Supercomputers / Analysis / Machine balance
Analysis 16
Machine balance: memory and storage per flop
Computing power has grown much faster than the memory and storage built around it. Across the
272 systems in this dataset with a recorded memory size and a measured Rmax, the median machine
first run in 2000 to 2004 carried 759 GB of memory for every TFlop/s it measured,
about 0.76 bytes per flop/s. For machines first run from 2025
the median is 8.1 GB, or 0.008 bytes. Storage followed the same direction: from
21 PB of parallel storage per PFlop/s in 2000 to 2004 to 0.65 PB.
Totals, not bandwidth. This compares how much memory and storage a machine has with
how fast it computes, not how fast data moves, which is what most codes are actually limited by and
which few operators publish. Recorded memory totals do not always say whether accelerator memory is
included, so the accelerated points may be understated or overstated by that amount. Only systems
with a measured Rmax appear: 272 for memory and 261 for storage, of 1019 with an Rmax.
Researched with AI assistance, checked by scripts.
These rows were researched by AI agents reading public operator, laboratory and vendor
sources, appended by scripts and checked by automated validation; no subject-matter expert
has reviewed them (how this site is made ). A
mechanical second-reader pass has since fetched every
cited source: 1,350 of 1,581 systems have a readable citation that names
them, and 64 are genuinely weakly sourced. Every claim carries its source and a
confidence tier. Treat anything below
verified as a lead, not a citation.
Memory per unit of measured performance, by first-light year
Log-scale dot plot of each system’s total memory divided by its measured HPL performance, by the year it first ran, coloured by whether it has accelerators, with a line through the group medians.
1 10 100 1000 10k 1990 1995 2000 2005 2010 2015 2020 2025
median per group
Touchstone Delta (Caltech)
CPU-only
576 GB of memory per TFlop/s, 1991
Numerical Wind Tunnel
CPU-only
359 GB of memory per TFlop/s, 1993
ZIB Cray T3D
CPU-only
632 GB of memory per TFlop/s, 1994
KEK VPP500/80
CPU-only
207 GB of memory per TFlop/s, 1995
MPI/IPP Garching Cray T3E
CPU-only
293 GB of memory per TFlop/s, 1996
CP-PACS
CPU-only
348 GB of memory per TFlop/s, 1996
ANU Fujitsu VPP300
CPU-only
505 GB of memory per TFlop/s, 1996
Cambridge Hitachi SR2201/96
CPU-only
1.11e+3 GB of memory per TFlop/s, 1996
Hitachi SR2201/1024
CPU-only
1.16e+3 GB of memory per TFlop/s, 1996
KISTI Cray T3E-900
CPU-only
201 GB of memory per TFlop/s, 1997
Osaka University SX-4
CPU-only
259 GB of memory per TFlop/s, 1997
ASCI Blue-Pacific
CPU-only
1.21e+3 GB of memory per TFlop/s, 1998
DWD Cray T3E1200
CPU-only
179 GB of memory per TFlop/s, 1999
UTokyo SR8000/128
CPU-only
1.17e+3 GB of memory per TFlop/s, 1999
KEK SR8000-F1/100
CPU-only
484 GB of memory per TFlop/s, 2000
ISSP SR8000-F1/60
CPU-only
1.11e+3 GB of memory per TFlop/s, 2000
ASCI White
CPU-only
1.22e+3 GB of memory per TFlop/s, 2000
Lemieux
CPU-only
547 GB of memory per TFlop/s, 2001
MPI/IPP Garching IBM p690
CPU-only
731 GB of memory per TFlop/s, 2001
UTokyo SR8000/MPP
CPU-only
1.35e+3 GB of memory per TFlop/s, 2001
Earth Simulator
CPU-only
279 GB of memory per TFlop/s, 2002
Cheetah (ORNL)
CPU-only
476 GB of memory per TFlop/s, 2002
Bluesky
CPU-only
1.05e+3 GB of memory per TFlop/s, 2002
CPTEC NEC SX-6
CPU-only
3.03e+3 GB of memory per TFlop/s, 2002
Phoenix (ORNL)
CPU-only
134 GB of memory per TFlop/s, 2003
EMSL MPP2
CPU-only
788 GB of memory per TFlop/s, 2003
Intimidata
CPU-only
36.3 GB of memory per TFlop/s, 2004
Columbia
CPU-only
521 GB of memory per TFlop/s, 2004
Mammouth-Ms
CPU-only
552 GB of memory per TFlop/s, 2004
JUMP
CPU-only
898 GB of memory per TFlop/s, 2004
DataStar
CPU-only
1.40e+3 GB of memory per TFlop/s, 2004
ISM Altix 3700
CPU-only
1.68e+3 GB of memory per TFlop/s, 2004
KMA Cray X1E
CPU-only
255 GB of memory per TFlop/s, 2005
Topdawg
CPU-only
401 GB of memory per TFlop/s, 2005
Thunderbird
CPU-only
517 GB of memory per TFlop/s, 2005
BigBen
CPU-only
521 GB of memory per TFlop/s, 2005
HLRS NEC SX-8
CPU-only
1.03e+3 GB of memory per TFlop/s, 2005
JUBL
CPU-only
107 GB of memory per TFlop/s, 2006
blueice
CPU-only
436 GB of memory per TFlop/s, 2006
TSUBAME1.0
CPU-only
561 GB of memory per TFlop/s, 2006
Tera-10
CPU-only
568 GB of memory per TFlop/s, 2006
Kan Balam
CPU-only
593 GB of memory per TFlop/s, 2006
HECToR
CPU-only
136 GB of memory per TFlop/s, 2007
JUGENE
CPU-only
168 GB of memory per TFlop/s, 2007
Ice Storm
CPU-only
257 GB of memory per TFlop/s, 2007
Genius (Garching)
CPU-only
365 GB of memory per TFlop/s, 2007
HLRB-II
CPU-only
690 GB of memory per TFlop/s, 2007
Finis Terrae
CPU-only
1.45e+3 GB of memory per TFlop/s, 2007
Chebyshev
CPU-only
117 GB of memory per TFlop/s, 2008
System G (Virginia Tech)
CPU-only
154 GB of memory per TFlop/s, 2008
Kraken
CPU-only
177 GB of memory per TFlop/s, 2008
Rostand
CPU-only
193 GB of memory per TFlop/s, 2008
Merlin (Cardiff)
CPU-only
200 GB of memory per TFlop/s, 2008
Bluefire
CPU-only
212 GB of memory per TFlop/s, 2008
VIP (Garching)
CPU-only
230 GB of memory per TFlop/s, 2008
Akka
CPU-only
232 GB of memory per TFlop/s, 2008
Netuno
CPU-only
246 GB of memory per TFlop/s, 2008
T2K Tsukuba
CPU-only
268 GB of memory per TFlop/s, 2008
Galera
CPU-only
282 GB of memory per TFlop/s, 2008
Huygens
CPU-only
307 GB of memory per TFlop/s, 2008
Tachyon
CPU-only
353 GB of memory per TFlop/s, 2008
Jade (CINES)
CPU-only
374 GB of memory per TFlop/s, 2008
T2K Todai
CPU-only
377 GB of memory per TFlop/s, 2008
Lomonosov
Accelerated
102 GB of memory per TFlop/s, 2009
Monte Rosa
CPU-only
155 GB of memory per TFlop/s, 2009
Plasma Simulator SR16000
CPU-only
159 GB of memory per TFlop/s, 2009
Earth Simulator 2
CPU-only
163 GB of memory per TFlop/s, 2009
GPC (SciNet)
CPU-only
231 GB of memory per TFlop/s, 2009
TCS (SciNet)
CPU-only
272 GB of memory per TFlop/s, 2009
Tachyon II
CPU-only
279 GB of memory per TFlop/s, 2009
Vayu (NCI)
CPU-only
293 GB of memory per TFlop/s, 2009
VSC-1
CPU-only
316 GB of memory per TFlop/s, 2009
KISTI GAIA
CPU-only
394 GB of memory per TFlop/s, 2009
JSS FX1
CPU-only
904 GB of memory per TFlop/s, 2009
Tianhe-1A
Accelerated
102 GB of memory per TFlop/s, 2010
LOEWE-CSC
Accelerated
196 GB of memory per TFlop/s, 2010
Edge (LLNL)
Accelerated
199 GB of memory per TFlop/s, 2010
Stokes
CPU-only
210 GB of memory per TFlop/s, 2010
Tera-100
CPU-only
286 GB of memory per TFlop/s, 2010
Trestles (SDSC)
CPU-only
297 GB of memory per TFlop/s, 2010
Millipede
CPU-only
302 GB of memory per TFlop/s, 2010
CHEOPS (Cologne)
CPU-only
419 GB of memory per TFlop/s, 2010
Magellan NERSC
CPU-only
719 GB of memory per TFlop/s, 2010
Todi
Accelerated
31.8 GB of memory per TFlop/s, 2011
Dawson2 (UCLA)
Accelerated
65.6 GB of memory per TFlop/s, 2011
K computer
CPU-only
135 GB of memory per TFlop/s, 2011
Lonestar 4
CPU-only
176 GB of memory per TFlop/s, 2011
Zeus (Cyfronet)
Accelerated
197 GB of memory per TFlop/s, 2011
Amazon EC2 Cluster Compute Instances
CPU-only
268 GB of memory per TFlop/s, 2011
VSC-2
CPU-only
275 GB of memory per TFlop/s, 2011
ALPS (NCHC)
CPU-only
416 GB of memory per TFlop/s, 2011
Mustang (LANL)
CPU-only
444 GB of memory per TFlop/s, 2011
Keeneland
Accelerated
26.4 GB of memory per TFlop/s, 2012
Titan
Accelerated
40.4 GB of memory per TFlop/s, 2012
MVS-10P
Accelerated
44.3 GB of memory per TFlop/s, 2012
SANAM
Accelerated
50.5 GB of memory per TFlop/s, 2012
HA-PACS
Accelerated
80.6 GB of memory per TFlop/s, 2012
Triolith
CPU-only
85.9 GB of memory per TFlop/s, 2012
Sisu
CPU-only
86.4 GB of memory per TFlop/s, 2012
Sequoia
CPU-only
87.3 GB of memory per TFlop/s, 2012
KEK SAKURA
CPU-only
89.4 GB of memory per TFlop/s, 2012
Vulcan
CPU-only
89.5 GB of memory per TFlop/s, 2012
Avoca
CPU-only
90.8 GB of memory per TFlop/s, 2012
Fermi
CPU-only
91.6 GB of memory per TFlop/s, 2012
SuperMUC
CPU-only
99.4 GB of memory per TFlop/s, 2012
JUQUEEN
CPU-only
108 GB of memory per TFlop/s, 2012
Vilje
CPU-only
111 GB of memory per TFlop/s, 2012
Yellowstone
CPU-only
115 GB of memory per TFlop/s, 2012
CRIEPI SGI ICE X
CPU-only
116 GB of memory per TFlop/s, 2012
Cab
CPU-only
117 GB of memory per TFlop/s, 2012
Altamira (Cantabria)
Accelerated
140 GB of memory per TFlop/s, 2012
Oakleaf-FX
CPU-only
144 GB of memory per TFlop/s, 2012
Kyushu FX10
CPU-only
144 GB of memory per TFlop/s, 2012
Raijin (NCI)
CPU-only
163 GB of memory per TFlop/s, 2012
Virgo (IIT Madras)
CPU-only
205 GB of memory per TFlop/s, 2012
Darwin III
CPU-only
209 GB of memory per TFlop/s, 2012
Curie
CPU-only
237 GB of memory per TFlop/s, 2012
Abel (UiO)
CPU-only
246 GB of memory per TFlop/s, 2012
Taito (CSC)
CPU-only
247 GB of memory per TFlop/s, 2012
Gordon
CPU-only
293 GB of memory per TFlop/s, 2012
Olympus (PNNL)
CPU-only
379 GB of memory per TFlop/s, 2012
Stampede
Accelerated
52.2 GB of memory per TFlop/s, 2013
Blue Brain IV
CPU-only
90.8 GB of memory per TFlop/s, 2013
Spirit (AFRL)
CPU-only
102 GB of memory per TFlop/s, 2013
TEIDE-HPC
CPU-only
117 GB of memory per TFlop/s, 2013
Fionn
Accelerated
157 GB of memory per TFlop/s, 2013
ATERUI I
CPU-only
169 GB of memory per TFlop/s, 2013
Hydra (MPCDF)
Accelerated
195 GB of memory per TFlop/s, 2013
ARCHER
CPU-only
206 GB of memory per TFlop/s, 2013
Iridis 4
Accelerated
220 GB of memory per TFlop/s, 2013
OCuLUS (Paderborn)
CPU-only
238 GB of memory per TFlop/s, 2013
Miztli
CPU-only
488 GB of memory per TFlop/s, 2013
Cypress (Tulane)
Accelerated
35.9 GB of memory per TFlop/s, 2014
Lobachevsky
Accelerated
44.2 GB of memory per TFlop/s, 2014
Magnolia (Kyoto)
CPU-only
86.6 GB of memory per TFlop/s, 2014
L-CSC
Accelerated
129 GB of memory per TFlop/s, 2014
Shadow (Mississippi State)
Accelerated
155 GB of memory per TFlop/s, 2014
ForHLR I
CPU-only
199 GB of memory per TFlop/s, 2014
Maverick (TACC)
Accelerated
214 GB of memory per TFlop/s, 2014
VSC-3
CPU-only
218 GB of memory per TFlop/s, 2014
ISM Supercomputer System I
CPU-only
278 GB of memory per TFlop/s, 2014
Piz Kesch
Accelerated
10.6 GB of memory per TFlop/s, 2015
Cherry Creek
Accelerated
31.7 GB of memory per TFlop/s, 2015
Plasma Simulator FX100
CPU-only
34.1 GB of memory per TFlop/s, 2015
SORA-MA
CPU-only
34.8 GB of memory per TFlop/s, 2015
BEM
CPU-only
66.8 GB of memory per TFlop/s, 2015
SuperMUC Phase 2
CPU-only
69.0 GB of memory per TFlop/s, 2015
JAEA SGI ICE X
CPU-only
81.3 GB of memory per TFlop/s, 2015
Beskow
CPU-only
86.9 GB of memory per TFlop/s, 2015
Salomon
Accelerated
90.6 GB of memory per TFlop/s, 2015
Galileo
Accelerated
96.5 GB of memory per TFlop/s, 2015
BURA
Accelerated
103 GB of memory per TFlop/s, 2015
CoolMUC-2
CPU-only
142 GB of memory per TFlop/s, 2015
Shaheen II
CPU-only
143 GB of memory per TFlop/s, 2015
Prometheus
Accelerated
170 GB of memory per TFlop/s, 2015
Tryton
CPU-only
216 GB of memory per TFlop/s, 2015
Swierk Computing Centre
CPU-only
283 GB of memory per TFlop/s, 2015
ARIS
Accelerated
317 GB of memory per TFlop/s, 2015
Sunway TaihuLight
CPU-only
14.1 GB of memory per TFlop/s, 2016
Phoenix (Adelaide)
Accelerated
48.4 GB of memory per TFlop/s, 2016
Marconi
Accelerated
56.3 GB of memory per TFlop/s, 2016
Camphor 2
Accelerated
64.3 GB of memory per TFlop/s, 2016
Cheyenne
CPU-only
65.4 GB of memory per TFlop/s, 2016
Oakforest-PACS
CPU-only
66.2 GB of memory per TFlop/s, 2016
Laconia
Accelerated
76.4 GB of memory per TFlop/s, 2016
ForHLR II
CPU-only
96.1 GB of memory per TFlop/s, 2016
Cori
Accelerated
99.1 GB of memory per TFlop/s, 2016
BinAC
Accelerated
114 GB of memory per TFlop/s, 2016
SciPhi-XVI
CPU-only
119 GB of memory per TFlop/s, 2016
TX-Green
CPU-only
120 GB of memory per TFlop/s, 2016
Quartz (LLNL)
CPU-only
128 GB of memory per TFlop/s, 2016
Laurel 2
CPU-only
129 GB of memory per TFlop/s, 2016
QPACE3
CPU-only
131 GB of memory per TFlop/s, 2016
Okeanos
CPU-only
153 GB of memory per TFlop/s, 2016
Draco
Accelerated
172 GB of memory per TFlop/s, 2016
Owens
Accelerated
180 GB of memory per TFlop/s, 2016
Finis Terrae II
Accelerated
209 GB of memory per TFlop/s, 2016
CARL and EDDY
CPU-only
214 GB of memory per TFlop/s, 2016
NEMO bwForCluster
CPU-only
219 GB of memory per TFlop/s, 2016
Reedbush-U
CPU-only
228 GB of memory per TFlop/s, 2016
Caliburn
CPU-only
232 GB of memory per TFlop/s, 2016
BrENIAC
CPU-only
312 GB of memory per TFlop/s, 2016
TSUBAME3.0
Accelerated
18.3 GB of memory per TFlop/s, 2017
Reedbush-L
Accelerated
19.8 GB of memory per TFlop/s, 2017
Reedbush-H
Accelerated
37.5 GB of memory per TFlop/s, 2017
Joliot-Curie KNL
CPU-only
70.8 GB of memory per TFlop/s, 2017
Joliot-Curie SKL
CPU-only
78.2 GB of memory per TFlop/s, 2017
Occigen2
CPU-only
80.3 GB of memory per TFlop/s, 2017
BeBop
CPU-only
86.4 GB of memory per TFlop/s, 2017
Trinity
Accelerated
105 GB of memory per TFlop/s, 2017
Serrano (Sandia)
CPU-only
115 GB of memory per TFlop/s, 2017
Cayenne (Sandia)
CPU-only
117 GB of memory per TFlop/s, 2017
IMS Molecular Simulator
CPU-only
121 GB of memory per TFlop/s, 2017
Sator
CPU-only
134 GB of memory per TFlop/s, 2017
SuperMUC-NG
CPU-only
36.9 GB of memory per TFlop/s, 2018
Stratus (NSC)
CPU-only
63.0 GB of memory per TFlop/s, 2018
Tetralith
Accelerated
72.9 GB of memory per TFlop/s, 2018
Gaffney
CPU-only
77.0 GB of memory per TFlop/s, 2018
Koehr
CPU-only
77.0 GB of memory per TFlop/s, 2018
PALMA II
CPU-only
90.5 GB of memory per TFlop/s, 2018
Aleph
CPU-only
92.7 GB of memory per TFlop/s, 2018
COBRA
Accelerated
94.3 GB of memory per TFlop/s, 2018
Lemhi
CPU-only
94.5 GB of memory per TFlop/s, 2018
Kay
Accelerated
96.2 GB of memory per TFlop/s, 2018
Electra
CPU-only
106 GB of memory per TFlop/s, 2018
Niagara
CPU-only
113 GB of memory per TFlop/s, 2018
Eclipse (Sandia)
CPU-only
140 GB of memory per TFlop/s, 2018
ISM SGI 8600
CPU-only
153 GB of memory per TFlop/s, 2018
ATERUI II
CPU-only
185 GB of memory per TFlop/s, 2018
Cygnus
Accelerated
10.4 GB of memory per TFlop/s, 2019
SWARM (Wuhan University)
Accelerated
11.4 GB of memory per TFlop/s, 2019
VSC-4
CPU-only
39.0 GB of memory per TFlop/s, 2019
Satori (MIT)
Accelerated
43.8 GB of memory per TFlop/s, 2019
NREL Eagle
CPU-only
61.0 GB of memory per TFlop/s, 2019
Oakbridge-CX
CPU-only
65.7 GB of memory per TFlop/s, 2019
Orion (MSU HPC2)
CPU-only
94.0 GB of memory per TFlop/s, 2019
MagicCube3
Accelerated
102 GB of memory per TFlop/s, 2019
Attaway
CPU-only
105 GB of memory per TFlop/s, 2019
Aitken
CPU-only
166 GB of memory per TFlop/s, 2019
Lyapunov
Accelerated
5.35 GB of memory per TFlop/s, 2020
Fugaku
CPU-only
12.4 GB of memory per TFlop/s, 2020
Raijin
Accelerated
25.6 GB of memory per TFlop/s, 2020
Hawk
CPU-only
51.7 GB of memory per TFlop/s, 2020
NYU Greene
Accelerated
69.4 GB of memory per TFlop/s, 2020
Lichtenberg II
CPU-only
81.6 GB of memory per TFlop/s, 2020
Expanse
Accelerated
88.7 GB of memory per TFlop/s, 2020
Cumulus Cascade Lake
CPU-only
91.5 GB of memory per TFlop/s, 2020
Wisteria/BDEC-01
CPU-only
11.9 GB of memory per TFlop/s, 2021
AlphaCentauri
Accelerated
15.1 GB of memory per TFlop/s, 2021
Dragão
Accelerated
22.3 GB of memory per TFlop/s, 2021
Tursa
Accelerated
34.4 GB of memory per TFlop/s, 2021
Karolina
Accelerated
37.0 GB of memory per TFlop/s, 2021
Vega
Accelerated
47.0 GB of memory per TFlop/s, 2021
Earth Simulator 4
Accelerated
61.2 GB of memory per TFlop/s, 2021
Narval
Accelerated
71.6 GB of memory per TFlop/s, 2021
Levante
CPU-only
84.3 GB of memory per TFlop/s, 2021
Ares
Accelerated
85.7 GB of memory per TFlop/s, 2021
Tryton Plus
CPU-only
96.0 GB of memory per TFlop/s, 2021
Setonix
Accelerated
17.0 GB of memory per TFlop/s, 2022
Lucia
Accelerated
44.5 GB of memory per TFlop/s, 2022
Hercules (MSU HPC2)
CPU-only
239 GB of memory per TFlop/s, 2022
Aurora
Accelerated
10.5 GB of memory per TFlop/s, 2023
AOBA-S
Accelerated
30.0 GB of memory per TFlop/s, 2023
Shaheen III
CPU-only
49.6 GB of memory per TFlop/s, 2023
Improv
CPU-only
93.9 GB of memory per TFlop/s, 2023
Nautilus
Accelerated
97.9 GB of memory per TFlop/s, 2023
ASPIRE 2A (GPU Partition)
Accelerated
143 GB of memory per TFlop/s, 2023
rzAdams
Accelerated
2.63 GB of memory per TFlop/s, 2024
El Capitan
Accelerated
3.01 GB of memory per TFlop/s, 2024
TSUBAME4.0
Accelerated
4.99 GB of memory per TFlop/s, 2024
Adastra 2
Accelerated
5.53 GB of memory per TFlop/s, 2024
Helios
Accelerated
20.5 GB of memory per TFlop/s, 2024
JMA PRIMERGY CX2550 M7
CPU-only
23.7 GB of memory per TFlop/s, 2024
AFI-NITY II
CPU-only
48.0 GB of memory per TFlop/s, 2024
Minerva
Accelerated
57.1 GB of memory per TFlop/s, 2024
Joule 3.0
Accelerated
99.0 GB of memory per TFlop/s, 2024
Viper-CPU
CPU-only
101 GB of memory per TFlop/s, 2024
Marvin
Accelerated
143 GB of memory per TFlop/s, 2024
Tuolumne
Accelerated
2.77 GB of memory per TFlop/s, 2025
SAKURAONE CYC
Accelerated
4.27 GB of memory per TFlop/s, 2025
Miyabi
Accelerated
5.17 GB of memory per TFlop/s, 2025
Nano 4 GB200
Accelerated
5.23 GB of memory per TFlop/s, 2025
VISION (TAMUS)
Accelerated
5.46 GB of memory per TFlop/s, 2025
SAKURAONE CYB
Accelerated
5.52 GB of memory per TFlop/s, 2025
AI-Farabium
Accelerated
5.58 GB of memory per TFlop/s, 2025
ABCI 3.0
Accelerated
10.6 GB of memory per TFlop/s, 2025
Torch
Accelerated
31.1 GB of memory per TFlop/s, 2025
Space HPC
Accelerated
34.9 GB of memory per TFlop/s, 2025
NASA Athena
CPU-only
49.2 GB of memory per TFlop/s, 2025
Rorqual
Accelerated
84.2 GB of memory per TFlop/s, 2025
Teton
CPU-only
37.9 GB of memory per TFlop/s, 2026
ASPIRE 2B (CPU Partition)
CPU-only
106 GB of memory per TFlop/s, 2026
rzAdams
Each dot is one system: total memory in GB divided by measured performance in TFlop/s, on a logarithmic axis. Divide by 1,000 for bytes per flop/s. The line joins the medians of groups with at least five systems. Table view: Memory per unit of measured performance First light Systems Median GB per TFlop/s Accelerated (n) CPU-only (n) Before 2000 14 432 - 432 (14) 2000 to 2004 18 759 - 759 (18) 2005 to 2009 42 276 102 (1) 279 (41) 2010 to 2014 67 155 102 (21) 188 (46) 2015 to 2019 78 95 91 (27) 105 (51) 2020 to 2024 39 48 36 (26) 82 (13) 2025 to date 14 8.1 5.5 (11) 49 (3)
How this was calculated
Memory per TFlop/s is the system's total memory in TB divided by its measured HPL Rmax in PFlop/s,
which is the same number as GB per TFlop/s. Storage per PFlop/s is the parallel storage capacity in
PB divided by Rmax in PFlop/s. Memory and storage come from each system's profile, each with its own
source; Rmax comes from the system's row. Systems are placed at their first-light year.
Systems recorded as upgraded are left out, as is Pleiades, because their Rmax belongs to a later and
larger configuration than the year they would be plotted at. A memory or storage figure stated for a
later configuration of a machine that is not marked upgraded will still be divided by that machine's
recorded Rmax; the group medians are robust to a few of these. “Accelerated” means the system has at
least one accelerator in the component graph. A least-squares fit of the logarithm of memory per TFlop/s
on year falls tenfold every 18 years (R² 0.54), but the medians in the table
show the fall is recent rather than steady.
Storage per unit of measured performance, by first-light year
Log-scale dot plot of each system’s parallel storage capacity divided by its measured HPL performance, by the year it first ran, coloured by whether it has accelerators, with a line through the group medians.
0.01 0.1 1 10 100 1000 1990 1995 2000 2005 2010 2015 2020 2025
median per group
KISTI Cray Y-MP C916
CPU-only
14.8 PB of storage per PFlop/s, 1993
CP-PACS
CPU-only
1.44 PB of storage per PFlop/s, 1996
KISTI Cray T3E-900
CPU-only
3.02 PB of storage per PFlop/s, 1997
Osaka University SX-4
CPU-only
21.1 PB of storage per PFlop/s, 1997
ASCI Blue-Pacific
CPU-only
37.3 PB of storage per PFlop/s, 1998
DWD Cray T3E1200
CPU-only
1.49 PB of storage per PFlop/s, 1999
ASCI White
CPU-only
32.4 PB of storage per PFlop/s, 2000
Bluesky
CPU-only
9.80 PB of storage per PFlop/s, 2002
Cheetah (ORNL)
CPU-only
17.3 PB of storage per PFlop/s, 2002
Earth Simulator
CPU-only
19.5 PB of storage per PFlop/s, 2002
CPTEC NEC SX-6
CPU-only
63.1 PB of storage per PFlop/s, 2002
Phoenix (ORNL)
CPU-only
2.14 PB of storage per PFlop/s, 2003
Weta Digital BladeCenter Cluster
CPU-only
34.2 PB of storage per PFlop/s, 2003
EMSL MPP2
CPU-only
52.1 PB of storage per PFlop/s, 2003
Thunder
CPU-only
9.53 PB of storage per PFlop/s, 2004
DataStar
CPU-only
22.0 PB of storage per PFlop/s, 2004
Topdawg
CPU-only
2.57 PB of storage per PFlop/s, 2005
KMA Cray X1E
CPU-only
5.60 PB of storage per PFlop/s, 2005
blueice
CPU-only
13.5 PB of storage per PFlop/s, 2006
Tera-10
CPU-only
18.9 PB of storage per PFlop/s, 2006
TSUBAME1.0
CPU-only
28.8 PB of storage per PFlop/s, 2006
Kan Balam
CPU-only
31.4 PB of storage per PFlop/s, 2006
HECToR
CPU-only
1.51 PB of storage per PFlop/s, 2007
Abe
CPU-only
6.38 PB of storage per PFlop/s, 2007
HLRB-II
CPU-only
11.7 PB of storage per PFlop/s, 2007
Finis Terrae
CPU-only
27.8 PB of storage per PFlop/s, 2007
Chebyshev
CPU-only
1.27 PB of storage per PFlop/s, 2008
Netuno
CPU-only
2.46 PB of storage per PFlop/s, 2008
Galera
CPU-only
2.82 PB of storage per PFlop/s, 2008
Jade (CINES)
CPU-only
3.89 PB of storage per PFlop/s, 2008
VIP (Garching)
CPU-only
4.98 PB of storage per PFlop/s, 2008
Rostand
CPU-only
9.43 PB of storage per PFlop/s, 2008
T2K Todai
CPU-only
12.1 PB of storage per PFlop/s, 2008
Tachyon
CPU-only
12.2 PB of storage per PFlop/s, 2008
T2K Kyodai
CPU-only
15.8 PB of storage per PFlop/s, 2008
Huygens
CPU-only
19.1 PB of storage per PFlop/s, 2008
Akka
CPU-only
21.7 PB of storage per PFlop/s, 2008
Lomonosov
Accelerated
1.94 PB of storage per PFlop/s, 2009
Tachyon II
CPU-only
3.86 PB of storage per PFlop/s, 2009
GPC (SciNet)
CPU-only
5.35 PB of storage per PFlop/s, 2009
Vayu (NCI)
CPU-only
6.33 PB of storage per PFlop/s, 2009
Plasma Simulator SR16000
CPU-only
7.91 PB of storage per PFlop/s, 2009
KISTI GAIA
CPU-only
14.4 PB of storage per PFlop/s, 2009
Nagoya FX1
CPU-only
40.3 PB of storage per PFlop/s, 2009
JSS FX1
CPU-only
99.5 PB of storage per PFlop/s, 2009
Tianhe-1A
Accelerated
0.779 PB of storage per PFlop/s, 2010
Millipede
CPU-only
4.04 PB of storage per PFlop/s, 2010
Supernova
CPU-only
4.30 PB of storage per PFlop/s, 2010
LOEWE-CSC
Accelerated
7.92 PB of storage per PFlop/s, 2010
Tera-100
CPU-only
19.0 PB of storage per PFlop/s, 2010
CHEOPS (Cologne)
CPU-only
28.7 PB of storage per PFlop/s, 2010
Magellan NERSC
CPU-only
34.2 PB of storage per PFlop/s, 2010
VSC-2
CPU-only
0.471 PB of storage per PFlop/s, 2011
Dawson2 (UCLA)
Accelerated
2.28 PB of storage per PFlop/s, 2011
K computer
CPU-only
3.90 PB of storage per PFlop/s, 2011
Lonestar 4
CPU-only
4.77 PB of storage per PFlop/s, 2011
ALPS (NCHC)
CPU-only
6.06 PB of storage per PFlop/s, 2011
Cielo
CPU-only
9.01 PB of storage per PFlop/s, 2011
RWTH Compute Cluster 2011
CPU-only
13.6 PB of storage per PFlop/s, 2011
HA-PACS
Accelerated
1.20 PB of storage per PFlop/s, 2012
Virgo (IIT Madras)
CPU-only
2.30 PB of storage per PFlop/s, 2012
Abel (UiO)
CPU-only
2.46 PB of storage per PFlop/s, 2012
Oakleaf-FX
CPU-only
3.07 PB of storage per PFlop/s, 2012
Sequoia
CPU-only
3.20 PB of storage per PFlop/s, 2012
Kyushu FX10
CPU-only
3.46 PB of storage per PFlop/s, 2012
SuperMUC
CPU-only
4.14 PB of storage per PFlop/s, 2012
CRIEPI SGI ICE X
CPU-only
4.95 PB of storage per PFlop/s, 2012
Fermi
CPU-only
5.59 PB of storage per PFlop/s, 2012
Oakley
Accelerated
7.56 PB of storage per PFlop/s, 2012
Raijin (NCI)
CPU-only
10.2 PB of storage per PFlop/s, 2012
Yellowstone
CPU-only
12.7 PB of storage per PFlop/s, 2012
Gordon
CPU-only
18.3 PB of storage per PFlop/s, 2012
Altamira (Cantabria)
Accelerated
26.9 PB of storage per PFlop/s, 2012
Olympus (PNNL)
CPU-only
39.1 PB of storage per PFlop/s, 2012
NIG Supercomputer 2012
CPU-only
60.3 PB of storage per PFlop/s, 2012
PARAM Yuva II
Accelerated
0.515 PB of storage per PFlop/s, 2013
ATERUI I
CPU-only
1.02 PB of storage per PFlop/s, 2013
Cascade (PNNL)
Accelerated
1.06 PB of storage per PFlop/s, 2013
OCuLUS (Paderborn)
CPU-only
2.65 PB of storage per PFlop/s, 2013
ARCHER
CPU-only
2.68 PB of storage per PFlop/s, 2013
Stampede
Accelerated
2.71 PB of storage per PFlop/s, 2013
Hydra (MPCDF)
Accelerated
3.90 PB of storage per PFlop/s, 2013
Fionn
Accelerated
3.99 PB of storage per PFlop/s, 2013
QUARTETTO
Accelerated
4.03 PB of storage per PFlop/s, 2013
Iridis 4
Accelerated
4.40 PB of storage per PFlop/s, 2013
Spirit (AFRL)
CPU-only
4.75 PB of storage per PFlop/s, 2013
Blue Brain IV
CPU-only
5.59 PB of storage per PFlop/s, 2013
Lichtenberg I
Accelerated
5.93 PB of storage per PFlop/s, 2013
Miztli
CPU-only
8.13 PB of storage per PFlop/s, 2013
TEIDE-HPC
CPU-only
9.49 PB of storage per PFlop/s, 2013
Lobachevsky
Accelerated
0.415 PB of storage per PFlop/s, 2014
Lomonosov-2
Accelerated
0.783 PB of storage per PFlop/s, 2014
Zenobe
CPU-only
1.05 PB of storage per PFlop/s, 2014
COMA (PACS-IX)
Accelerated
2.01 PB of storage per PFlop/s, 2014
Cypress (Tulane)
Accelerated
4.53 PB of storage per PFlop/s, 2014
Edison
CPU-only
4.57 PB of storage per PFlop/s, 2014
Deepthought2
Accelerated
6.11 PB of storage per PFlop/s, 2014
ISM Supercomputer System I
CPU-only
6.94 PB of storage per PFlop/s, 2014
Camellia (Kyoto)
Accelerated
7.88 PB of storage per PFlop/s, 2014
Magnolia (Kyoto)
CPU-only
9.77 PB of storage per PFlop/s, 2014
Avitohol
Accelerated
0.363 PB of storage per PFlop/s, 2015
Galileo
Accelerated
1.46 PB of storage per PFlop/s, 2015
Salomon
Accelerated
1.46 PB of storage per PFlop/s, 2015
BEM
CPU-only
1.58 PB of storage per PFlop/s, 2015
HOKUSAI GreatWave
Accelerated
2.12 PB of storage per PFlop/s, 2015
JAEA SGI ICE X
CPU-only
2.59 PB of storage per PFlop/s, 2015
Beskow
CPU-only
2.78 PB of storage per PFlop/s, 2015
Plasma Simulator FX100
CPU-only
4.21 PB of storage per PFlop/s, 2015
Prometheus
Accelerated
5.99 PB of storage per PFlop/s, 2015
DKRZ Mistral
CPU-only
6.64 PB of storage per PFlop/s, 2015
Swierk Computing Centre
CPU-only
6.85 PB of storage per PFlop/s, 2015
Hummel
CPU-only
9.20 PB of storage per PFlop/s, 2015
Aziz
CPU-only
9.47 PB of storage per PFlop/s, 2015
Shaheen II
CPU-only
9.57 PB of storage per PFlop/s, 2015
Sango
Accelerated
11.9 PB of storage per PFlop/s, 2015
Shirokane3
Accelerated
39.4 PB of storage per PFlop/s, 2015
NEMO bwForCluster
CPU-only
1.43 PB of storage per PFlop/s, 2016
Marconi
Accelerated
1.64 PB of storage per PFlop/s, 2016
Draco
Accelerated
1.90 PB of storage per PFlop/s, 2016
Oakforest-PACS
CPU-only
1.92 PB of storage per PFlop/s, 2016
Cobalt (CCRT)
Accelerated
1.92 PB of storage per PFlop/s, 2016
CARL and EDDY
CPU-only
2.03 PB of storage per PFlop/s, 2016
BinAC
Accelerated
2.15 PB of storage per PFlop/s, 2016
GOSAT-2 RCF2
Accelerated
2.60 PB of storage per PFlop/s, 2016
Met Office Cray XC40
CPU-only
2.84 PB of storage per PFlop/s, 2016
Finis Terrae II
Accelerated
3.52 PB of storage per PFlop/s, 2016
Lengau
CPU-only
4.85 PB of storage per PFlop/s, 2016
Lonestar 5
CPU-only
4.90 PB of storage per PFlop/s, 2016
Camphor 2
Accelerated
6.14 PB of storage per PFlop/s, 2016
ECMWF Cray XC40
CPU-only
8.27 PB of storage per PFlop/s, 2016
Cheyenne
CPU-only
8.35 PB of storage per PFlop/s, 2016
TX-Green
CPU-only
9.68 PB of storage per PFlop/s, 2016
Reedbush-U
CPU-only
11.0 PB of storage per PFlop/s, 2016
ASPIRE 1
CPU-only
12.9 PB of storage per PFlop/s, 2016
Laurel 2
CPU-only
29.2 PB of storage per PFlop/s, 2016
Biowulf
CPU-only
38.1 PB of storage per PFlop/s, 2016
Piz Daint
Accelerated
0.415 PB of storage per PFlop/s, 2017
Joliot-Curie SKL
CPU-only
1.23 PB of storage per PFlop/s, 2017
TSUBAME3.0
Accelerated
1.96 PB of storage per PFlop/s, 2017
Occigen2
CPU-only
2.05 PB of storage per PFlop/s, 2017
Mogon II
CPU-only
3.81 PB of storage per PFlop/s, 2017
Joliot-Curie KNL
CPU-only
3.82 PB of storage per PFlop/s, 2017
Trinity
Accelerated
3.87 PB of storage per PFlop/s, 2017
AIST AI Cloud
Accelerated
4.16 PB of storage per PFlop/s, 2017
Reedbush-L
Accelerated
6.22 PB of storage per PFlop/s, 2017
Reedbush-H
Accelerated
6.30 PB of storage per PFlop/s, 2017
Raad2
Accelerated
6.67 PB of storage per PFlop/s, 2017
Cedar (SFU)
Accelerated
8.51 PB of storage per PFlop/s, 2017
COBRA
Accelerated
0.891 PB of storage per PFlop/s, 2018
Lemhi
CPU-only
1.20 PB of storage per PFlop/s, 2018
HPC4
Accelerated
1.23 PB of storage per PFlop/s, 2018
Kay
Accelerated
1.50 PB of storage per PFlop/s, 2018
Nurion
CPU-only
1.53 PB of storage per PFlop/s, 2018
Summit
Accelerated
1.68 PB of storage per PFlop/s, 2018
Stratus (NSC)
CPU-only
2.20 PB of storage per PFlop/s, 2018
Niagara
CPU-only
2.50 PB of storage per PFlop/s, 2018
Taiwania
Accelerated
2.57 PB of storage per PFlop/s, 2018
Mustang (AFRL)
Accelerated
2.61 PB of storage per PFlop/s, 2018
Gaffney
CPU-only
2.80 PB of storage per PFlop/s, 2018
Koehr
CPU-only
2.80 PB of storage per PFlop/s, 2018
ATERUI II
CPU-only
3.11 PB of storage per PFlop/s, 2018
ISM SGI 8600
CPU-only
3.82 PB of storage per PFlop/s, 2018
CLAIX-2018
Accelerated
4.03 PB of storage per PFlop/s, 2018
ITO Subsystem A
CPU-only
5.29 PB of storage per PFlop/s, 2018
Grand Chariot
CPU-only
8.00 PB of storage per PFlop/s, 2018
Aleph
CPU-only
9.88 PB of storage per PFlop/s, 2018
AFI-NITY
CPU-only
11.2 PB of storage per PFlop/s, 2018
NVIDIA DGX SuperPOD (V100)
Accelerated
0.0585 PB of storage per PFlop/s, 2019
Jean Zay V100
Accelerated
0.223 PB of storage per PFlop/s, 2019
Cygnus
Accelerated
1.58 PB of storage per PFlop/s, 2019
Frontera
CPU-only
1.87 PB of storage per PFlop/s, 2019
Pangea III
Accelerated
2.80 PB of storage per PFlop/s, 2019
Puhti
Accelerated
2.81 PB of storage per PFlop/s, 2019
NREL Eagle
CPU-only
2.89 PB of storage per PFlop/s, 2019
Oakbridge-CX
CPU-only
2.89 PB of storage per PFlop/s, 2019
MagicCube3
Accelerated
3.79 PB of storage per PFlop/s, 2019
Hera
CPU-only
6.01 PB of storage per PFlop/s, 2019
CARA
Accelerated
9.77 PB of storage per PFlop/s, 2019
Béluga
Accelerated
12.2 PB of storage per PFlop/s, 2019
Betzy
Accelerated
1.63 PB of storage per PFlop/s, 2020
Grace (TAMU HPRC)
Accelerated
1.89 PB of storage per PFlop/s, 2020
Hawk
CPU-only
2.17 PB of storage per PFlop/s, 2020
Toubkal
CPU-only
2.53 PB of storage per PFlop/s, 2020
Dogwood
CPU-only
2.60 PB of storage per PFlop/s, 2020
Raijin
Accelerated
4.07 PB of storage per PFlop/s, 2020
NYU Greene
Accelerated
4.31 PB of storage per PFlop/s, 2020
Flow
CPU-only
4.53 PB of storage per PFlop/s, 2020
Freeman (ERDC)
CPU-only
5.30 PB of storage per PFlop/s, 2020
davinci-1
Accelerated
5.80 PB of storage per PFlop/s, 2020
Flow Type II
Accelerated
6.15 PB of storage per PFlop/s, 2020
Expanse
Accelerated
7.66 PB of storage per PFlop/s, 2020
IARA
Accelerated
0.410 PB of storage per PFlop/s, 2021
Berzelius
Accelerated
0.571 PB of storage per PFlop/s, 2021
ARCHER2
CPU-only
0.737 PB of storage per PFlop/s, 2021
Raven-GPU
Accelerated
0.812 PB of storage per PFlop/s, 2021
Topaze
Accelerated
0.920 PB of storage per PFlop/s, 2021
Fritz
CPU-only
0.978 PB of storage per PFlop/s, 2021
Wisteria/BDEC-01
CPU-only
1.17 PB of storage per PFlop/s, 2021
Tursa
Accelerated
1.53 PB of storage per PFlop/s, 2021
Discoverer
Accelerated
1.65 PB of storage per PFlop/s, 2021
MeluXina
CPU-only
1.90 PB of storage per PFlop/s, 2021
ABCI 2.0
Accelerated
2.12 PB of storage per PFlop/s, 2021
Baskerville
Accelerated
2.35 PB of storage per PFlop/s, 2021
CARO
CPU-only
2.43 PB of storage per PFlop/s, 2021
Cactus
CPU-only
2.60 PB of storage per PFlop/s, 2021
Vega
Accelerated
2.89 PB of storage per PFlop/s, 2021
Narval
Accelerated
4.24 PB of storage per PFlop/s, 2021
Earth Simulator 4
Accelerated
6.01 PB of storage per PFlop/s, 2021
Levante
CPU-only
12.9 PB of storage per PFlop/s, 2021
Setonix
Accelerated
0.515 PB of storage per PFlop/s, 2022
Adastra
Accelerated
0.564 PB of storage per PFlop/s, 2022
Lucia
Accelerated
1.10 PB of storage per PFlop/s, 2022
Noctua 2
Accelerated
1.43 PB of storage per PFlop/s, 2022
LANTA
Accelerated
1.47 PB of storage per PFlop/s, 2022
Delta
Accelerated
2.49 PB of storage per PFlop/s, 2022
Anvil
CPU-only
2.81 PB of storage per PFlop/s, 2022
Komondor
Accelerated
3.23 PB of storage per PFlop/s, 2022
fortytwo
Accelerated
0.151 PB of storage per PFlop/s, 2023
Aurora
Accelerated
0.217 PB of storage per PFlop/s, 2023
AOBA-S
Accelerated
0.261 PB of storage per PFlop/s, 2023
Shaheen III
CPU-only
0.897 PB of storage per PFlop/s, 2023
MareNostrum 5 ACC
Accelerated
1.79 PB of storage per PFlop/s, 2023
Lichtenberg II Phase 2
Accelerated
1.94 PB of storage per PFlop/s, 2023
HEMUS
Accelerated
2.66 PB of storage per PFlop/s, 2023
Deucalion
CPU-only
2.68 PB of storage per PFlop/s, 2023
JMA PRIMEHPC FX1000
CPU-only
3.16 PB of storage per PFlop/s, 2023
Nautilus
Accelerated
6.67 PB of storage per PFlop/s, 2023
ASPIRE 2A (GPU Partition)
Accelerated
10.5 PB of storage per PFlop/s, 2023
ASPIRE 2A (CPU Partition)
CPU-only
13.6 PB of storage per PFlop/s, 2023
MUSICA
Accelerated
0.126 PB of storage per PFlop/s, 2024
Alps
Accelerated
0.253 PB of storage per PFlop/s, 2024
Jean Zay H100
Accelerated
0.830 PB of storage per PFlop/s, 2024
Helios
Accelerated
0.914 PB of storage per PFlop/s, 2024
Gautschi AI
Accelerated
0.964 PB of storage per PFlop/s, 2024
amplitUDE
Accelerated
1.08 PB of storage per PFlop/s, 2024
TSUBAME4.0
Accelerated
1.12 PB of storage per PFlop/s, 2024
ASPIRE 2A+
Accelerated
2.11 PB of storage per PFlop/s, 2024
Forerunner 1
CPU-only
2.61 PB of storage per PFlop/s, 2024
Viper-CPU
CPU-only
2.80 PB of storage per PFlop/s, 2024
Joule 3.0
Accelerated
4.01 PB of storage per PFlop/s, 2024
Arunika
CPU-only
4.04 PB of storage per PFlop/s, 2024
Minerva
Accelerated
4.05 PB of storage per PFlop/s, 2024
AFI-NITY II
CPU-only
6.74 PB of storage per PFlop/s, 2024
JUPITER
Accelerated
0.0200 PB of storage per PFlop/s, 2025
Isambard-AI Phase 2
Accelerated
0.0924 PB of storage per PFlop/s, 2025
VISION (TAMUS)
Accelerated
0.230 PB of storage per PFlop/s, 2025
Miyabi
Accelerated
0.241 PB of storage per PFlop/s, 2025
AI-Farabium
Accelerated
0.335 PB of storage per PFlop/s, 2025
Viper-GPU
Accelerated
0.386 PB of storage per PFlop/s, 2025
Olivia
Accelerated
0.402 PB of storage per PFlop/s, 2025
ABCI 3.0
Accelerated
0.517 PB of storage per PFlop/s, 2025
Hunter
Accelerated
0.789 PB of storage per PFlop/s, 2025
Space HPC
Accelerated
0.805 PB of storage per PFlop/s, 2025
Otus
Accelerated
1.16 PB of storage per PFlop/s, 2025
Fir
Accelerated
2.51 PB of storage per PFlop/s, 2025
Trillium
Accelerated
3.06 PB of storage per PFlop/s, 2025
Nibi
Accelerated
3.13 PB of storage per PFlop/s, 2025
Nano 4 GB200
Accelerated
4.84 PB of storage per PFlop/s, 2025
Rorqual
Accelerated
9.90 PB of storage per PFlop/s, 2025
Arrhenius
Accelerated
0.434 PB of storage per PFlop/s, 2026
ASPIRE 2B (CPU Partition)
CPU-only
6.29 PB of storage per PFlop/s, 2026
JUPITER NVIDIA DGX SuperPOD (V100) Isambard-AI Phase 2
Each dot is one system: parallel storage in PB divided by measured performance in PFlop/s, on a logarithmic axis. The line joins the medians of groups with at least five systems. Table view: Storage per unit of measured performance First light Systems Median PB per PFlop/s Accelerated (n) CPU-only (n) Before 2000 6 8.9 - 8.9 (6) 2000 to 2004 10 21 - 21 (10) 2005 to 2009 29 9.4 1.9 (1) 11 (28) 2010 to 2014 55 4.5 3.9 (20) 4.9 (35) 2015 to 2019 79 2.9 2.6 (37) 3.8 (42) 2020 to 2024 64 2.1 1.6 (43) 2.6 (21) 2025 to date 18 0.65 0.52 (17) 6.3 (1)
By workload, for machines first run since 2020 Workload Systems GB memory per TFlop/s Scientific computing (HPC) 27 52 Mixed HPC and AI 22 26 AI training 3 5.6 Classified 1 3.0
Workload Systems PB storage per PFlop/s Scientific computing (HPC) 40 2.5 Mixed HPC and AI 33 1.8 AI training 9 0.52
Medians. Groups with only a handful of systems are shown for completeness and say little on their own.
The leanest machines The read
An old rule of thumb asked for a byte of memory for every flop/s. The median machine in this dataset
was already below that before 2000, at 0.43 bytes, and the gap has widened: the median machine first run in
2025 to date has about 94 times less memory per flop than the median
machine of 2000 to 2004. Most of that fall is recent. The table shows the medians drifting down slowly
through the 2000s and 2010s and dropping sharply among the machines first run in the 2020s, which are
mostly accelerated.
Accelerators explain much of it. In the 2025 to date group, CPU-only machines carry 8.9 times as much memory per flop as accelerated ones (49 against 5.5 GB per TFlop/s). The leanest machines on record are rzAdams, Tuolumne, El Capitan, all accelerated systems built around fast memory beside each processor rather than a large pool. For a code that needs to hold a large problem in memory, a low ratio is a real constraint, whatever the
machine's speed.
The storage ratio says something similar about the AI build-out. Training clusters move enormous
volumes of data but keep comparatively little parallel storage beside each flop:
among machines first run since 2020, the AI training systems here have a median of 0.52 PB per PFlop/s against 2.5 for scientific systems. Most large AI clusters record neither a measured Rmax nor a storage total, so this comparison rests on
the minority that do.