Cray has landed a £48m deal to construct Blighty's 28-petaFLOPS Archer2 supercomputer, which will use second-generation AMD Epyc processors.
The contract was confirmed on Monday in an email from the British government's UK Research and Innovation (UKRI) organization to boffins.
The 750,000-ish-core supercomputer is significantly more powerful than its predecessor, the 2.6-petaFLOPS Cray XC30 Archer, which has 118,000 CPU cores inside racks of Intel Xeon E5 v2 processors. The UK's most powerful publicly known super, an eight-petaFLOPS 241,920-core Cray-Intel machine, meanwhile, is operated by the Meteorological Office.
All three are dwarfed by the world's fastest-known beasts, such as America's 201-petaFLOPS IBM-Nvidia Summit, and the upcoming generation of publicly known exascale monsters.
Archer2 will be housed at the University of Edinburgh in the same room used by Archer, and is due to be switched on by HPE-owned Cray in May 2020, three months after the predecessor is powered down for good. Although based in Scotland, Archer2 is expected to be made available to scientists and engineers across the UK.
The Archer series is aimed at running simulations of chemical reactions and materials, turbulence, protein folding, and other such tasks that require thousands upon thousands of processor cores crunching numbers in parallel.
Look at me. Look at me. I'm the El Capitan now: Cray to build US govt's $600m cray-cray exascale nuke app superREAD MORE
The new super will, we're told, feature 5,848 compute nodes, each packing two 64-core AMD Epyc Rome processors clocked at 2.2GHz, totaling 748,544 CPU cores and 1.57PB of system memory, all held in 23 liquid-cooled Shasta Mountain cabinets. As for storage, eggheads will have access to 14.5PB of Lustre-managed space across four file-systems, 1.1PB of all-flash Lustre BurstBuffer, disaster recovery provisions, and other bits and bytes.
Networking will arrive in the form of Cray's 100Gbps Slingshot technology. The system is designed to be a CPU-only beast, though it will include 16 AMD GPUs connected to four compute nodes just in case someone needs the acceleration.
"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," this week's UKRI memo reads.
"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.
"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."
You can find more details on the Archer2 project on our high-performance computing sister site The Next Platform, published earlier this year when the contract was put out for tender. ®