America's Argonne lab buys AI super from HPE, AMD, Nvidia while waiting for Intel
44-PFLOPS Polaris 'testbed' ordered ahead of Chipzilla-powered Aurora
Uncle Sam's Argonne National Laboratory has ordered Polaris – a supercomputer to be built from AMD Epyc processors and Nvidia GPUs – to test its applications in anticipation of eventually getting the Intel-delayed Aurora super.
Polaris will be capable of reaching 44 petaFLOPS of performance at double-precision or FP64, we're told, making it about as fast as the eighth most-powerful supercomputer in the latest top-500 list.
In terms of its AI prowess, it is said it can achieve a maximum performance of 1.4 exaFLOPs at mixed-precision or FP16.
It will be built using HPE’s Apollo Gen10 Plus architecture, and contain 560 second and third-generation AMD Epyc processors as well as 2,240 Nvidia A100 GPUs. Polaris is expected to be up and running early next year at Argonne, which is run by the US Dept of Energy.
The machine has been described as a “testbed supercomputer,” a stepping stone that will help Argonne “prepare critical workloads for future exascale systems”.
The lab is no doubt referring to the years-late Aurora supercomputer promised by Intel and HPE-owned Cray. Once planned to go live in 2018 with 180-PFLOPS of performance, Aurora's deployment stalled after its brains – Intel's esoteric Xeon Phi processors – turned out to be a dead end, forcing a rethink of its design.
Intel and Cray were thus given another crack at Aurora, albeit with a much higher performance target of 1,000 PFLOPS or one exFLOPS with FP64. This beast was supposed to be completed this year, but it's not going to arrive before 2022, as it will be filled with Intel's Sapphire Rapids Xeon Scalable x86 processors – due to land next year – and Xeon Xe GPUs, code-named Ponte Vecchio. And so you can see that while waiting for Aurora to be built, Argonne's scientists need a system to try out their code on, hence Polaris.
Our sister site The Next Platform has more detail on Aurora and Polaris right here.
"Argonne’s use of different technologies from multiple vendors allows researchers to prepare key application software that will run on the Intel-powered Aurora system," an Intel spokesperson told The Register on Wednesday.
"Intel continues to work closely with Argonne National Lab and the DOE to deliver the Aurora Exascale supercomputer and are committed to a 2022 delivery."
Polaris will be used to crunch data in a variety of science projects, ranging from modelling tumor cells to discover potential cancer treatments, and predicting biochemical reactions for fuel, to analyzing particle collisions in the search for dark matter.
- Singapore goes Cray-cray in the best way, picks HPE for new 10 PFLOPS super 'puter
- Tesla shows off the AI supercomputer training what it hopes will one day be an actual self-driving car
- IBM pulls up the ladder behind some supercomputer customers
- South Korea to build an exascale supercomputer running local CPUs by 2030
“Beyond getting us ready for Aurora, Polaris will further provide a platform to experiment with the integration of supercomputers and large-scale experiment facilities, like the Advanced Photon Source, making HPC available to more scientific communities,” said Michael Papka, director at the Argonne Leadership Computing Facility at the lab.
“Polaris will also provide a broader opportunity to help prototype and test the integration of HPC with real-time experiments and sensor networks.”
The Register has asked Argonne National Laboratory and HPE for further comment. ®