Forward-looking: Earlier this month, it was reported that the US once again owns the world's fastest supercomputer---the 200-petaflop Summit. While the machine currrently being developed by HPE won't have that kind of performance, it will boast one record: it'll be the world's largest ARM-based supercomputer.
HPE has announced a collaboration with Sandia National Laboratories and the US Department of Energy to deliver the ARM-based system, called Astra. It will be part of the DoE National Nuclear Security Administration's Vanguard project, used to run advanced modeling and simulation workloads for areas such as national security, energy and science.
Astra will be based on HPE's Apollo 70 system and be made up of 145,000 cores across 2,592 dual-socket nodes. Each node will be equipped with two 28-core Cavium ThunderX2 processors, which offer eight memory channels, running at 2.0 GHz. This allows the system to be more power efficient and denser than traditional market offerings, and brings 33% better memory performance.
Once it's up and running, Astra will be able to reach a peak performance of 2.3 petaflops, putting it into the upper fifth of the top 500 supercomputers list.
Every processor in Astra has access to a giant shared pool of memory, as opposed to typical systems where relatively small amounts of memory are tethered to each processor.
"For one processor to access data not held in its own memory, the computer must play an inefficient game of "Mother May I," so to speak. One processor must request access from another processor to get anything accomplished. What's worse, the relationship between storage and memory is also inefficient. In fact, in today's computers it's estimated that 90 percent of work is devoted to moving information between tiers of memory and storage," writes HPE.
"...we think that should change. We want to give all the processors equal access to a big pool of shared memory, thereby eliminating all the back and forth, and we want to build this pool out of new types of non-volatile memory."
HPE is also providing a power-efficient cooling system called MCS-300, which improves operational environment for HPC systems while being cheaper to run compared to air-cooled options. There's also an optimized storage solution that offers a high performing, all-flash Lustre filesystem for extreme bandwidth and reduced latency on data access.