Supercomputers/Analysis/Accelerator share
Analysis 01
Accelerator share, weighted by FLOPS
Across the 52 systems in this dataset, AMD supplies 50% of installed FLOPS as of 2025, against 26% for NVIDIA. Machine-count share (the figure normally quoted) produces a different ordering, because it counts one exascale system and one departmental cluster as one machine each. Weighting by capability is the correction, and it is the whole argument of this page.
Phase 1 seed dataset, compiled by hand. These rows were built from public operator, laboratory and vendor sources. A mechanical second-reader pass has since fetched every cited source: 32 of 52 systems have a readable citation that names them, and 4 are genuinely weakly sourced. Every claim carries its source and a confidence tier. Treat anything below verified as a lead, not a citation.
Read the shape, not the spikes. This is the installed base of the 52 systems in this dataset: a curated seed chosen for supply-chain diversity, not a census of the market. At that sample size one machine entering or leaving service moves a share line by tens of points, so year-to-year jumps are composition effects rather than market events. The multi-year trends are the part worth quoting; the individual steps usually have one system's name on them, and the table view under the chart will tell you which.
- AMD
- NVIDIA
- Intel
- NUDT
- IBM
- No accelerator (CPU-only)
Table view: Accelerator share of installed FLOPS
| Year | AMD | NVIDIA | Intel | NUDT | IBM | No accelerator (CPU-only) | Installed total |
|---|---|---|---|---|---|---|---|
| 1993 | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% | 59.7 GFlop/s |
| 1994 | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% | 59.7 GFlop/s |
| 1995 | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% | 59.7 GFlop/s |
| 1996 | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% | 59.7 GFlop/s |
| 1997 | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% | 1.13 TFlop/s |
| 1998 | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% | 1.13 TFlop/s |
| 1999 | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% | 1.13 TFlop/s |
| 2000 | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% | 6.07 TFlop/s |
| 2001 | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% | 6.07 TFlop/s |
| 2002 | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% | 55.81 TFlop/s |
| 2003 | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% | 55.81 TFlop/s |
| 2004 | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% | 534.01 TFlop/s |
| 2005 | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% | 534.01 TFlop/s |
| 2006 | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% | 534.01 TFlop/s |
| 2007 | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% | 528.00 TFlop/s |
| 2008 | 0.0% | 0.0% | 0.0% | 0.0% | 12.9% | 87.1% | 7.95 PFlop/s |
| 2009 | 0.0% | 0.0% | 0.0% | 0.0% | 10.6% | 89.4% | 9.71 PFlop/s |
| 2010 | 0.0% | 21.0% | 0.0% | 0.0% | 8.4% | 70.6% | 12.24 PFlop/s |
| 2011 | 0.0% | 11.3% | 0.0% | 0.0% | 4.5% | 84.2% | 22.75 PFlop/s |
| 2012 | 0.0% | 30.5% | 0.0% | 0.0% | 1.6% | 68.0% | 66.10 PFlop/s |
| 2013 | 0.0% | 15.8% | 24.1% | 24.1% | 0.8% | 35.2% | 127.5 PFlop/s |
| 2014 | 0.0% | 15.9% | 24.3% | 24.3% | 0.0% | 35.5% | 126.5 PFlop/s |
| 2015 | 0.0% | 16.0% | 24.4% | 24.4% | 0.0% | 35.3% | 126.1 PFlop/s |
| 2016 | 0.0% | 8.6% | 19.2% | 13.2% | 0.0% | 59.0% | 233.1 PFlop/s |
| 2017 | 0.0% | 16.5% | 25.2% | 10.2% | 0.0% | 48.0% | 299.7 PFlop/s |
| 2018 | 0.0% | 53.7% | 13.0% | 5.3% | 0.0% | 28.1% | 582.3 PFlop/s |
| 2019 | 0.0% | 51.6% | 12.5% | 5.1% | 0.0% | 30.9% | 605.9 PFlop/s |
| 2020 | 0.0% | 35.4% | 6.6% | 2.7% | 0.0% | 55.3% | 1.14 EFlop/s |
| 2021 | 0.0% | 40.2% | 6.1% | 2.5% | 0.0% | 51.2% | 1.23 EFlop/s |
| 2022 | 54.0% | 23.0% | 2.4% | 1.0% | 0.0% | 19.7% | 3.21 EFlop/s |
| 2023 | 34.5% | 31.0% | 21.4% | 0.6% | 0.0% | 12.6% | 5.03 EFlop/s |
| 2024 | 48.5% | 27.5% | 14.8% | 0.4% | 0.0% | 8.7% | 7.16 EFlop/s |
| 2025 | 49.6% | 25.9% | 15.2% | 0.4% | 0.0% | 8.9% | 7.01 EFlop/s |
How this was calculated
For each calendar year, every system in service that year contributes its last publicly reported Rmax. A system is in service from its first-operational date until its decommissioning date, or to the present if it has not been retired.
Where a system has more than one accelerator supplier, which happens when a machine is rebuilt with different silicon, as Tianhe-2 was after the 2015 export controls, its FLOPS are split evenly between them rather than counted twice.
Systems with no accelerator are kept in the denominator as their own band. Fugaku, Hawk and ARCHER2 are real machines representing real installed capability, and dropping them is the most common way this chart is drawn wrong elsewhere: it silently redefines the question from “share of installed FLOPS” to “share of accelerated FLOPS” and inflates every supplier.
Suppliers are the company that shipped the part at build time, rolled up through
vendor_aliases to the parent that owns that business today. Beyond eight suppliers
the tail folds into a single “Other” band rather than being given a ninth colour.
7 systems carry no Rmax and are excluded from this chart entirely: Blue Waters, Isambard-AI, JUPITER, Venado, Shaheen III, Meta GenAI cluster (RoCE), Meta GenAI cluster (InfiniBand). Their component graphs are still on their own pages.
This is the installed base of the systems in this dataset, a curated seed of 52 machines, not a census of the market. The shape is real; the absolute totals are a sample.
The read
The chart has three regimes. Until roughly 2010 there is no accelerator band at all worth speaking of: leadership compute was CPUs, and the only question was whose. The middle period, 2012 to 2020, is NVIDIA establishing that accelerators are how you buy FLOPS, first through Titan and then through the CORAL systems. The most recent regime is the one that machine-count share hides completely: AMD's share is built out of a very small number of very large machines, and a single procurement moves the line several points.
That concentration is the finding, not the ordering. A share this sensitive to individual contracts is not a market position in the sense a reader might assume from a supplier chart; it is a statement about who won the last two exascale procurements. The company pages carry the same caution in a different form, as a concentration percentage.
The CPU-only band is worth watching on its own. It does not decay to zero, and the systems in it are not legacy: Fugaku was the fastest machine in the world with no accelerator at all, and ARCHER2 and Hawk were both bought deliberately without one because their users run codes that do not vectorise onto GPUs. This chart measures supply, not suitability.