NASA today announced it has brought into operation its most powerful supercomputer ever - one that can tackle every one of its current needs both within the stratosphere and without. Only, their new supercomputer, Aitken, isn't all that new. But like any good enthusiast will tell you, spreading out upgrades - even between storied rivals such as Intel and AMD - is the way to go.
NASA has taken a different approach from most supercomputing clients, as The Register covers. Instead of ordering monolithic, self-contained, days-numbered systems such as the exascale-class Frontier, NASA takes a modular approach akin to recent European supercomputing designs such as LUMI. In addition, the networking of additional computing nodes ensures that you can partition the cost across budgets and extend the operational life you can squeeze from these systems.
"This sizeable enhancement — a 16% increase in performance since its previous expansion, and a 49% increase since last year, when the system was ranked at number 72 on the June 2021 Top500 list — translates to solving larger problems with faster results for important NASA research projects in aeronautics, space exploration, Earth science, and astrophysics," said Michelle Moyer with the NASA Ames Research Center.
NASA's approach has been in the making already - the company has been taking lessons from this modular approach for over fourteen years, as it has upgraded its "Pleiades" supercomputer since its initial power-on back in 2008.
And in a testament to how long AMD's High-Performance Computing (HPC) chops have come, the new addition - four HPE "Apollo" racks based on AMD's "Rome" architecture - have joined the system's Cascade Lake-based heart. But, of course, when supercomputers add another architecture to their portfolio, you know how far the tables have turned.
Aitken's added four racks might sound measly, but remember that compute density has been increasing at an astonishing pace. Those are enough to bring out the power on 308,000 AMD Zen 2, 7 nm-manufactured "Rome" cores distributed through 512 computing nodes. The system's initial design featured "only" 46,080 Intel Cascade Lake cores, also provided by HPE (1,152 40-core nodes distributed between 4 E-Cell systems).
Density is so high that the supercomputer fits in what could be any junior high's sports materials shed, taking up just one acre of land. And yet, the new addition pushes the system's theoretical peak performance to 13.12 PetaFlops per second, with 1.27 PetaBytes of memory supporting all that processing grunt.
Is this improvement enough to call it a Big Bang in NASA's supercomputing capabilities? You tell us.
