UK's Arm-based Isambard 2 supercomputer powers off for good

Isambard 3 and AI sibling set to pick up the torch

The UK's Isambard 2, one of the early Arm-based supercomputers, has officially retired after just a few years of operation. It is superseded by the more powerful Isambard 3 and Isambard-AI, just as British supercomputing enters an uncertain period for funding.

Isambard 2 ended service at 9am on September 30, but started winding down days earlier to drain the system queues before final switch-off.

The move was heralded by Simon McIntosh-Smith, professor of High Performance Computing at the University of Bristol, saying: "Isambard 3 takes over today using Nvidia Grace Arm-based CPUs provided by HPE."

According to McIntosh-Smith, Isambard 2 is being retired after six years of service, but he appeared to be referring to the date that the first incarnation of Isambard went live.

He told The Register: "Isambard 3 early access use began last week and will ramp up over the coming months. Once the system has operated successfully for a few months, we'll pass acceptance and go into full production use."

That system was unveiled back in 2017 as the world's first Arm-based production supercomputer. It was based on the Cray XC50 architecture and powered by 10,496 ThunderX2 cores.

Based on the success of that system, the GW4 Alliance – comprising the universities of Bath, Bristol, Exeter, and Cardiff – was able to secure funding from the Engineering and Physical Sciences Research Council (EPSRC) in February 2020 to create Isambard 2, touted at the time as the largest Arm-based supercomputer in Europe.

The professor gave The Reg some impressive Isambard 2 stats, noting that "since May 2018" the university had tracked:

Total core Hours of all jobs: 640M
Total number of jobs executed = 321k
Largest job = 20,992 cores (basically the whole machine)
Average size of each job = 345 cores (each node is 64 cores, being dual socket ThunderX2, so this is 5.4 nodes on average)

He added: "The 640M core hours impressed me for a modestly sized machine!"

Isambard 2 was built on the existing clusters, but more than doubled the number of cores to 21,504 across 336 nodes, and also incorporated a partition of 72 nodes powered by Fujitsu's A64FX CPUs – the same chips that power the Fugaku supercomputer, formerly the most powerful system in the world.

Its successor, Isambard 3, was announced last year and swaps the ThunderX2 cores for Nvidia's Grace superchip, described as the first Arm-based server CPUs that are optimized specifically for HPC from a top-tier silicon vendor.

Installed in a purpose-built facility at the Bristol & Bath Science Park, Isambard 3 features 384 Nvidia Grace chips, each of which includes 144 cores, for a grand total of 55,296 compute cores and capable of 2.7 petaFLOPS of double precision 64-bit floating point performance while consuming less than 270 kilowatts of power, according to Nvidia.

The UK government offered up £225 million ($299 million) later in 2023 for the building of a separate AI supercomputer, known as Isambard-AI. This is based on Nvidia's Grace-Hopper chips, which combine an Arm CPU with the company's GPU silicon.

As detailed by our partner site The Next Platform, the GW4 Alliance was essentially given carte blanche to build the biggest machine it could, with the main limiting factor being the available power at the site, which allowed for Isambard-AI to comprise 5,448 Nvidia GH200 Grace-Hopper chips, making it the most powerful computer in the UK.

Phase 1 of Isambard-AI deployment was completed in May, comprising just 168 of its full complement of those Grace-Hopper superchips. However, even this was enough to have the system reach position 128 in the TOP500 list of the high-performance systems, as published at the ISC High Performance 2024 event.

However, the rest of the UK high performance compute arena isn't going quite so swimmingly. In August, it was announced that an Edinburgh-based exascale computer planned by the previous government has been canceled by the incoming administration.

The same announcement saw £500 million ($664 million) extra funding earmarked for the AI Research Resource program dropped, which has a knock-on effect on another supercomputing program, the Dawn system being built by Intel, Dell, and the University of Cambridge.

Dawn, announced last November, was being built in two phases. The first phase is complete and judged to be about on a par with the UK's previous fastest supercomputer, Archer2, currently ranked 39th in the Top500, as The Register detailed last year. Phase 2 was anticipated to boost performance by 10x, but it is unclear whether this will now be funded.

According to the Financial Times, one senior government figure claimed that the Edinburgh exascale project "made little strategic sense" as it was not focused on AI.  

This kind of thinking is difficult to comprehend when researchers in the US have access to two exascale systems, the EU is readying one, and China probably has several already. Perhaps, as The Next Platform hinted, UK scientists will have to hope that an upgrade will result in Isambard 4 turning out to be the nation's first true exascale machine. ®

More about

TIP US OFF

Send us news


Other stories you might like