Reference
Glossary
25 terms used across this site, defined once. Jump straight to one, or read the technical guide for the longer version of most of these.
- Confidence tiers (verified / reported / estimated)
-
Verified means two genuinely independent organisations name the same specific fact. Reported means one source states it, or two sources that are not independent (a press release and a page that copies it). Estimated means we derived the figure ourselves, such as a GPU count from a node count.
Confidence tiers → - CORAL / CORAL-2
-
The US Department of Energy's joint procurement program (Collaboration of Oak Ridge, Argonne and Lawrence Livermore) that produced Summit and Sierra (CORAL) and Frontier, Aurora and El Capitan (CORAL-2). Systems from the same round often share an architecture.
- Direct liquid cooling (DLC)
-
Cooling that brings liquid to a cold plate on or near the chip itself, rather than cooling room air. Needed once rack density passes roughly 20 kW, the practical ceiling for air cooling; current racks are commonly rated at several hundred kW.
Power and cooling → - Dragonfly topology
-
A network topology used by several modern interconnects (Slingshot, some InfiniBand deployments) that keeps the number of hops between any two nodes low even at large scale, at the cost of needing more cabling than a simpler tree.
- FLOP/s
-
Floating-point operations per second. A PFlop/s is 10^15 per second; an EFlop/s is 10^18. The unit says nothing about precision on its own, which is why flops_basis and the arithmetic format (FP64, FP8, and so on) matter alongside it.
- flops_basis
-
The field on every system row stating what kind of number Rmax and Rpeak are: hpl_measured (an actual benchmark run), vendor_peak (a manufacturer specification), operator_stated (the operator's own figure, not necessarily a benchmark) or our_estimate. Never assume a figure is measured without checking this field.
- FP64 / FP32 / BF16 / FP8
-
Floating-point formats, from 64 bits per number down to 8. Scientific simulation generally needs FP64 for correctness; neural network training tolerates BF16 or FP8, which is faster and cheaper per operation. The same chip can differ by a factor of thirty in FLOP/s between formats, which is why comparing two systems requires knowing which format each figure uses.
AI clusters versus HPC systems → - Green500
-
TOP500's companion ranking, the same systems sorted by power efficiency (GFlop/s per watt) instead of raw Rmax.
- HBM (High Bandwidth Memory)
-
Stacked memory placed physically next to (or on the same package as) a processor, trading capacity for far higher bandwidth than conventional DRAM. Standard on modern accelerators; the generation (HBM2e, HBM3, HBM3e) affects both bandwidth and how much memory an accelerator can hold.
- HHI (Herfindahl-Hirschman Index)
-
A market-concentration measure: the sum of every supplier's squared percentage share. Ranges from near 0 (many small suppliers) to 10000 (one supplier holds everything). US antitrust guidance treats a market above 2500 as highly concentrated.
Supply-chain concentration → - HPCG
-
The High Performance Conjugate Gradients benchmark, a second TOP500-associated ranking meant to correct for HPL rewarding peak arithmetic over memory bandwidth and irregular access, which most real simulation codes are bound by instead.
- HPL (LINPACK)
-
The High-Performance LINPACK benchmark: solving a large dense system of linear equations. It has been the TOP500 sorting metric since 1993, mostly because it scales predictably and every architecture can run it, not because it resembles most real workloads.
Measuring performance → - Immersion cooling
-
Submerging hardware directly in a dielectric (non-conductive) fluid rather than routing liquid through cold plates. Less common than direct liquid cooling in this dataset, but used at some facilities for very high density.
- Integrator
-
The company that assembles and delivers a system, which may differ from the company that operates it (the site) or the companies that supply its chips. HPE, Lenovo, Dell and Atos/Eviden are common integrators in this dataset.
- Interconnect
-
The network fabric linking a system's nodes: InfiniBand, HPE Slingshot, Fujitsu's Tofu, or a vendor's proprietary fabric. Distinct from the network a node uses to talk to storage or the outside world.
The interconnect → - NISQ
-
Noisy Intermediate-Scale Quantum: the current era of quantum hardware, physicist John Preskill's term for devices with tens to a few hundred qubits and error rates too high for reliable error correction by default.
Quantum computers versus supercomputers → - NVLink / NVLink domain
-
NVIDIA's chip-to-chip interconnect, faster than a general-purpose network fabric, used to link GPUs within a tight cluster (a rack-scale NVLink domain such as the GB200 NVL72) before traffic has to cross onto the slower general interconnect.
- PUE
-
Power Usage Effectiveness: total facility power divided by the power delivered to the computing equipment itself. A PUE of 1.1 means 10% of the electricity a site draws goes to cooling, lighting and other overhead rather than the machine. Lower is better; 1.0 is the physical floor.
- Qubit
-
The basic unit of quantum information: a two-level quantum system that can be placed in a superposition of both states, unlike a classical bit which is only ever 0 or 1.
Quantum computers versus supercomputers → - Rmax
-
The measured performance of a system on the LINPACK (HPL) benchmark, in FLOP/s. This is what a TOP500 rank is actually sorted by. A vendor peak figure is not Rmax; see flops_basis.
Measuring performance → - Rpeak
-
The theoretical maximum FLOP/s a system's hardware could produce if every core ran at full clock speed on every cycle with no overhead. Always higher than Rmax; the ratio between them (HPL efficiency) is itself a fact worth comparing across systems.
- Stockpile stewardship
-
The US National Nuclear Security Administration's program of certifying the nuclear arsenal through simulation rather than live testing, which the US has not conducted since the 1990s. Several systems in this dataset are built for exactly this workload.
Showcase: weapons physics → - TDP
-
Thermal Design Power: the amount of heat a chip is designed to dissipate under sustained load, in watts. Used as a proxy for power draw per chip, though actual draw depends on the workload and can run above or below TDP.
- TOP500
-
The twice-yearly ranking of the world's fastest measured supercomputers by Rmax, published since 1993. This site never republishes the list; system rows store only a pointer (edition and rank) linking to the canonical entry.
Methodology → - workload_class
-
The field on each system row: hpc (traditional simulation), ai_training, ai_inference, mixed, or classified. A categorical fact from the operator or a clear statement of mission, not our guess at what runs on a machine day to day.
Showcase →