{ "id": "isambard-ai", "url": "/supercomputers/isambard-ai", "name": "Isambard-AI", "site": "University of Bristol (Bristol Centre for Supercomputing)", "country": "GB", "compute_atlas_url": "https://computeatlas.net/supercomputers/systems/isambard-ai", "sourcing": "University of Bristol's Isambard-AI announcement, 2026 'Inside Isambard-AI' article and Cray User Group 2024 paper (arXiv 2410.11199), the BriCS system-specification documentation, Nvidia architecture write-ups and the TOP500 entry, researched 2026-10-03", "reading_note": "Unmarked companies come from the operator's or a vendor's own pages (bristol.ac.uk, docs.isambard.ac.uk, arxiv.org paper by Bristol authors, nvidia.com, top500.org); reported companies come from named trade press or analyst reporting whose article body was read in full, not just the headline.", "specs": { "accelerator": "NVIDIA GH200 Grace Hopper Superchip (Hopper H100-class GPU, 96 GB HBM3 per GPU); 4 per node, 5,448 in total", "host_cpu": "NVIDIA Grace (72 Arm Neoverse V2 cores) in each superchip, 288 cores per node", "integrator": "HPE (HPE Cray EX4000 and EX2500 liquid-cooled cabinets; TOP500 lists the system as HPE Cray EX254n)", "nodes": "1,362 nodes: 42 in Phase 1 (one HPE Cray EX2500 cabinet) plus 1,320 in Phase 2 (12 HPE Cray EX4000 cabinets)", "memory": "Per node: 4 x 120 GB LPDDR5X (460 GB usable) plus 4 x 96 GB HBM3 (384 GB)", "interconnect": "HPE Slingshot 11, 4 x 200 Gb/s Cassini NICs per node (800 Gb/s injection)", "storage": "All-flash HPE Cray ClusterStor E1000 Lustre, about 20.3 PiB usable (44 flash servers, 1,056 x 30.72 TB NVMe SSDs); VAST Data Platform flash system, 3.56 PB native usable", "cooling": "Direct liquid cooling in HPE Cray EX cabinets (coolant-and-water closed loop through coolant distribution units, cold plates on compute and network blades), inside HPE modular data centres; PUE around 1.08", "power_mw": "Under 5 MW design limit for the full system", "rmax_pflops": "216.50 PFlop/s (TOP500 June 2026 list)", "rpeak_pflops": "278.58 PFlop/s (TOP500 June 2026 list)", "status_note": "Phase 1 (42 nodes, 168 GH200) installed March 2024 in an HPE modular data centre; Phase 2 (1,320 nodes, 5,280 GH200) installed early 2025; full system launched July 2025 and delivered through BriCS for the UK AI Research Resource; TOP500 lists the full system as 'Isambard-AI phase 2' (installation year 2025)" }, "not_confirmed": "The Bristol sources name HPE, Nvidia and VAST Data as suppliers but do not itemise the parts inside their products. No source read names the maker of the HBM3 stacks or LPDDR5X DRAM on the GH200 modules, the 30.72 TB NVMe drives in the ClusterStor Lustre system, the 3.84 TB node-local SSDs, the cabinet cold plates, quick disconnects and rectifiers, the Slingshot cables and optics, or the hardware inside the VAST enclosures. The modular data centre is HPE's own POD design (the Phase 1 module is described as a DC20), but the manufacturers of its chillers, pump skids, UPS, transformers and DCIM software are not named; the paper's acknowledgements thank \"Contour\" and the building contractor \"Oakland\" without saying what either supplied, so neither is placed. Facility figures that have no named supplier are given here instead of as lines. The Phase 1 DC20 heat-removal plant was sized for 711 kW with two free-cooling chillers; the whole site was designed within a 5 MW power limit and about 550 m2 of car park at the National Composites Centre; the 2026 Bristol article describes the data centre as four large containers joined together and gives a power usage effectiveness of around 1.08. Waste-heat reuse is designed for but is described as a future intention, so no heat-reuse line is made. The TOP500 entry carries no Linpack power figure, so the 5 MW number is a design limit, not a measurement.", "lines": [ { "n": 1, "line": "System design, integration and cabinets (HPE Cray EX4000 for Phase 2, HPE Cray EX2500 for Phase 1, direct liquid cooled)", "parts": [ { "id": "4.01", "name": "Rack-scale systems", "url": "/parts/servers-racks-and-integration/rack-scale-systems-nvl72-class-mi400-helios-class" } ], "companies": [ { "id": "hpe", "name": "HPE", "url": "/companies/hpe", "reported": false, "note": "system integrator and cabinet designer; HPE was named as the builder in the 2023 award and the contract was placed in October 2023" } ], "unresolved": [], "sources": [ "https://www.bristol.ac.uk/news/2023/november/supercomputer-announcement.html", "https://arxiv.org/pdf/2410.11199" ] }, { "n": 2, "line": "Compute node (4x GH200 superchips with 4x Slingshot NICs; 42 Phase 1 nodes plus 1,320 Phase 2 nodes, 1,362 in total)", "parts": [ { "id": "4.03", "name": "GPU baseboards & compute trays", "url": "/parts/servers-racks-and-integration/gpu-baseboards-and-compute-trays-hgx-bianca" } ], "companies": [ { "id": "hpe", "name": "HPE", "url": "/companies/hpe", "reported": false, "note": "HPE Cray EX compute blades; TOP500 lists the system as HPE Cray EX254n" }, { "id": "nvidia", "name": "Nvidia", "url": "/companies/nvidia", "reported": false, "note": "GH200 superchips on each node" } ], "unresolved": [ "maker of the 3.84 TB node-local SSDs fitted in a subset of nodes (not named)" ], "sources": [ "https://docs.isambard.ac.uk/specs/", "https://www.bristol.ac.uk/research/centres/bristol-supercomputing/articles/2026/inside-isambard-ai.html", "https://arxiv.org/pdf/2410.11199", "https://top500.org/system/180388/" ] }, { "n": 3, "line": "GH200 Grace Hopper Superchip (Hopper GPU plus 72-core Grace CPU joined by NVLink-C2C, 5,448 in the full system)", "parts": [ { "id": "1.01", "name": "Merchant GPUs / accelerators", "url": "/parts/compute-silicon-and-packaging/merchant-gpus-accelerators" }, { "id": "1.05", "name": "Server CPUs", "url": "/parts/compute-silicon-and-packaging/server-cpus" } ], "companies": [ { "id": "nvidia", "name": "Nvidia", "url": "/companies/nvidia", "reported": false, "note": null }, { "id": "arm", "name": "Arm", "url": "/companies/arm", "reported": false, "note": "Neoverse V2 core design used in the Grace CPU" } ], "unresolved": [], "sources": [ "https://docs.isambard.ac.uk/specs/", "https://developer.nvidia.com/blog/nvidia-grace-hopper-superchip-architecture-in-depth/", "https://www.bristol.ac.uk/research/centres/bristol-supercomputing/articles/2026/inside-isambard-ai.html" ] }, { "n": 4, "line": "Hopper GPU die fabrication (TSMC 4N process)", "parts": [ { "id": "1.09", "name": "Leading-edge foundry", "url": "/parts/compute-silicon-and-packaging/leading-edge-foundry" } ], "companies": [ { "id": "tsmc", "name": "TSMC", "url": "/companies/tsmc", "reported": false, "note": "fabricates the GH100 GPU die on its 4N process, per Nvidia's Hopper architecture write-up; the Isambard pages do not name the foundry" } ], "unresolved": [ "advanced-packaging flow and substrate suppliers for the GH200 module (none named in any source read)" ], "sources": [ "https://developer.nvidia.com/blog/nvidia-hopper-architecture-in-depth/" ] }, { "n": 5, "line": "GPU memory (96 GB HBM3 per Hopper GPU, 384 GB per node)", "parts": [ { "id": "2.01", "name": "HBM", "url": "/parts/memory-and-storage/hbm-hbm3e-hbm4" } ], "companies": [ { "id": "sk-hynix", "name": "SK Hynix", "url": "/companies/sk-hynix", "reported": true, "note": "reported as the primary HBM3 supplier for Nvidia's H100-generation GPUs; no source read confirms it for the GH200 units installed at Bristol" } ], "unresolved": [ "which HBM3 vendors supplied the stacks on Bristol's GH200 modules (the Bristol pages and Nvidia's write-ups do not say; TrendForce also notes Samsung entered Nvidia's HBM3 supply in late 2023)" ], "sources": [ "https://www.trendforce.com/presscenter/news/20240313-12075.html", "https://docs.isambard.ac.uk/specs/" ] }, { "n": 6, "line": "System memory (4x 120 GB LPDDR5X per node, 460 GB usable, on the Grace Hopper modules)", "parts": [ { "id": "2.02", "name": "DRAM", "url": "/parts/memory-and-storage/dram-ddr5-rdimm-mrdimm-lpddr5x-for-grace" } ], "companies": [ { "id": "nvidia", "name": "Nvidia", "url": "/companies/nvidia", "reported": false, "note": "the LPDDR5X sits on the Grace Hopper module, per Nvidia's architecture write-up" } ], "unresolved": [ "DRAM vendor for the LPDDR5X (not named by Bristol, HPE or Nvidia in anything read)" ], "sources": [ "https://docs.isambard.ac.uk/specs/", "https://developer.nvidia.com/blog/nvidia-grace-hopper-superchip-architecture-in-depth/" ] }, { "n": 7, "line": "Slingshot network interface (HPE Cassini NIC, 200 Gb/s, one per superchip, 4 per node)", "parts": [ { "id": "1.06", "name": "NICs / SuperNICs", "url": "/parts/compute-silicon-and-packaging/nics-supernics" } ], "companies": [ { "id": "hpe", "name": "HPE", "url": "/companies/hpe", "reported": false, "note": "Slingshot Cassini NICs, cooled by the same cold plates as the compute blades" } ], "unresolved": [], "sources": [ "https://docs.isambard.ac.uk/specs/", "https://arxiv.org/pdf/2410.11199" ] }, { "n": 8, "line": "Slingshot 11 switch blades and fabric (high-performance networking switch blades inside the HPE Cray EX cabinets)", "parts": [ { "id": "3.03", "name": "Switch systems & white-box ODMs", "url": "/parts/networking-and-interconnect/switch-systems-and-white-box-odms" } ], "companies": [ { "id": "hpe", "name": "HPE", "url": "/companies/hpe", "reported": false, "note": "designs and supplies the Slingshot switches; the Bristol 2023 release says the networking was \"co-developed at HPE's Bristol labs\"" } ], "unresolved": [ "copper and optical cable vendors for the fabric (not named)", "fabric topology (not stated in any source read)" ], "sources": [ "https://arxiv.org/pdf/2410.11199", "https://www.bristol.ac.uk/news/2023/november/supercomputer-announcement.html" ] }, { "n": 9, "line": "Flash parallel file system (HPE Cray ClusterStor E1000 running Lustre; about 20.3 PiB usable, 44 flash servers, 1,056 x 30.72 TB NVMe SSDs)", "parts": [ { "id": "2.10", "name": "AI storage systems", "url": "/parts/memory-and-storage/ai-storage-systems-parallel-file-systems-object" } ], "companies": [ { "id": "hpe", "name": "HPE", "url": "/companies/hpe", "reported": false, "note": "ClusterStor E1000 hardware and Lustre integration" } ], "unresolved": [ "maker of the 30.72 TB NVMe SSDs (not named in any source read)" ], "sources": [ "https://arxiv.org/pdf/2410.11199", "https://www.bristol.ac.uk/news/2023/november/supercomputer-announcement.html" ] }, { "n": 10, "line": "Multi-tenant flash storage platform (VAST Data Platform: 4 C-Box protocol servers with 16 C-Nodes, 3 D-Box flash enclosures, 3.56 PB native usable)", "parts": [ { "id": "2.10", "name": "AI storage systems", "url": "/parts/memory-and-storage/ai-storage-systems-parallel-file-systems-object" } ], "companies": [ { "id": "vast-data", "name": "VAST Data", "url": "/companies/vast-data", "reported": false, "note": "software-defined, multi-protocol storage platform (supports object storage)" } ], "unresolved": [ "maker of the flash drives and of the enclosure hardware inside the C-Boxes and D-Boxes (not named)" ], "sources": [ "https://arxiv.org/pdf/2410.11199" ] }, { "n": 11, "line": "In-cabinet direct liquid cooling (cold plates on compute and network blades, dripless quick connects, coolant distribution units serving up to four cabinets each, up to 400 kW per cabinet)", "parts": [ { "id": "5.01", "name": "Direct-to-chip cold plates", "url": "/parts/liquid-cooling-it-side-chip-to-cdu/direct-to-chip-cold-plates" }, { "id": "5.05", "name": "Coolant distribution units", "url": "/parts/liquid-cooling-it-side-chip-to-cdu/coolant-distribution-units-in-rack-in-row" } ], "companies": [ { "id": "hpe", "name": "HPE", "url": "/companies/hpe", "reported": false, "note": "cabinet, blade cold plates and CDUs are part of the HPE Cray EX design" } ], "unresolved": [ "makers of the cold plates, quick disconnects and CDU internals (not named in any source read)" ], "sources": [ "https://arxiv.org/pdf/2410.11199", "https://www.bristol.ac.uk/research/centres/bristol-supercomputing/articles/2026/inside-isambard-ai.html" ] }, { "n": 12, "line": "Modular data centre (HPE Performance Optimized Data Center POD: IT module, UPS and power modules, heat-removal plant and pump skid; Phase 1 in a DC20 module, Phase 2 in a second, larger module)", "parts": [ { "id": "9.07", "name": "Prefab / modular data centers & power/cooling skids", "url": "/parts/building-shell-site-and-construction/prefab-modular-data-centers-and-power-cooling-skids" } ], "companies": [ { "id": "hpe", "name": "HPE", "url": "/companies/hpe", "reported": false, "note": "HPE's POD modular data centre design, delivered with the Isambard systems; the DC20 arrived on trucks on 11 March 2024 and assembly took 24 hours" } ], "unresolved": [ "manufacturers of the chillers, pump skids, UPS, transformers and DCIM software inside the POD (not named; see outro)", "what the paper's acknowledged parties \"Contour\" and \"Oakland\" (building contractors) supplied" ], "sources": [ "https://www.bristol.ac.uk/news/2023/november/supercomputer-announcement.html", "https://arxiv.org/pdf/2410.11199", "https://www.bristol.ac.uk/research/centres/bristol-supercomputing/articles/2026/inside-isambard-ai.html" ] }, { "n": 13, "line": "Cluster management software (HPE Cray Systems Management, Kubernetes-based, with multi-tenancy)", "parts": [ { "id": "11.02", "name": "Cluster management & GPU orchestration", "url": "/parts/software-tools-and-services-touching-the-hardware-optional/cluster-management-and-gpu-orchestration-out-of-scope-for-a-part" } ], "companies": [ { "id": "hpe", "name": "HPE", "url": "/companies/hpe", "reported": false, "note": "HPE Cray Systems Management (CSM) and Slingshot fabric manager" } ], "unresolved": [], "sources": [ "https://arxiv.org/pdf/2410.11199" ] } ] }