equipped with NVIDIA Tesla GPGPUs.
The Prometheus architecture responds to the diverse needs of researchers by providing computing resources organized in partitions:
- classical cluster of computing servers with highly efficient CPU nodes equipped with Intel Xeon Haswell and Intel Xeon Gold processors,
- cluster of servers equipped with graphic accelerators GPGPU NVIDIA Tesla K40 XL,
- acceleration partition including GPGPU NVIDIA K80 cards as well as Intel Xeon Phi 7120P and Nallatech FPGA accelerators,
- partition dedicated to calculations related to artificial intelligence, equipped with GPGPU NVIDIA Tesla V100 graphics accelerators. It is worth mentioning that this partition is a system with computing power over 4 PFlops for tensor operations and 256 TFlops for standard calculations performed on double precision numbers, which makes it the fastest dedicated solution for artificial intelligence available for the needs of scientists in Poland.
Prometheus has been installed in a high-tech computing room, exclusively adapted for its operation. The supercomputer's proper functioning is additionally supported by the accompanying infrastructure, including such systems as guaranteed power supply with an additional generator, modern air-conditioning and gas extinguishing.