Cray to build next-gen ARCHER2 UK supercomputer featuring 11,696 AMD Epyc Rome CPUs

New AMD Epyc-based supercomputer expected to become operational from 6th May 2020

The UK Research and Innovation Centre has awarded the contract to build the ARCHER2 national supercomputer to Cray.

The next-gen ARCHER2 supercomputer will replace the existing Archer supercomputer deployed at the University of Edinburgh.

According to the Centre, the current ARCHER supercomputer will be shut down on 18th February 2020, and ARCHER2 will become operational 78 days later from 6th May 2020.

The new supercomputer will be built in the same room as the out-going ARCHER system, which means there will be a downtime of between two and three months for the users of the supercomputer. The new system will also be stress tested for about 30 days, so potential user will have to wait for one more month before the computing system is ready for access.

The Archer supercomputer system was unveiled in 2013, and is being used since then by researchers who need huge computational power for their research work. Based on Cray XC30 design, this supercomputer was once among the world's top 20 supercomputers.

ARCHER's successor ARCHER2 is expected to be the world's fifth most powerful supercomputer and also the most powerful CPU-only supercomputer. It will be equipped with 11,696 AMD Epyc Rome CPUs that will enable it to deliver approximately 1.5 million processing threads.

ARCHER2 will be fitted into 23 Shasta Mountain cabinets, with 5,848 compute nodes inside. Each of these 5,848 nodes will have two AMD EPYC Rome 64C/128T CPUs, with each clocked at 2.2GHz.

The system will also have a 1.1PB Lustre BurstBuffer file system, a 14.5PB of Lustre work storage and a 1+1PB backup/recovery file system.

If we talk about software, ARCHER2 will come with Cray Programming Environment and specific optimising libraries and compilers for AMD's EPYC Rome.

The ARCHER2 is expected to perform 8.7x for CP2K, 9.5x for OpenSBLI, 11.3x for CASTEP, 12.9x for GROMACS, and 18.0x for HadGEM3. The peak performance of the new supercomputer is estimated at ~28PFLOP/s.