Horizon: every supplier in the machine

Texas Advanced Computing Center (University of Texas at Austin), United States

Accelerator

NVIDIA Blackwell GPUs on the Grace Blackwell GB200 NVL4 platform, 2 per node, 2,000 GPU nodes (4,000 GPUs); 185 GiB HBM3 listed per GPU

Host CPU

GPU nodes: 1 NVIDIA Grace (72 cores) per node. CPU partition (planned): 4,750 nodes with two 88-core NVIDIA Vera sockets each

Integrator

Dell Technologies (Dell Integrated Rack Scalable Systems with direct-liquid-cooled PowerEdge servers)

Nodes

2,000 Grace Blackwell nodes in 28 racks of 72 nodes (144 GPUs, about 215 kW each); 4,750 Vera nodes planned in 66 racks of 72 dual-socket nodes (over 100 kW each); 128 cabinets in all, including 34 storage, switching and management racks

Memory

GB200 node: Grace CPU with LPDDR5X (120 GiB per CPU in the user guide's text, 240 GiB in its table) plus 2 Blackwell GPUs with 185 GiB HBM3 each. Vera node memory not published

Interconnect

NVIDIA XDR InfiniBand (Quantum-X800): 800 Gb/s per Grace Blackwell node and 400 Gb/s per Vera node, two non-blocking fat trees (2,000 and 4,750 ports) joined by a third level; about 9,000 endpoints and over 15,000 cables

Storage

VAST Data shared all-flash file system, about 400 PB usable for $HOME and $SCRATCH, 8 TB/s write and 16 TB/s read; storage back end designed and provided by VAST

Cooling

Direct-liquid-cooled racks with rear-door cooling, 7 coolant distribution units and PG-25 coolant; hall built for up to 250 kW per cabinet inside Sabey's Round Rock campus

Power

TACC estimates about 13 MW for the 128 cabinets (August 2026 slides); the hall is described as a 20 MW data center (user guide) or 15-20 MW (TACC announcement). No measured figure published

Rmax (HPL)

Not on the TOP500 in any source read. TACC reports native HPL of 124.3 PFlop/s on about 3,700 GPUs and expects about 130 PFlop/s from the full GPU partition (August 2026 slides)

Rpeak

Not a list figure. TACC's quoted FP64 peak has varied: 300 PFlop/s (Dell and TACC, November 2025), 360 PFlop/s (TACC system page), 160 PFlop/s FP64 for the GPU component alone (user guide)

Status

Installation announced 17 November 2025; as of the August 2026 slides all 4,000 GPUs are on the floor (3,856 powered and cooled, 3,072 tested with HPL and healthy) and only early users have access; the user guide (24 July 2026) says Horizon is still limited to internal users; the Vera CPU partition is expected in early 2027. Production start is given as Spring 2026 (TACC announcement) and Fall 2026 (TACC system page)

Unmarked companies come from the operator's own pages (tacc.utexas.edu, docs.tacc.utexas.edu, lccf.tacc.utexas.edu), its executive director's presentation, or the named vendor's own press release or site (dell, sabeydatacenters.com); reported companies come from named trade press whose article body was read in full, not just the headline. No company on this page is marked reported. This is a map of who plays where, not a confirmed bill of materials: TACC's Horizon system page and user guide (updated August 2026), TACC's 17 November 2025 announcement, a TACC presentation dated August 2026, TACC's Ranch page, Dell's November 2025 press release, and Sabey Data Centers' news pages, researched 2026-10-22

This page is the parts-first view. For Horizon's full system record on the sister site Compute Atlas, an open dataset of supercomputer and AI-cluster supply chains, see Horizon on Compute Atlas.

TACC names Dell, NVIDIA, VAST Data, Spectra Logic, Versity and Sabey Data Centers as partners, and its executive director's slides also mention a vendor called Cambridge without saying what it supplies. No source read names the maker of the Blackwell GPUs' HBM, the Grace and Vera LPDDR memory, the flash drives that make up the 400 PB file system, the InfiniBand cables and optics, the in-rack and rear-door heat exchangers, the seven coolant distribution units, or the facility's electrical and cooling plant, so none of those get a line. Several headline figures changed between TACC's pages (peak FP64 of 300, 360 or 160 petaflops depending on the page, 15-20 MW or 20 MW, Spring or Fall 2026 production, 4,750 or 4,752 CPU nodes) and the numbers in the specs follow the most recent operator document, with the differences noted. TACC's own slides say the CPU partition will arrive later and smaller than planned, with more memory per node, and put this down to market conditions. Power and cooling outside the cabinet sit in Layer 6 and Layer 7.

Machine-readable: Markdown ·JSON ·every line of every machine (CSV)

Other machines:El Capitan · Frontier · Aurora · JUPITER · Alps · LUMI · Leonardo · Isambard-AI · Venado · Perlmutter · Eagle (Microsoft Azure) · Eos · Fugaku · MareNostrum 5 · Doudna (NERSC-10). See also the AI rack teardowns and thesupercomputer overview.