Section
Supercomputers
The ranking is a solved problem and we are not competing on it. What nobody publishes is the layer underneath: which company supplied each accelerator, CPU, fabric and cabinet; which corporate parent that supplier rolls up to after thirty years of consolidation; and which listed ticker therefore carries the exposure. Every system here is a node in a parts graph, not a row in a table.
Phase 1 seed dataset — under review. These rows were compiled by hand from public operator, laboratory and vendor sources, and have not yet been through a second-reader verification pass. Every claim carries its source and a confidence tier; treat anything below verified as a lead, not a citation.
- Systems
- 21
- Countries
- 6
- Aggregate Rmax
- 6.60 EFlop/s
- Aggregate power
- 231.53 MW
- Span
- 1993–2024
Accelerator share, weighted by FLOPS
Machine-count share is the figure everyone quotes, and it understates whoever supplies the largest systems. Weighted by installed Rmax across the systems in this dataset, the picture is different. Supplier is rolled up to the corporate parent that owns it today.
| Accelerator supplier (today's parent) | Installed Rmax | Share | Distribution |
|---|---|---|---|
| AMD | 3.47 EFlop/s | 52.7% | |
| NVIDIA | 1.50 EFlop/s | 22.7% | |
| Intel | 1.04 EFlop/s | 15.8% | |
| No accelerator (CPU-only) | 547.8 PFlop/s | 8.3% | |
| NUDT | 30.72 PFlop/s | 0.5% | |
| IBM | 1.03 PFlop/s | 0.0% |
How this was calculated
Each system contributes its last publicly reported Rmax once. Where a system has more than one accelerator supplier — which happens when a machine was rebuilt with different silicon, as Tianhe-2 was — its FLOPS are split evenly between them rather than double-counted. Systems with no accelerator edge, such as Fugaku and the K computer, are shown as a CPU-only bucket rather than dropped, because dropping them is the most common way this chart is drawn wrong elsewhere.
Suppliers are rolled up through vendor_aliases to the parent that owns them
today, so Mellanox's EDR fabric counts toward NVIDIA and Cray's Gemini counts toward HPE.
Aggregates over a 21-system seed are illustrative, not market share.
The full treatment, over the complete historical backfill, is
Phase 2.
Systems
Ranked and unranked systems carried in one taxonomy, newest first. Confidence is per-system and every claim on the page behind it carries its own source.
The Rank column links out to the canonical entry on top500.org. We store rank and edition as a pointer and never reproduce a ranked table — see sourcing policy.
Where to go next
Explorer
Query the graph directly: systems by supplier, role, country, era or fabric.
Analysis
The eight derived views the incumbent ranking structurally cannot produce.
Parts taxonomy
Every accelerator, CPU, fabric, cabinet and cooling architecture, with the systems using it.
Methodology
Confidence tiers, sourcing policy, entity resolution, and what we refuse to publish.