Following a procurement exercise, UK Research and Innovation (UKRI) are pleased to announce that Cray, an HPE company, has been awarded the contract to supply the hardware for the next national supercomputer, ARCHER2. Powered by AMD EPYC processors, ARCHER2 will be deployed at the University of Edinburgh.
Needless to say, ARCHER2 represents a significant step forwards in capability for the UK science community, with the system expected to sit among the fastest fully general purpose (CPU only) systems when it comes into service in May 2020.
ARCHER2 should be capable on average of over eleven times the science throughput of ARCHER, based on benchmarks which use five of the most heavily used codes on the current service. As with all new systems, the relative speedups over ARCHER vary by benchmark. The ARCHER2 science throughput codes used for the benchmarking evaluation are estimated to reach 8.7x for CP2K, 9.5x for OpenSBLI, 11.3x for CASTEP, 12.9x for GROMACS, and 18.0x for HadGEM3.
Hardware specifications:
- 28 PFLOP/s peak performance
- 5,848 compute nodes, each with dual AMD Rome 64 core CPUs at 2.2GHz, for 748,544 cores in total and 1.57 PBytes of total system memory
- 23x Shasta Mountain direct liquid cooled cabinets
- 14.5 PBytes of Lustre work storage in 4 file systems
- 1.1 PByte all-flash Lustre BurstBuffer file system
- 1+1 PByte home file system in Disaster Recovery configuration using NetApp FAS8200
- Cray next-generation Slingshot 100Gbps network in a diameter-three dragonfly topology, consisting of 46 compute groups, 1 I/O group and 1 Service group
- Shasta River racks for management and post processing
- Test and Development System (TDS) platform, to be installed in advance
- Collaboration platform with 4 x compute nodes attached to 16 x Next Generation AMD GPUs
Software stack:
- Cray Programming Environment including optimizing compilers and libraries for the AMD Rome CPU
- Cray Linux Environment optimized for the AMD CPU blade based on SLES 15
- Shasta Software Stack
- SLURM work load manager
- CrayPat as profiler
- GDB4HPC as debugger
Sign up for our insideHPC Newsletter