HPE is building the world's largest Arm-based supercomputer, Astra – 2.5 petaFLOPS from 2,592 HPE Apollo 70s – for Sandia National Labs in the US, where it will run advanced modeling and simulation workloads in areas including national security and energy.
Apollo 70s have four compute nodes in a 2U case. The CPUs are Cavium Arm v8-A 64-bit Thunder X2 with up to 32 cores and 8 x DDR4 channels.
Astra involves 2,592 dual-CPU servers with more than 145,000 cores. Liquid cooling will keep these hot little cores in working order.
- Message Passing Interface (MPI)
- MCS30 liquid cooling unit
- Performance Cluster Manager software
- Apollo 4520 all-flash storage with a Lustre filesystem
Ironically, the 4520s are dual x86 (E5-2600 v4 Series) server systems. It seems Arm CPUs can only take you so far. They have 23 large form factor (3.5-inch) drive bays, designed back in the bulk capacity disk era.
HPE has said it is a major stepping stone to exascale. That's interesting timing when IBM has just delivered the 200 petaFLOPS Summit system – which ostensibly would need a five-times scale-up to reach the fabled 1,000 petaFLOPS exascale level.
Might we assume Arm CPUs could feature in an HPE exascale system? It's not a fantastical idea. The Fujitsu Post-K exascale computing project in Japan is based on Arm v8 processors with extensions, making up scalable custom CPU cores. The total Post-K node count is estimated to be more than 10,000.
El Reg thinks HPE could be hinting that Arm CPUs will feature in its exascale computing development work. ®