Supercomputer Aurora: With the exception of Intel's chips, the exascale system is now being built

0
348

For Vision 2022, Intel also gave another look at the Aurora supercomputer. Progress can be seen, but one thing is missing: Intel. The first test modules are now being delivered, but it is actually no longer possible to use the system productively this year. This is also the aim of the application phase.

The almost endless delay of the supercomputer Aurora will not be eliminated for Vision 2022, as Intel's in-house exhibition is called in the spring of this year. Intel made that clear on stage by explaining that the infrastructure including cooling, the DOAS storage system and the management nodes were in place. But the crucial elements are still missing: the compute nodes.

Intel confirms this with the image description of the system: Already installed are the DAOS storage system, management nodes, and cooling infrastructure. There is still no trace of the compute nodes, each with two Sapphire Rapids processors and six Ponte Vecchio chips. There will be thousands of nodes in total. Intel explained again that this requires both 20,000 Sapphire Rapids chips in the HBM2e version and 60,000 Ponte Vecchio accelerators.

Argonne installation (Image: Intel)

As part of the keynote speech on Vision 2022, Raja Koduri shared the stage with the head of the Argonne Lab, where the supercomputer will be located. He explained that Aurora is only now being built (from 1:21:00). The first validation systems have now finally been delivered by Intel, it is said. Now the final interaction can be tested.

A blade in the Aurora supercomputer (Image: Intel)

This aligns with what other senior Intel employees have said. In the quarterly report two weeks ago, Intel emphasized that the delivery of Sapphire Rapids had started. The head of the data center division, Sandra Riviera, substantiated this with the same statement and the appropriate wafer yesterday on stage. This involves the delivery of initial SKUs, i.e. the first chips that traditionally go to lighthouse projects – something like Aurora. The real product launch will then happen sometime in the coming weeks and months. The fact that Sapphire Rapids is undoubtedly late has already become clear in recent months and years. According to the original plan, Sapphire Rapids will end up being around 1.5 years late.

Researchers can now apply to adapt their software for the supercomputer during the test phase or to develop completely new ones, so that the productive work can begin immediately after the official starting signal. On paper, the launch will probably still be held in 2022 in order to be able to celebrate at least a small success after endless delays. But the system will probably not be accessible to “everyone” until 2023.