ISC23: Details of the Intel supercomputer Aurora, Falcon Shores finally only in 2025

Intel’s exascale flagship project Aurora also made the June 2023 list of the top 500 fastest supercomputers in the world. The Intel Xeon Datacenter CPU Max and the Datacenter GPU Max really do exist – at least Intel seems to want to defend itself against evil tongues and not even let rumors to the contrary arise.

The manufacturer showed both components in separate demos at the stand in the exhibition area at the International Supercomputer Conference 23 in Hamburg in selected tests, in which they were of course able to shine. There were also more precise specifications for the Aurora supercomputer itself.

Formerly known as the Ponte Vecchio, the accelerator that carries the brunt of Aurora’s computations is slated to be available in June from Dell and Lenovo in x4 OAMs, i.e. boards with four accelerator mounts. Supermicro and Inspur will bring the 8-socket version from July. This was confirmed by Jeff McVeigh, who took the stage at the presentation by Intel’s Corporate Vice President and General Manager of the Super Compute Group invited Ray Pang, Head of Technology & Business Enablement at Supermicro. He showed a flexibly configurable rack server with slots for up to twelve PCIe cards.



Intel’s availability roadmap for the rest of 2023.

(Image: Intel)

According to the plan, the Intel Data Center GPU Max 1100 will be available as a PCIe 5.0 plug-in card from August. It consumes 300 watts, has 56 Xe cores and 48 GB of HBM stack memory. The larger relatives 1450 and 1550 with up to 128 Xe cores, 128 GB of memory and 600 watts of power loss will only be offered via the Intel Developer Cloud at least this year.

So far, there has only been loose talk of more than 2 exaflops when it comes to the Aurora specification. Intel has now announced further details. More than 60,000 blades are said to have already been delivered and installed, of which Aurora will receive a total of 10,624. With two processors and six accelerators, that’s 21,248 CPUs and 63,744 GPUs.

The system will have 10.9 petabytes of DDR5 memory at 5.95 PB/s and an additional 1.36 PB of HBM stack memory (30.5 PB/s) in the Xeon Max CPUs and 8.16 PB of HBM memory in the GPUs (208.9 PB/s). For mass storage, Aurora has 1024 storage nodes, whose 230 Pbytes have a total transfer rate of 31 Tbytes/s.



Aurora supercomputer specifications as of May 2023.

(Image: Intel)

Jeff McVeigh confirmed that Intel’s former prestige project codenamed Falcon Shores is now finally a “2025 Product”. “Falcon Shores is a new architecture that brings together x86 cores and an Xe GPU in one processor socket. This architecture is planned for 2024 and is expected to deliver more than five times the performance per watt, five times the compute density and five times the memory capacity as well as bandwidth” Intel wrote in its press release at the time.

According to McVeigh, the reason for the shift lies in the enormous impact that so-called generative AI has had. This “ChatGPT moment” made Intel reconsider whether it was really the right time to integrate the various types of arithmetic units that were still missing, especially those for AI. Intel would thus commit itself to hardware, while the workloads, i.e. the tasks to be expected, are currently subject to rapid change.

Another reason, according to McVeigh, is that Intel still sees some potential in Ponte Vecchio. However, the performance data shown mostly related to the Data Center GPU Max 1550 with its 600 watt TDP, which, for example, beat one of Nvidia’s current PCIe cards H100 with 350 watts in a series of benchmarks by a factor of 1.2 (QMCPack) to 2, 1 (autodock) beats. Should we wait two more years with a successor?


More from c't magazine

More from c't magazine


More from c't magazine

More from c't magazine


(csp)

To home page

source site