COMPUTE ATLAS Supercomputer supply-chain graph
1153 systems 482 sites Sourced data

Blog/2026-10-07

· Compute Atlas

Isambard-AI versus Alps: two GH200 routes

Bristol's Isambard-AI has 5,448 GH200 chips (168 in phase 1, 5,280 in phase 2) in a 5 MW modular site; CSCS's Alps has 10,752 in Lugano. Alps and phase 1 both score near 60 GFlops per watt in HPL, but the facilities differ.

Isambard-AI in Bristol and Alps at CSCS in Lugano are two large public Grace Hopper machines in Europe, and they are built on almost the same silicon: NVIDIA GH200 superchips in HPE Cray EX cabinets with Slingshot-11 networking. Alps has about twice as many chips (10,752 against 5,448) and about twice the measured Linpack result (434.9 against 216.5 petaflops). What differs is the building around them, the way each was funded and the people each is meant to serve. What is not publicly known for either is a measured, audited power figure for the production system; the numbers we have are a mix of vendor design values, a site grid supply and one benchmark run each. This article compares what the operators and the TOP500 list state, and flags where sources disagree. Our entries are isambard-ai (phase 1), isambard-ai-phase2 and alps.

How Isambard-AI came to exist

Isambard-AI was funded as part of the UK’s AI Research Resource (AIRR). NVIDIA’s launch-day blog of 17 July 2025 gives the government investment as 225 million pounds and says the Secretary of State for Science, Innovation and Technology cut the ribbon at the Bristol Centre for Supercomputing. The University of Bristol’s June 2026 account says investment was confirmed in late 2023 through the Department for Science, Innovation and Technology (DSIT) and UK Research and Innovation under AIRR. The same page separates a second pot: the Sovereign AI programme, which it says was funded with 500 million pounds from the UK government and gives portfolio companies dedicated access to the machine. We read these as capital for the system (225 million pounds) and a later programme that pays for access, not a larger machine budget, but the Bristol page does not draw that line explicitly.

The speed is the headline claim. Bristol’s ISC 2025 post says the system was built in under 18 months, against three to five years for a typical funded supercomputer, helped by early supplier involvement, concurrent project streams, a containerised datacentre and pre-integrated hardware. NVIDIA’s blog says roughly two years from conception to deployment. Those two statements are consistent (18 months is the build, two years is concept to operation), but the endpoint varies by source: the 2025 Bristol post says phase 2 would reach full capacity in summer 2025, NVIDIA says fully operational by June 2025, and the 2026 Bristol timeline says acceptance passed in mid 2025 with real users from August. ITPro’s report adds an inauguration date of 17 July 2025 and a contracted life of five years, possibly stretching to six or seven.

Phase 1, phase 2 and the counts

The two phases are the same node design at very different scale. The Isambard documentation lists phase 1 as 42 nodes with 168 GH200 superchips and phase 2 as 1,320 nodes with 5,280, each node holding four superchips, each superchip capped at 660 watts. Together that is 5,448 superchips in 1,362 nodes, which matches Bristol’s own total. Each GH200 pairs a 72-core Grace CPU (120 GB LPDDR5 on Bristol’s account) with an H100 GPU carrying 96 GB of HBM3, and a node presents about 864 GB of memory in one address space. Four 200 Gbps Slingshot 11 links connect each node, and storage is all flash: 20 PiB of Cray ClusterStor and 3.5 PiB of VAST, per the 2025 Bristol post.

Phase 1 arrived first, in summer 2024, and let early users test the platform. By the 2025 post, almost 60 projects had used about 100,000 GPU hours on it. On the TOP500, phase 1 appears with 7.42 petaflops Rmax and a listed power of 117.08 kW (82.40 kW in an optimised run), first at rank 128 in June 2024. Phase 2 is a separate submission at 216.5 petaflops Rmax and 278.58 Rpeak, ranked 11th in November 2025 and 13th in June 2026. The TOP500 page lists no power for phase 2, which is why our phase 2 row carries no HPL power. The AI figure that Bristol and the preprint by the system’s designers quote, over 21 exaflops, is an 8-bit peak. The 64-bit figure over 200 petaflops is consistent with the HPL result, so the usable headline for traditional simulation is the TOP500 number, not the 21.

Cooling, power and what PUE says

Isambard-AI is housed in a modular data centre, not a building. Bristol describes what looks like four large shipping containers joined together, based on HPE’s MODPOD design, with no offices; ITPro says the build used six components assembled in HPE’s Czech facility. We cannot reconcile four against six from these pages and report both. Bristol credits the modular approach with about 72 percent lower construction emissions than a conventional data centre (a Bristol claim without a stated baseline in the page).

Cooling is direct liquid cooling: a coolant and water mix flows in a closed loop through the blades and is rejected outside. NVIDIA and ITPro both describe hybrid cooling towers running more than 90 percent dry, which means little evaporative water use. Bristol gives PUE as “around 1.08”, NVIDIA says below 1.1 and ITPro 1.1. None says whether the value is a design target, a measured annual mean or a snapshot at load. The site grid supply is 5 MW per the 2025 Bristol post, ITPro and the preprint (which says “under 5MW”). Bristol says the power comes from renewable UK sources. Heat reuse is a plan, not a reported operation: Bristol says the liquid design “unlocks the pathway” to district heating, and ITPro says waste heat will warm the host National Composites Centre and may later go to local businesses and homes.

Efficiency rankings also differ by source. Bristol says phase 1 entered the May 2024 Green500 at number 2, a statement that matches the Green500 pointer our data records for that list; ITPro cites fourth place for the full system. They are probably different submissions on different lists, which we have not resolved.

Alps in Lugano

Alps is a different kind of project. CSCS’s Alps page describes a general-purpose research infrastructure that is geo-distributed (Lugano, Lausanne, a Paul Scherrer Institute archive and ECMWF in Bologna) and built from virtual clusters, with separate ones for MeteoSwiss forecasting, the user lab and machine learning. The Grace Hopper partition is 2,688 nodes with four sockets each, 10,752 GH200 in all, installed stepwise between January and June 2024 and formally inaugurated in September 2024. The same page lists other partitions that Isambard-AI has no counterpart to: 1,024 AMD EPYC Rome nodes, 144 NVIDIA A100 nodes, 128 AMD MI300A nodes and 24 MI250X nodes, plus 100 PB of hard-disk scratch and tape archives. The page cites 434.9 petaflops sustained and 10 exaflops BF16 for the whole infrastructure.

The timeline is a useful contrast with Bristol’s. The installation history on the same page begins in October 2020 with the first 1,024 AMD nodes, records the April 2021 announcement with HPE and NVIDIA of what it called the world’s most powerful AI-capable supercomputer, and reaches the Grace Hopper nodes only in the first half of 2024. Alps was therefore a staged, multi-year build in which the GH200 partition arrived after the original CPU nodes and several storage upgrades; Isambard-AI was a single funded build delivered in about two years. The staged route meant Alps already had a CPU partition and a host organisation with an established user community when the AI partition arrived, while Isambard-AI opened a phase 1 pilot to early users in 2024 and its full user base only from August 2025.

The June 2026 TOP500 list shows Alps at rank 10 with 434.90 petaflops Rmax, 574.84 Rpeak, 2,121,600 cores and 7,124 kW. Divide Rmax by that power and the benchmark run delivered about 61 gigaflops per watt (our arithmetic). Doing the same for Isambard-AI phase 1 (7.42 petaflops at 117.08 kW) gives about 63. The two designs are therefore close to level on benchmark efficiency at the compute-rack level, which is expected for two builds of the same chip and cabinet.

CSCS’s facility is where Alps differs from Bristol. Its 2022 efficiency article says the Lugano data centre, opened in 2012, runs on lake water pumped from Lake Lugano, with a first cooling circuit designed for up to 14 MW of high-performance machines and a second for 7 MW of smaller systems, and a PUE below 1.2. It also says CSCS buys 100 percent hydropower, used about 37 gigawatt hours in 2021 (about 4 MW average), and that about 35 percent of the centre’s demand came from partner systems. Heat goes to the CSCS building and, with the local utility, to the city of Lugano and a university campus. Two caveats: that PUE is a 2022 statement for the whole centre, so it is a bound for Alps and not a measurement of it, and the article predates Alps entering production.

The practical comparison on cooling is a Swiss facility built for 21 MW of mixed machines and lake water against a British 5 MW purpose-built enclosure with cooling towers. Bristol’s PUE claim (1.08) is lower than CSCS’s stated bound (below 1.2), but they describe different boundaries and dates, so the ordering is not established.

Purpose and access

Both machines are public, but the access models differ. Isambard-AI allocates time through AIRR calls, prioritising scientific, economic or societal impact, and through Sovereign AI to companies funded by the programme. The Bristol article names UK-language model training, cardiomyopathy protein modelling and pollution mapping among uses. Alps serves weather forecasting, the national user lab and ML. Its most visible AI output is Apertus: ETH Zurich’s release of 2 September 2025 says CSCS invested over 10 million GPU hours on Alps to train the 8 billion and 70 billion parameter models, which the technical report says used 15 trillion tokens. Neither the release nor the abstract says how many GPUs ran the largest job at once.

What we could not confirm

We could not confirm: a measured PUE for Isambard-AI or Alps in operation; a power figure for Isambard-AI phase 2 from TOP500; the funding behind Alps (we found no figure on the pages fetched); how many of Alps’ 10,752 superchips are assigned to AI versus weather and user-lab partitions; what fraction of Isambard-AI’s AIRR allocation goes to industry; and the UK government’s original 300 million pound AIRR total, which we saw only in a search summary and did not fetch. We also did not verify actual utilisation of either machine.

What to watch

Three things would settle the open points. A published annual energy and water report from Bristol would turn PUE from a claim into a number. A phase 2 HPL submission with a measured power value would let the two sites be compared like for like on the Green500. And any follow-on funding for Isambard-AI or a successor, which the UK’s Sovereign AI spending suggests is coming, will show whether the modular, quickly delivered model is repeated or whether the next machine moves into a permanent building. For Alps, the next test is whether its multi-tenant design lets one national resource serve forecasting, open science and open-model training without the compute-hungry workload crowding out the others.

Sources

deep-diveai-clusterspower-and-coolingprocurementsovereignty