Compute·Atlas

Supercomputers/Analysis/Interconnect share

Analysis 02

Interconnect share over time

The fabric layer consolidated twice in twelve months and almost nobody charted it. NVIDIA completed its acquisition of Mellanox in April 2020; HPE completed its acquisition of Cray in September 2019. Four independent fabric suppliers became two corporate owners, and because a spec string names the product rather than the parent, the change is invisible unless you resolve the entities first. As of 2025, Hewlett Packard Enterprise accounts for 72% of installed FLOPS by fabric and NVIDIA for 19%.

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.

  • Hewlett Packard Enterprise
  • NVIDIA
  • Fujitsu
  • NRCPC
  • Intel
  • NUDT
  • IBM
  • Other (3 suppliers)
Interconnect share of installed FLOPS, 1993–2025 Stacked area chart of each interconnect supplier's share of installed Rmax by year, with suppliers rolled up to the corporate parent that owns them today. Hewlett Packard EnterpriseNVIDIAFujitsuNRCPCIntelNUDTIBMOther (3 suppliers) 0%25%50%75%100%19931998200320082013201820232025 HPENVIDIAFujitsu 1993 HPE 0.0% NVIDIA 0.0% Fujitsu 0.0% NRCPC 0.0% Intel 0.0% NUDT 0.0% IBM 0.0% Other 100.0%1994 HPE 0.0% NVIDIA 0.0% Fujitsu 0.0% NRCPC 0.0% Intel 0.0% NUDT 0.0% IBM 0.0% Other 100.0%1995 HPE 0.0% NVIDIA 0.0% Fujitsu 0.0% NRCPC 0.0% Intel 0.0% NUDT 0.0% IBM 0.0% Other 100.0%1996 HPE 0.0% NVIDIA 0.0% Fujitsu 0.0% NRCPC 0.0% Intel 0.0% NUDT 0.0% IBM 0.0% Other 100.0%1997 HPE 0.0% NVIDIA 0.0% Fujitsu 0.0% NRCPC 0.0% Intel 94.7% NUDT 0.0% IBM 0.0% Other 5.3%1998 HPE 0.0% NVIDIA 0.0% Fujitsu 0.0% NRCPC 0.0% Intel 94.7% NUDT 0.0% IBM 0.0% Other 5.3%1999 HPE 0.0% NVIDIA 0.0% Fujitsu 0.0% NRCPC 0.0% Intel 94.7% NUDT 0.0% IBM 0.0% Other 5.3%2000 HPE 0.0% NVIDIA 0.0% Fujitsu 0.0% NRCPC 0.0% Intel 17.6% NUDT 0.0% IBM 81.4% Other 1.0%2001 HPE 0.0% NVIDIA 0.0% Fujitsu 0.0% NRCPC 0.0% Intel 17.6% NUDT 0.0% IBM 81.4% Other 1.0%2002 HPE 0.0% NVIDIA 0.0% Fujitsu 0.0% NRCPC 0.0% Intel 1.9% NUDT 0.0% IBM 8.8% Other 89.2%2003 HPE 0.0% NVIDIA 0.0% Fujitsu 0.0% NRCPC 0.0% Intel 1.9% NUDT 0.0% IBM 8.8% Other 89.2%2004 HPE 0.0% NVIDIA 0.0% Fujitsu 0.0% NRCPC 0.0% Intel 0.2% NUDT 0.0% IBM 90.5% Other 9.3%2005 HPE 0.0% NVIDIA 0.0% Fujitsu 0.0% NRCPC 0.0% Intel 0.2% NUDT 0.0% IBM 90.5% Other 9.3%2006 HPE 0.0% NVIDIA 0.0% Fujitsu 0.0% NRCPC 0.0% Intel 0.2% NUDT 0.0% IBM 90.5% Other 9.3%2007 HPE 0.0% NVIDIA 0.0% Fujitsu 0.0% NRCPC 0.0% Intel 0.0% NUDT 0.0% IBM 90.6% Other 9.4%2008 HPE 0.0% NVIDIA 87.8% Fujitsu 0.0% NRCPC 0.0% Intel 0.0% NUDT 0.0% IBM 11.8% Other 0.5%2009 HPE 18.1% NVIDIA 71.9% Fujitsu 0.0% NRCPC 0.0% Intel 0.0% NUDT 0.0% IBM 9.7% Other 0.4%2010 HPE 14.4% NVIDIA 57.0% Fujitsu 0.0% NRCPC 0.0% Intel 0.0% NUDT 21.0% IBM 7.7% Other 0.0%2011 HPE 7.7% NVIDIA 30.7% Fujitsu 46.2% NRCPC 0.0% Intel 0.0% NUDT 11.3% IBM 4.1% Other 0.0%2012 HPE 29.3% NVIDIA 10.6% Fujitsu 15.9% NRCPC 0.0% Intel 0.0% NUDT 3.9% IBM 40.4% Other 0.0%2013 HPE 15.2% NVIDIA 5.5% Fujitsu 8.2% NRCPC 0.0% Intel 0.0% NUDT 50.2% IBM 20.9% Other 0.0%2014 HPE 15.3% NVIDIA 4.7% Fujitsu 8.3% NRCPC 0.0% Intel 0.0% NUDT 50.6% IBM 21.1% Other 0.0%2015 HPE 15.3% NVIDIA 4.7% Fujitsu 8.3% NRCPC 0.0% Intel 0.0% NUDT 50.8% IBM 20.8% Other 0.0%2016 HPE 14.3% NVIDIA 2.6% Fujitsu 4.5% NRCPC 39.9% Intel 0.0% NUDT 27.5% IBM 11.3% Other 0.0%2017 HPE 24.9% NVIDIA 2.0% Fujitsu 3.5% NRCPC 31.0% Intel 8.4% NUDT 21.4% IBM 8.8% Other 0.0%2018 HPE 12.8% NVIDIA 42.8% Fujitsu 1.8% NRCPC 16.0% Intel 11.1% NUDT 11.0% IBM 4.5% Other 0.0%2019 HPE 12.3% NVIDIA 45.0% Fujitsu 1.7% NRCPC 15.4% Intel 10.7% NUDT 10.6% IBM 4.3% Other 0.0%2020 HPE 5.0% NVIDIA 35.1% Fujitsu 38.8% NRCPC 8.2% Intel 5.7% NUDT 5.6% IBM 1.6% Other 0.0%2021 HPE 12.6% NVIDIA 33.5% Fujitsu 35.8% NRCPC 7.5% Intel 5.2% NUDT 5.2% IBM 0.0% Other 0.0%2022 HPE 58.9% NVIDIA 20.4% Fujitsu 13.8% NRCPC 2.9% Intel 2.0% NUDT 2.0% IBM 0.0% Other 0.0%2023 HPE 57.7% NVIDIA 29.3% Fujitsu 8.8% NRCPC 1.8% Intel 1.1% NUDT 1.3% IBM 0.0% Other 0.0%2024 HPE 70.4% NVIDIA 20.6% Fujitsu 6.2% NRCPC 1.3% Intel 0.7% NUDT 0.9% IBM 0.0% Other 0.0%2025 HPE 72.0% NVIDIA 18.9% Fujitsu 6.3% NRCPC 1.3% Intel 0.6% NUDT 0.9% IBM 0.0% Other 0.0%
Share of installed Rmax by fabric supplier. Mellanox counts toward NVIDIA from 2020 and Cray toward HPE from 2019, because the question is who owns the business now.
Download this chart's data (CSV) ↓
Table view: Interconnect share of installed FLOPS
YearHewlett Packard EnterpriseNVIDIAFujitsuNRCPCIntelNUDTIBMOther (3 suppliers)Installed total
19930.0%0.0%0.0%0.0%0.0%0.0%0.0%100.0%59.7 GFlop/s
19940.0%0.0%0.0%0.0%0.0%0.0%0.0%100.0%59.7 GFlop/s
19950.0%0.0%0.0%0.0%0.0%0.0%0.0%100.0%59.7 GFlop/s
19960.0%0.0%0.0%0.0%0.0%0.0%0.0%100.0%59.7 GFlop/s
19970.0%0.0%0.0%0.0%94.7%0.0%0.0%5.3%1.13 TFlop/s
19980.0%0.0%0.0%0.0%94.7%0.0%0.0%5.3%1.13 TFlop/s
19990.0%0.0%0.0%0.0%94.7%0.0%0.0%5.3%1.13 TFlop/s
20000.0%0.0%0.0%0.0%17.6%0.0%81.4%1.0%6.07 TFlop/s
20010.0%0.0%0.0%0.0%17.6%0.0%81.4%1.0%6.07 TFlop/s
20020.0%0.0%0.0%0.0%1.9%0.0%8.8%89.2%55.81 TFlop/s
20030.0%0.0%0.0%0.0%1.9%0.0%8.8%89.2%55.81 TFlop/s
20040.0%0.0%0.0%0.0%0.2%0.0%90.5%9.3%534.01 TFlop/s
20050.0%0.0%0.0%0.0%0.2%0.0%90.5%9.3%534.01 TFlop/s
20060.0%0.0%0.0%0.0%0.2%0.0%90.5%9.3%534.01 TFlop/s
20070.0%0.0%0.0%0.0%0.0%0.0%90.6%9.4%528.00 TFlop/s
20080.0%87.8%0.0%0.0%0.0%0.0%11.8%0.5%7.95 PFlop/s
200918.1%71.9%0.0%0.0%0.0%0.0%9.7%0.4%9.71 PFlop/s
201014.4%57.0%0.0%0.0%0.0%21.0%7.7%0.0%12.24 PFlop/s
20117.7%30.7%46.2%0.0%0.0%11.3%4.1%0.0%22.75 PFlop/s
201229.3%10.6%15.9%0.0%0.0%3.9%40.4%0.0%66.10 PFlop/s
201315.2%5.5%8.2%0.0%0.0%50.2%20.9%0.0%127.5 PFlop/s
201415.3%4.7%8.3%0.0%0.0%50.6%21.1%0.0%126.5 PFlop/s
201515.3%4.7%8.3%0.0%0.0%50.8%20.8%0.0%126.1 PFlop/s
201614.3%2.6%4.5%39.9%0.0%27.5%11.3%0.0%233.1 PFlop/s
201724.9%2.0%3.5%31.0%8.4%21.4%8.8%0.0%299.7 PFlop/s
201812.8%42.8%1.8%16.0%11.1%11.0%4.5%0.0%582.3 PFlop/s
201912.3%45.0%1.7%15.4%10.7%10.6%4.3%0.0%605.9 PFlop/s
20205.0%35.1%38.8%8.2%5.7%5.6%1.6%0.0%1.14 EFlop/s
202112.6%33.5%35.8%7.5%5.2%5.2%0.0%0.0%1.23 EFlop/s
202258.9%20.4%13.8%2.9%2.0%2.0%0.0%0.0%3.21 EFlop/s
202357.7%29.3%8.8%1.8%1.1%1.3%0.0%0.0%5.03 EFlop/s
202470.4%20.6%6.2%1.3%0.7%0.9%0.0%0.0%7.16 EFlop/s
202572.0%18.9%6.3%1.3%0.6%0.9%0.0%0.0%7.01 EFlop/s
How this was calculated

Identical construction to the accelerator chart: each in-service system contributes its last reported Rmax to every year it was live, split evenly where a system has more than one fabric supplier.

Every system in this dataset has at least one interconnect edge, so unlike the accelerator chart there is no “none” band: a machine without a fabric is not a machine.

The rollup is the whole point of this chart. Suppliers are recorded as the company that shipped the fabric at build time and then resolved through vendor_aliases to today's parent. Read without that step, Summit's dual-rail EDR belongs to a company that no longer exists independently, and the concentration this chart shows does not appear.

One caveat that cuts against the chart: vertically integrated systems understate their supplier here. IBM's BlueGene fabrics and Fujitsu's Tofu are on the processor package, and appear in no spec string as a third-party product. They are recorded as IBM and Fujitsu edges, but a dataset built only from fabric-vendor mentions would miss them entirely.

Installed base of the 52 systems in this dataset, not a census of the market.

The read

Three structural facts sit in this chart. The first is that the fabric layer has always been more concentrated than the accelerator layer, and for most of the period it was concentrated in vendors who also supplied the machine: Cray's SeaStar, Gemini and Aries; IBM's torus networks; Fujitsu's Tofu. The independent merchant fabric is a relatively recent and relatively brief phenomenon.

The second is that the 2019–2020 acquisitions removed the last two large independent fabric businesses in the same twelve months. After them, the two companies that supply most accelerator FLOPS also supply most fabric FLOPS. Accelerator share and fabric share stopped being independent variables, which matters to anyone modelling concentration risk from either one alone.

The third is Omni-Path, which is the counterexample worth keeping. Intel discontinued it in 2019 and it did not die: Cornelis Networks took it over in 2020 and continues it. It is the only case in this dataset of a fabric outliving its vendor's interest in it, and it is why the parts table records Cornelis as the current supplier of record while the edges from Stampede2, SuperMUC-NG and TSUBAME3.0 still correctly name Intel as the supplier at build.