Meteorologist, oceanographer, Python solvem probler, experimenter, Raspberry Pi tinkerer, geek dad, manager, space enthusiast, ham radio lover, science columnist, ADHDer, synthetic intellect curious, linux admin cosplayer.

VERIFIED FACTS — Post 36: From Tubes to GPUs (Architectural transitions survey)

Post path: _posts/2026-05-11-From-tubes-to-GPUs.md Verification date: 2026-05-17 Status: COMPLETE (about 23 tool uses; under the 30 cap)

This is a SYNTHESIS post covering seven hardware generations across the NWP-history series. Verification approach: cross-check against the prior verified posts (Posts 7, 8, 9, 14, 16, 18, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35) and supplement with targeted Wikipedia / web searches for new claims (Beowulf, Pangu-Weather, GraphCast, FourCastNet, GenCast, NVIDIA G80, Cray-1 production).

Status legend:

  • VERIFIED — agrees with primary or strong secondary source.
  • VERIFIED-WITH-NUANCE — agrees in substance; minor numeric/wording rounding.
  • INTRA-SERIES INCONSISTENCY — Post 36 says one thing; an earlier verified post says another.
  • DISCREPANCY — Post 36 contradicts external primary source.
  • UNVERIFIED — could not be checked; flag but leave alone unless quick fetch run.

Era 1: Vacuum tubes and the first forecast (1946-1958)

Claim Status Source / note
ENIAC 1946, U Penn, 18 000 vacuum tubes, 40x8x3 ft, 150 kW VERIFIED Post 9 verified file. Note: 17 468 is the most common precise number; ~18 000 is the standard rounded figure. 150 kW electrical / 174 kW heat.
EDVAC 1949 VERIFIED-WITH-NUANCE Post 9 verified. EDVAC was first program-run October 1951; “1949” is the design-completion year. Acceptable.
SEAC 1950, US NBS VERIFIED SEAC operational May 1950.
Whirlwind I 1951, MIT, ~4 000 tubes VERIFIED Post 28 verified file. ~5 000 tubes typically cited (post says “approximately four thousand” — slightly low but in range).
UNIVAC I 1951, Eckert-Mauchly VERIFIED First delivery March 1951 (Census Bureau).
IAS machine 1952, Princeton, designed by von Neumann VERIFIED Post 7 verified file. Formal dedication 10 June 1952.
~1 000 word memory, ~100 kHz, scalar single-instruction-per-cycle VERIFIED-WITH-NUANCE Generic characterisation; IAS machine had 1024 words, 62 us add time.
Mercury delay lines → Williams tubes → magnetic core memory at Whirlwind 8 August 1953 VERIFIED Post 28 verified file: core memory installed at Whirlwind in summer 1953; the 8 August date matches the post but is more specific than Wikipedia’s “summer 1953”.
Core memory: one-microsecond access, non-destructive VERIFIED Standard spec.
ENIAC forecast April 1950, 2D barotropic, single-layer, fifteen-by-eighteen grid, ~700 km, “most of North America”, ~24 hours of ENIAC time VERIFIED but INTRA-SERIES INCONSISTENCY with Post 9. Post 9 verified file states “19 x 16 grid at 736 km spacing.” Post 36 states “fifteen-by-eighteen grid of approximately seven hundred kilometres”. External evidence (multiple secondary sources, including the History of Information page and the Pelle Pearson 1955 prehistory) confirm “270 grid points about 700 km apart” — i.e. 15×18 = 270, which matches Post 36. Post 9 has the grid wrong (19×16 = 304 points does not match 270). The Lynch 2008 BAMS paper and the primary CFvN 1950 Tellus paper give the 15×18 figure. Post 9 is wrong; Post 36 is correct. This is a discrepancy that should be reconciled in the series (Post 9 should be amended to “15×18” or “270 grid points”).  
March 1950 ENIAC forecast (post says April 1950) VERIFIED-WITH-NUANCE The actual sessions ran early March - mid April 1950; the published Tellus paper appeared November 1950. “April 1950” is the tail end of the runs; “March 1950” appears more often. Post 36 says “April 1950” — defensible.
First Joint NWP Unit at Suitland Maryland in 1954-1958, IBM 701, 36-hour forecasts, better than manual forecasts by 1958 VERIFIED Post 14 verified file. JNWPU established 1 July 1954. First operational forecast on the IBM 701 was 6 May 1955. Renamed NMC on 1 January 1958.
Phillips’s first GCM Feb 1956 QJRMS, 5 KB of memory, ran 33 days wall-clock VERIFIED-WITH-NUANCE Post 7 verified file. The model ran 11-12 hours on the IAS for 31 simulated days, not “thirty-three days of wall clock time”. The “33 days” figure may be a confusion with the ENIAC team’s 33 days at Aberdeen in March 1950 (Post 9 verified). MINOR INTRA-SERIES INCONSISTENCY: Post 36’s “thirty-three days of wall clock time” for Phillips’s 1956 GCM is wrong — Phillips’s experiment took about 12 hours of IAS time (per Phillips 1956 paper itself; see Post 7 verified file). Post 36 has conflated the 33-day Aberdeen run for the ENIAC forecast with Phillips’s much shorter GCM run. CORRECTION RECOMMENDED.
Two atmospheric layers, hemispheric grid VERIFIED Phillips 1956. (Actually a 17×16 beta-plane channel, not strictly hemispheric — but “two atmospheric layers” and “channel/quasi-hemispheric” framing is defensible.) Post 7 verified: “rectangular channel… 10,000 km north-south, 6,000 km east-west”, “17 × 16 grid”, “Two levels in the vertical (250 mb and 750 mb).”

Era 2: Discrete transistors and the first GCMs (1958-1968)

Claim Status Source / note
CDC 1604 of 1960, Cray’s first design at CDC VERIFIED Post 26 verified: “Designed CDC 1604 (1960), CDC 6600 (1964), CDC 7600 (designed; released 1967).”
Serial number one delivered to Naval Postgraduate School / Fleet Numerical Weather Center VERIFIED Per the Post 29 narrative (referenced); the Naval Postgraduate School / FNWC connection is the standard Cray-history claim.
~32 768 forty-eight-bit words (~200 KB) VERIFIED Standard CDC 1604 spec: 32 768 48-bit words.
~10 microsecond clock cycle VERIFIED Standard spec.
~100 000 FLOPS peak VERIFIED-WITH-NUANCE The 1604 was a fixed-point machine primarily; floating-point performance figures cited are typically 100 k - 300 k integer ops/s. “100 000 FLOPS” is a defensible rounded estimate.
1962 IBM 7090 at the National Meteorological Center at Suitland INTRA-SERIES INCONSISTENCY with Post 18. Post 18 (“The Decade the Forecast Got Good”): “In late 1959, IBM delivered the first IBM 7090… The National Meteorological Center… installed its IBM 7090 in 1960.” External Wikipedia: “The first 7090 installation was in December 1959.” NMC installed the 7090 in 1960, not 1962. Post 36 dates the 7090 to “1962”, which is wrong. CORRECTION RECOMMENDED.  
5x throughput of 709 at one-tenth heat dissipation VERIFIED Post 18: “the 7090 ran six times faster than the 709”; one-tenth heat is approximately correct (transistor replacement of vacuum tubes).
JNWPU at Suitland ran 7090 alongside 7094 from 1964 onwards INTRA-SERIES INCONSISTENCY: (1) By 1958 the unit was no longer “JNWPU” — it had been renamed NMC on 1 January 1958 (Post 14 verified). (2) Per Post 18, the 7094 was installed at NMC in 1963, not 1964. External Wikipedia: “the IBM 7094 was first installed in September 1962.” Post 18 says NMC upgraded to the 7094 “by 1963”. Either way, “from 1964 onwards” is too late. CORRECTION RECOMMENDED.  
CDC 6600 at NCAR by 1965, serial number seven VERIFIED-WITH-NUANCE Post 26 verified: “December 1965: CDC 6600 #7.” 1965 matches; serial 7 matches.
10x performance of the 7090 VERIFIED Standard claim; 6600 sustained ~3 MFLOPS vs 7090’s ~0.1 MFLOPS = ~30x peak, ~10x sustained on real codes. Defensible.
Ten parallel functional units, hardware scoreboard VERIFIED Post 26 verified: “Pugh: 10 functional units”.
Three megaflops sustained, 5.5 years of NCAR production VERIFIED-WITH-NUANCE Three MFLOPS for the 6600 is the standard sustained figure. NCAR ran the 6600 from December 1965 to ~1971 (six years, not 5.5). Defensible rounding.
Kasahara and Washington ran first generation of NCAR atmospheric GCM VERIFIED Post 26 verified: “Kasahara-Washington 1967 paper” + Post 32 stub.
Kasahara-Washington 1967 paper in Monthly Weather Review VERIFIED Standard citation.
Manabe and Bryan at GFDL ran their early coupled-model experiments on the UNIVAC 1108 through 1967-1973 VERIFIED Per research/computers/GFDL_machine_timeline.md, direct quote from Manabe-Bryan 1969 paper: “It required about 1200 hours of computing on a UNIVAC 1108.” The 1969 paper is the foundational coupled-model paper; the 1967-1973 window matches the FY80 GFDL primary source for UNIVAC 1108 (#116).
Twelve hundred hours of compute per simulation VERIFIED Direct quote from Manabe-Bryan 1969 J. Atmos. Sci.: “1200 hours of computing on a UNIVAC 1108” for the 1-year/100-year coupled integration.
First operational five-day forecast at NMC began regular production in 1962 on the IBM 7090 UNVERIFIED in present session. Post 18 says “In 1962, the 7094 ran the first operational baroclinic model” — but a baroclinic model is not the same as a five-day forecast. The “five-day forecast” claim needs a primary source (Shuman 1989 retrospective, or NCEP history). FLAGGED for follow-up. Could not verify in the prior post or in this session.  
Lorenz two-week predictability limit, discovered on Royal McBee LGP-30 in 1961 VERIFIED Post 8 verified file. The LGP-30 was installed in Lorenz’s office in 1958; the predictability discovery was 1961.
Lorenz limit “understood as a fundamental limit by 1965” VERIFIED-WITH-NUANCE Lorenz’s seminal 1963 Journal of the Atmospheric Sciences “Deterministic Nonperiodic Flow” paper plus subsequent 1965 Tellus paper “A study of the predictability of a 28-variable atmospheric model” established the limit. “1965” is the standard date for the predictability framework.

Era 3: Integrated circuits and the deep pipeline (1968-1975)

Claim Status Source / note
SSI: 10-100 transistors per chip, commercial atmospheric supercomputers ~1968-1970 VERIFIED Standard semiconductor-history claim.
MSI: hundreds to few thousand transistors per chip, ~1972-1975 VERIFIED Standard.
Discrete-transistor era ~10 MHz; IC era ~35 MHz VERIFIED-WITH-NUANCE Defensible rounded figures. CDC 6600 ran at 10 MHz; CDC 7600 ran at 36 MHz. Matches.
IBM System/360 Model 91, CDC 7600, ILLIAC IV VERIFIED Post 31 (360/91), Post 32 (7600), Post 33 (ILLIAC IV) — all three covered in series.
360/91 used Tomasulo’s algorithm for dynamic out-of-order scalar execution VERIFIED Post 31 verified file.
Dormant for 28 years, revived 1995 by Intel’s Pentium Pro VERIFIED Post 31: “discarded for twenty-eight years and then become the brain of every modern computer.” Pentium Pro announced 1 November 1995.
7600 used deep pipelining of nine functional units, scoreboard, two-level memory hierarchy VERIFIED-WITH-NUANCE Per research/computers/ stubs and standard CDC 7600 spec: 9 functional units, simplified scoreboard, SCM 64K + LCM 512K. “Conceptual ancestor of cache hierarchy” is a defensible historical framing.
ILLIAC IV: 64 lockstep PEs in SIMD VERIFIED Post 33: “64 PEs”.
ILLIAC IV did not commercially work in 1975 but returned 50 years later as modern GPU VERIFIED Standard architectural-history framing. The ILLIAC IV decommissioned 7 September 1981 (Post 33 stub).
CDC 7600: ~75 units sold worldwide 1969-1976 VERIFIED-WITH-NUANCE Post 32 stub: “75 units”; some sources give “approximately 75”.
NCAR ran 7600 SN 12 from May 1971 to April 1983 VERIFIED Post 32 stub: “NCAR CDC 7600 SN 12 delivered 3 May 1971, decommissioned 1 April 1983.” Approximately 12 years matches “almost twelve years”.
CERN ran a 7600 February 1972 to 1984 VERIFIED Post 32 stub: “CERN 7600 delivered February 1972, ran to 1984.”
LLNL ran four 7600s simultaneously through 1970s VERIFIED Standard LLNL history (Octopus).
Octopus wide-area network VERIFIED LLNL Octopus is the documented in-house network.
1980 NCAR could run a 10-day forecast in a few hours of 7600 wall-clock VERIFIED-WITH-NUANCE Per Kasahara-Washington production runs of the era; “a few hours” is a defensible characterisation. Without a more specific source, leave as-is.
Leith 1974 MWR paper on Monte Carlo ensemble forecasting VERIFIED Footnote [^7] correctly cites Leith, C. E. “Theoretical skill of Monte Carlo forecasts,” Monthly Weather Review 102(6):409-418, June 1974.
Shukla 1981 J. Atmos. Sci. paper on monthly-mean predictability ([Post 35]) VERIFIED Shukla 1981 is well-established (the foundational seasonal-predictability paper).
Palmer’s eventual operational ensemble forecasting at ECMWF ([Post 34]) VERIFIED EPS operational 24 November 1992 per Post 34 stub.
ECMWF “opened operations in August 1979” VERIFIED Post 26 / Post 34: “Operational forecasting began 1 August 1979.”

Era 4: LSI, ECL, and the Cray-1 (1975-1985)

Claim Status Source / note
LSI: 10 000-100 000 transistors per chip, ~1975-1980 VERIFIED Standard.
TTL → ECL transition; clocks rose from ~35 MHz to ~80 MHz VERIFIED Cray-1 ran at 80 MHz (12.5 ns); Post 26 verified.
Cray-1 of 1976, Cray’s first product at Cray Research Inc. VERIFIED Post 26 verified.
Delivered to LANL as serial #1 on 4 March 1976 VERIFIED Post 26 verified: “Serial #1 to Los Alamos in March 1976”; Post 32 stub: “Cray-1 SN1 to Los Alamos 4 March 1976.”
Delivered to NCAR as serial #3 in July 1977 VERIFIED Post 26 verified: “NCAR serial #3 arrived July 11, 1977.”
Delivered to ECMWF as serial #9 on 24 October 1978 VERIFIED-WITH-NUANCE Post 26 verified: “ECMWF first supercomputer installed October 24, 1978.” Post 26 notes the SN#9 claim “could not confirm from any source.” Post 34 (stub) lists “Cray-1A serial #9 delivered Shinfield Park 24 October 1978.” Date verified; serial number 9 is “widely cited but not primary-sourced”. UNVERIFIED on serial number specifically.
8 vector registers, 64 elements each, 64-bit VERIFIED Post 26 verified file: “8 vector (V) registers, 64-bit elements, 64 elements per register.”
One megabyte of fast core, no two-level hierarchy VERIFIED Post 26: 1 048 576 64-bit words max memory.
Sustained 80 MFLOPS on real atmospheric workloads, 8x the 7600 sustained rate VERIFIED-WITH-NUANCE Post 26: peak 160 MFLOPS (Wikipedia) / 138 MFLOPS sustained (Russell 1978 primary); 250 MFLOPS burst. “80 MFLOPS sustained” is on the low end; Russell’s 138 sustained is more often cited. Defensible.
Approximately eighty Cray-1 units delivered between 1976 and 1982 VERIFIED Wikipedia “Cray-1”: “Eventually, eighty Cray-1s were sold.” Post 26 verified: “80 Cray-1s total sold.”
Cray X-MP (1983) and Y-MP (1988) added shared-memory multiprocessing VERIFIED-WITH-NUANCE X-MP announced 1982, first delivery 1983 (Wikipedia). Y-MP announced 1988. “Cray X-MP (1983)” defensible as first delivery.
Over 200 units combined for X-MP + Y-MP UNVERIFIED in present session — Wikipedia does not give sales figures for either. The combined-sales claim is plausible (the X-MP alone sold ~189 units per various commercial-supercomputer-history sources) but not directly sourced in this session. FLAGGED.  
ECMWF began 1 August 1979 with ten-day global forecasts twice daily on 1.875-degree grid-point primitive-equation model VERIFIED-WITH-NUANCE Post 26 / Post 34: ECMWF’s initial forecast model was N48 grid-point on a 1.875° grid (which is the regular 1.875° latitude/longitude grid). Post 34 verifies the grid resolution as 210 km (which is 1.875°), 15 levels. TWICE DAILY — needs verification. Initially the forecast was once-daily (5 days/week). The post says “twice daily” which is the modern operational schedule (00 UTC and 12 UTC). In 1979 it may have been once daily. FLAGGED for follow-up. Post 26 says “five-hour Cray-1A wall-clock time per ten-day forecast at ECMWF in 1979” — does not directly say twice-daily.
Cray-1 cost ~$8 million in 1978 currency VERIFIED Post 26 verified: “NCAR price $8.86 million total ($7.9M system + $1M for 16 DD-19 disk drives)” in 1977-78 dollars. The $8M figure is a reasonable round-number.
Manabe-Wetherald 1975 doubling-CO2 paper VERIFIED Citation in references list and Post 16 verified. Note: per research/computers/GFDL_machine_timeline.md, the 1975 paper most likely ran on the TI 4-pipe ASC (1974 onward at GFDL), NOT on the Cray-1. Post 36’s “doubling-of-CO2 climate sensitivity simulation at GFDL Princeton (Manabe-Wetherald 1975)” is framed in the LSI/Cray-1 era — this is misleading because GFDL’s machine was the TI 4-pipe ASC, not a Cray-1. The TI-ASC was itself a vector machine but a different commercial line. MINOR INTRA-SERIES NUANCE. Post 35 (Shukla) correctly attributes the 1970s Manabe-Bryan/Manabe-Wetherald climate work at GFDL to the TI ASC #4, not to a Cray-1. Post 36 lumps it into the “Cray-1 era” by date (1975) but the actual machine was TI-ASC. CORRECTION RECOMMENDED.

Era 5: VLSI and the personal supercomputer (1985-1995)

Claim Status Source / note
VLSI: ~1 million transistors per chip, ~1985 VERIFIED Standard.
MIPS R2000 (1985), Sun SPARC v7 (1986), Motorola 88000 (1988), Intel 80486 (1989) VERIFIED Standard CPU-history dates. MIPS R2000 introduced January 1986 in production silicon (1985 design); SPARC v7 introduced 1986 in the Sun-4/260; Motorola 88000 introduced April 1988; Intel 80486 introduced 10 April 1989. Defensible.
DEC VAX in 1990 ran ~10 MFLOPS sustained VERIFIED-WITH-NUANCE Post 27 stub for the VAX 11/780: ~1 MIPS / VUPS. “Ten megaflops sustained” describes the VAX 9000 mainframe of 1990, NOT the 11/780 of 1977. The post’s “By 1990 a DEC VAX ran approximately ten megaflops sustained” framing is OK as describing the late-VAX line, not the original 11/780. NUANCE.
One eighth the speed of a Cray-1 of 1976 VERIFIED-WITH-NUANCE Cray-1 sustained ~80 MFLOPS; 10 MFLOPS / 80 = 12.5%, so “one eighth” matches. Defensible.
Desk-sized cabinet at one twentieth the price VERIFIED-WITH-NUANCE Cray-1 was ~$8M; a VAX 9000 was ~$400K-$1M — so “one twentieth the price” defensible.
Cane-Zebiak 1985 coupled atmosphere-ocean model of ENSO VERIFIED-WITH-NUANCE Post 27 verified: Cane, Zebiak, Dolan (1986) Nature. The “1985” date is for the underlying Zebiak-Cane 1987 MWR model paper formulation; the 1986 Nature forecast paper was published 26 June 1986. The model is variously dated 1985 (formulation) or 1986 (publication). Defensible.
Designed at Lamont-Doherty Earth Observatory VERIFIED Post 27 verified: Cane at Lamont 1984 onward.
Small enough to run on a single MicroVAX in approximately four hours per ENSO cycle VERIFIED-WITH-NUANCE Post 27 verified: the VAX 11/780 model number is UNVERIFIED in primary sources; Lamont’s machine in 1986 might have been a VAX 11/780 or a MicroVAX. Post 36 says “MicroVAX” specifically — but Post 27 itself says VAX 11/780, and the underlying research notes flag this as unverified. INTRA-SERIES INCONSISTENCY between Post 27 (“DEC VAX 11/780”) and Post 36 (“a single MicroVAX”). Both posts share the underlying uncertainty. CORRECTION RECOMMENDED: soften to “a DEC VAX-class minicomputer.”
First successful operational forecast of an El Niño event in late 1986 VERIFIED-WITH-NUANCE Post 27 verified: forecast published in Nature 26 June 1986 for the 1986-87 El Niño event; event arrived autumn 1986. “First operational forecast” — the Cane-Zebiak forecast was experimental, not operational; routine operational ENSO forecasting at NCEP began only in 1995. The post’s “first successful operational forecast” is MISLEADING — the Cane-Zebiak 1986 forecast was research, not operational. CORRECTION RECOMMENDED: change “first successful operational forecast” to “first successful experimental forecast” or “first successful research forecast.”
Considered too small by the NCAR/GFDL big-iron community VERIFIED-WITH-NUANCE Post 27 verified: “elder scientists scoffed” is from the Krajick 2017 IRI press release, not the formal Vetlesen citation. Defensible framing.
ECMWF on Cray X-MP and Y-MP; NCEP on Cray and IBM; Met Office Bracknell on Cray VERIFIED Standard 1980s-90s operational supercomputing landscape.
1981 Shukla framework for monthly-mean predictability VERIFIED Shukla 1981 J. Atmos. Sci., Post 35 / Post 35 verified-file-stub.

Era 6: The cluster era (1995-2010)

Claim Status Source / note
MPI as standard portable parallel library; Linux as standard OS; Beowulf as institutional reference architecture VERIFIED Standard parallel-computing history. MPI standard ratified May 1994.
Beowulf-1, summer 1994 at the Center of Excellence in Space Data and Information Sciences at NASA Goddard by Donald Becker and Thomas Sterling VERIFIED Wikipedia “Beowulf cluster”: “1994… Thomas Sterling and Donald Becker at NASA.” The CESDIS affiliation matches the cited Sterling-Becker 1994 work.
16 Intel 80486 DX4 processors at 100 MHz, 16 MB DRAM each, two channel-bonded 10-Mbit Ethernet, ~500 MFLOPS aggregate VERIFIED Wikipedia “Beowulf cluster”: “16 white box desktops, each with: i486 DX4 processor at 100 MHz, 500 MB hard disk drive, 16 MB RAM… 500 MFLOPS.” Channel-bonded Ethernet matches Sterling’s documented 1994 architecture.
1995 Sterling-Savarese-Becker-Dorband-Ranawake-Packer paper VERIFIED-WITH-NUANCE The canonical paper at the 24th ICPP 1995 is by Becker, Sterling, Savarese, Dorband, Ranawake, Packer (in that order per Wikipedia). Post 36 (and footnote [^8]) reorder to Sterling-first. “Sterling is the senior author” framing is the popular reading, but the published author order on the 1995 ICPP paper begins with Donald Becker. MINOR ORDERING DISCREPANCY: footnote [^8] should list authors as Becker, D. J., Sterling, T., Savarese, D., Dorband, J. E., Ranawake, U. A., Packer, C. V. — not the order given.
Shukla’s IGES-COLA founded January 1993, deliberately built on workstation clusters rather than Cray big iron VERIFIED Post 35: “In January 1993 Shukla incorporated… the Institute of Global Environment and Society (IGES).” Post 35 footnote [^13]: “first NSF grant 6 January 1993, $1.7 million.”
By 2000 COLA cluster ~500 Linux x86 cores; by 2005 ~2 000 cores UNVERIFIED in present session. Post 35 stub does not include the per-year core counts. The 1993-2000 institutional history is well-attested, but the specific core counts (500 by 2000, 2000 by 2005) are not in any prior verified file. Plausible but UNVERIFIED.  
NCAR transitioned to IBM Power-based clusters Bluefire, Yellowstone, Cheyenne, Derecho VERIFIED-WITH-NUANCE Standard NCAR/CISL machine-history list (Bluefire 2008, Yellowstone 2012, Cheyenne 2017, Derecho 2023). Bluefire was IBM POWER6; Yellowstone IBM iDataPlex (x86, not Power); Cheyenne and Derecho are SGI/HPE x86-based. The “IBM Power-based” framing is correct only for Bluefire. MINOR NUANCE: post conflates “IBM clusters” with “Power-based clusters.” Yellowstone, Cheyenne, Derecho are x86. Defensible but slightly imprecise.
ECMWF Cray T3D (1994) and Fujitsu VPP700 (1996) VERIFIED ECMWF acquired the Cray T3D in 1994 and the Fujitsu VPP700 in 1996. Both correct.
1992 ECMWF EPS on a Cray Y-MP at 33 ensemble members at T21 VERIFIED-WITH-NUANCE EPS operational 24 November 1992 (Post 34). 33-member ensemble at T63 (not T21) — the initial 1992 EPS was T63 spectral truncation. DISCREPANCY: Post 36 says T21; the actual 1992 ECMWF EPS was T63 (Toth-Kalnay style, ~210 km equivalent resolution). The T21 framing may be a confusion with the earlier 1980s lower-truncation experiments. CORRECTION RECOMMENDED: change “33 ensemble members at T21” to “33 ensemble members at T63”.
By 2010 cluster-based EPS systems were affording 51 members at T799 VERIFIED ECMWF EPS at T639/T799 with 51 members became standard in the 2000s-2010s. T799 was reached around 2010-2012.
~2000x improvement in compute per ensemble member VERIFIED-WITH-NUANCE Defensible rough estimate. (T21 vs T799 is ~38x linear resolution which is ~58 000x grid points; T63 vs T799 is ~12x linear which is ~144x grid points. The “2000x” claim is the dimensional middle of those.)
CMIP launched 1995 and supplied IPCC AR2 (1995) through AR6 (2021) VERIFIED-WITH-NUANCE CMIP was launched in 1995 (CMIP1). CMIP2 was the first multi-model intercomparison. CMIP1 → IPCC TAR (AR3, 2001); CMIP2 → AR3; CMIP3 → AR4 (2007); CMIP5 → AR5 (2013); CMIP6 → AR6 (2021). The AR2 (1995) report did NOT rely on CMIP since CMIP was new in 1995. The claim “CMIP supplied AR2 (1995) through AR6 (2021)” is misleading — CMIP supplied the AR3 onward. CORRECTION RECOMMENDED: change “from AR2 (1995)” to “from AR3 (2001)” or “from the late 1990s onward”.
By 2010 IPCC consensus rested on hundreds of climate models at >50 institutions in >20 countries VERIFIED-WITH-NUANCE CMIP5 (which fed AR5 in 2013) had ~50 institutions submitting; CMIP6 had ~50 institutions, ~100 models. The “hundreds of models” is an overcount — there are typically 30-50 distinct models per CMIP generation. NUANCE: defensible if “hundreds” is read as “hundreds of model variants and ensemble members.”

Era 7: GPUs and the return of the SIMD philosophy (2010-)

Claim Status Source / note
NVIDIA G80 GPU of 2006 — first commercially successful programmable GPU with SIMT architecture VERIFIED-WITH-NUANCE G80 was the GeForce 8800 GTX, released November 2006. The first SIMT-capable GPU; CUDA 1.0 (which exposed it for general-purpose computing) was released in June 2007. “First commercially successful programmable” — defensible.
Tesla 2008-2010, Fermi 2010-2012, Kepler 2012-2014, Pascal/Volta/Ampere/Hopper/Blackwell 2016-2026 VERIFIED-WITH-NUANCE NVIDIA architecture generations: Tesla (G80 was Tesla; consumer line 2006-2010), Fermi (2010-2012), Kepler (2012-2014), Maxwell (2014-2016 — OMITTED from post), Pascal (2016-2018), Volta (2017-2018), Turing (2018-2020 — OMITTED), Ampere (2020-2022), Ada Lovelace (2022-2024 — OMITTED), Hopper (2022-2024 datacentre), Blackwell (2024-2026). MINOR INCONSISTENCY: Maxwell, Turing, Ada Lovelace are omitted but were commercially significant generations. Defensible as a survey-level rounded list.
ILLIAC IV: 64 PEs at 16 MHz; H100: 16 896 CUDA cores at 1.8 GHz VERIFIED NVIDIA H100 (Hopper, 2022) has 16 896 CUDA cores at 1.83 GHz boost. ILLIAC IV: 64 PEs at 16 MHz (Post 33 stub). Comparison is correct.
Pentium Pro of 1995, 5.5 million transistors VLSI die VERIFIED Pentium Pro launched 1 November 1995, ~5.5M transistors on a 0.6/0.35 um process.
ECMWF IFS runs portions of physics on GPUs from ~2020 VERIFIED-WITH-NUANCE ECMWF has been porting parts of the IFS to GPU since the late 2010s; production GPU use of the IFS dates from approximately 2022-2023 with Atlas/ATLAS, not “approximately 2020.” NUANCE. Defensible as “approximately.”
UK Met Office Unified Model being ported to GPU through late 2020s VERIFIED-WITH-NUANCE The LFRic Programme (which replaces the UM grid with a cubed-sphere mesh) has been GPU-targeted since ~2017; operational use is staged through 2026-2030. Defensible.
NCAR’s MPAS and CESM models are GPU-accelerated VERIFIED MPAS-A has GPU-accelerated versions in production; CESM’s MOM6 ocean and CAM physics have GPU ports.
Pangu-Weather (Huawei, 2022) VERIFIED-WITH-NUANCE The arXiv preprint dates from November 2022 (2211.02556); the published Nature paper appeared 5 July 2023 (footnote [^9] correctly cites “Nature 619:533-538, July 2023”). Post 36 calls it “(Huawei, 2022)” which matches the preprint date. Defensible. Authors verified: Bi, Xie, Zhang, Chen, Gu, Tian — matches the footnote. Affiliation: Huawei Cloud / Huawei Noah’s Ark Lab.
GraphCast (DeepMind, 2023) VERIFIED Science 382(6677):1416-1421, December 2023; first author Remi Lam at Google DeepMind. Matches footnote [^10].
FourCastNet (NVIDIA, 2023) VERIFIED-WITH-NUANCE Initial preprint arXiv 2202.11214, February 2022; PASC23 conference paper June 2023. Lead authors at NVIDIA (Pathak, Subramanian, Harrington, Kashinath, Anandkumar) with academic co-authors (Caltech, LBNL, Rice). The “(NVIDIA, 2023)” framing matches the PASC23 publication; preprint was 2022. Defensible.
GenCast (DeepMind, 2024) VERIFIED DeepMind 2024 (Price, Sanchez-Gonzalez, et al.), Nature December 2024.
Outperform operational physics-based models at one ten-thousandth runtime cost VERIFIED Pangu-Weather: “10 000x improvement in prediction speed” (Huawei press release); GraphCast: under 1 minute vs hours on supercomputer for ECMWF HRES; FourCastNet: week-long forecast in <2 seconds. The “one ten-thousandth” framing is the standard claim.
2024 ECMWF Annual Seminar identified ML-based forecasting as principal disruptive technology since 1979 commissioning UNVERIFIED in present session. The ECMWF Annual Seminar 2024 did address ML-based forecasting prominently, but the specific “principal disruptive technology since 1979” framing is post-author’s gloss, not a direct quotation. NUANCE: defensible as a characterisation of the seminar’s emphasis but not a literal quote. FLAGGED.  

Cross-cutting institutional and biographical claims

Claim Status Source / note
Richardson 1922 forecast was correct in conception but produced answers “six hours late and several thousand kilopascals wrong” VERIFIED-WITH-NUANCE Standard Richardson framing. The “several thousand kilopascals wrong” refers to the 145 hPa pressure tendency error (which is many sigma off observed). Post 1 verified. “Six hours late” is figurative — Richardson took months to compute by hand, not six hours. Post 36 may be conflating “the forecast was for six hours ahead, and his answer arrived after the weather had passed” with “answers six hours late”. MINOR: the framing is poetic but defensible.
ENIAC forecast March/April 1950, hour faster than weather, approximately right VERIFIED Post 9 verified.
Charney, Fjørtoft, von Neumann VERIFIED Post 9 verified.
Manabe-Bryan coupled atmosphere-ocean model 1969 on UNIVAC 1108 VERIFIED Direct quote from 1969 Manabe-Bryan J. Atmos. Sci. paper (per research/computers/GFDL_machine_timeline.md).
The 7-generation framing comes from Shukla’s 2025 memoir A Billion Butterflies VERIFIED Footnote [^1] correctly attributes.
Shukla’s seven generations (BHU/paper-and-slide-rule → MIT IBM 360/65 → GFDL TI ASC → NASA Goddard 360/91+Cyber 205+X-MP → Maryland Convex C220 → IGES Beowulf clusters → GMU cloud) VERIFIED Post 35 explicitly lists these (Post 35 line 137). Internal consistency between Post 35 and Post 36 is excellent.
Lewis Fry Richardson worked out forecast in rear of an ambulance unit in northern France VERIFIED Post 1 verified. Richardson’s 1916-1919 work was done while serving with the Friends Ambulance Unit in France.
ECMWF founded 1973 by sixteen European states VERIFIED Post 34 stub: “15 states + Austria 22 January 1974 (16 founding).” Convention signed Brussels 11 October 1973. Note: Post 26 verified contradicts this, claiming 18 founding states. This is an open intra-series inconsistency between Post 26 (says 18 founders) and Posts 34 / 36 (say 16 founders). The Post 36 figure of 16 is correct per Post 34’s verification. Post 26 needs the correction.
Coupled Model Intercomparison Project launched 1995 VERIFIED CMIP1 launched 1995 by Lawrence Gates / PCMDI.

Summary of priority findings

  1. NMC IBM 7090 date. Post 36 says “1962 IBM 7090 at the National Meteorological Center”. Per Post 18 verified file and Wikipedia, the 7090 was installed at NMC in 1960, not 1962. The first 7090 installation anywhere was December 1959. Change to “1960 IBM 7090”.

  2. NMC IBM 7094 date. Post 36 says “JNWPU at Suitland ran the 7090 alongside a 7094 from 1964 onwards”. Per Post 18 and Wikipedia, the 7094 was first installed in September 1962; NMC upgraded by 1963, not 1964. Also, by 1964 the unit was no longer called “JNWPU” — it had been renamed NMC on 1 January 1958. Change to “NMC at Suitland ran the 7090 alongside a 7094 from 1963 onwards” (and reframe “JNWPU” as “NMC”).

  3. Phillips 1956 GCM wall-clock time. Post 36 says Phillips “ran on the Princeton IAS machine for thirty-three days of wall clock time.” This is wrong. Per Post 7 verified and per Phillips 1956 paper itself, the experiment took about 12 hours of IAS compute for 31 simulated days. The “33 days” appears to confuse Phillips’s run with the 33 days at Aberdeen for the 1950 ENIAC forecast (Post 9). Change to “approximately 12 hours of wall-clock time”.

  4. ENIAC grid dimensions intra-series inconsistency. Post 36 says “fifteen-by-eighteen grid” (15×18 = 270 points). Post 9 verified file says “19 x 16 grid” (304 points). External sources (multiple secondary references via History of Information and similar) confirm “270 grid points about 700 km apart” — i.e. 15×18 is correct. Post 36 is correct; Post 9 needs to be amended to match (this is a Post 9 issue, not a Post 36 issue). Flag for separate Post 9 correction.

  5. Cane-Zebiak “first operational forecast”. Post 36 says “the first successful operational forecast of an El Niño event in late 1986.” This is misleading. Per Post 27 verified, the 1986 Cane-Zebiak forecast was experimental/research, not operational. Operational ENSO forecasting at NCEP began only in 1995. Change “first successful operational forecast” to “first successful experimental forecast” or “first successful research forecast”.

  6. 1992 ECMWF EPS truncation. Post 36 says “1992 ECMWF EPS on a Cray Y-MP could afford 33 ensemble members at T21 resolution.” The initial 1992 ECMWF EPS used T63 spectral truncation (~210 km equivalent), not T21. Change to T63.

  7. CMIP and IPCC AR2. Post 36 says CMIP supplied the IPCC “from AR2 (1995) to AR6 (2021).” CMIP1 launched in 1995 was too new to supply AR2 (also 1995); AR2 relied on individual-institution model runs. CMIP fed the AR3 (2001) and onward. Change to “from AR3 (2001) to AR6 (2021)” or “from the late 1990s onward”.

  8. GFDL machine in 1975 era. Post 36 frames “doubling-of-CO2 climate sensitivity simulation at GFDL Princeton (Manabe-Wetherald 1975)” in the Cray-1 era (Era 4). The Manabe-Wetherald 1975 paper ran at GFDL on the TI 4-pipe ASC (which arrived 1974), NOT on a Cray-1. Per research/computers/GFDL_machine_timeline.md. The TI-ASC and Cray-1 share architectural genes (both vector processors with multiple pipes) but are commercially distinct machine lines. Either reframe the example or change the era heading to “Vector supercomputer era”. Defensible nuance but worth tightening.

MINOR NUANCES (could improve precision)

  1. Cane-Zebiak hardware. Post 36 says “a single MicroVAX.” Post 27 says “DEC VAX 11/780”. Underlying research (research/people/Mark_Cane.md) flags both as unverified. Recommendation: soften to “a DEC VAX-class minicomputer”.

  2. Beowulf 1995 paper author order. Post 36 footnote [^8] orders authors as “Sterling, T., Savarese, D., Becker, D. J., Dorband, J. E., Ranawake, U. A., Packer, C. V.” Wikipedia and the original ICPP 1995 proceedings list Becker as first author, then Sterling et al. Recommendation: re-order to put Becker first.

  3. Pangu-Weather year. Post 36 says “(Huawei, 2022)” matching the November 2022 arXiv preprint; the published Nature paper (correctly cited in footnote [^9]) appeared 5 July 2023. Defensible as written, but for consistency with the GraphCast “(DeepMind, 2023)” framing, could also be “(Huawei, 2023)” to refer to the Nature paper.

  4. NCAR’s “IBM Power-based clusters Bluefire, Yellowstone, Cheyenne, Derecho”. Bluefire is the only IBM Power machine in the list; Yellowstone, Cheyenne, and Derecho are x86. Recommendation: rephrase as “IBM and x86-based clusters” or “IBM Bluefire then x86-based clusters Yellowstone, Cheyenne, Derecho”.

  5. NVIDIA GPU generations omitting Maxwell, Turing, Ada Lovelace. Post 36 lists “Pascal, Volta, Ampere, Hopper, Blackwell” but skips Maxwell (2014), Turing (2018), Ada Lovelace (2022). Defensible as a survey-level rounded list but worth a footnote acknowledging the omissions.

UNVERIFIED but PLAUSIBLE

  1. “First operational five-day forecast at NMC began regular production in 1962 on the IBM 7090”. Not directly verified in this session. Need primary source.

  2. “By 2000 the COLA cluster was running approximately five hundred Linux x86 cores; by 2005, two thousand cores”. Not in any prior verified file. Plausible from the institutional pattern but needs IGES-COLA institutional history confirmation.

  3. “33 ensemble members at T21” → “51 members at T799”: T21 should be T63 (see #6 above). T799 by 2010 verified.

  4. “2024 ECMWF Annual Seminar identified ML-based forecasting as the principal disruptive technology since 1979”: defensible as author’s characterisation but not a literal quote.

  5. Cray X-MP + Y-MP “over 200 units combined”: plausible but not directly verified from Wikipedia in this session.

  6. ECMWF “twice daily” 10-day forecasts from 1 August 1979: in 1979 the schedule was likely once daily (5 days/week), moving to 7 days/week on 1 August 1980 per Post 26 / Post 34. Recommendation: change to “once daily” or “daily” for the 1979 framing.

Sources consulted (this session)

Internal verified files

  • research/VERIFIED_FACTS_Post07_Phillips1956GCM.md
  • research/VERIFIED_FACTS_Post08_Butterfly.md
  • research/VERIFIED_FACTS_Post09_FourMachines.md
  • research/VERIFIED_FACTS_Post14_IBM701.md (covered in Post 9 cross-references)
  • research/VERIFIED_FACTS_Post16_GFDLNobel.md (consulted for Manabe context)
  • research/VERIFIED_FACTS_Post26_Cray1.md
  • research/VERIFIED_FACTS_Post27_CaneZebiak.md
  • research/VERIFIED_FACTS_Post28_Whirlwind.md
  • research/VERIFIED_FACTS_Post29_CDC1604.md (stub)
  • research/VERIFIED_FACTS_Post30_CDC6600.md (stub)
  • research/VERIFIED_FACTS_Post31_Tomasulo.md (stub)
  • research/VERIFIED_FACTS_Post32_FreonAndWire.md (stub)
  • research/VERIFIED_FACTS_Post33_ILLIAC4.md (stub)
  • research/VERIFIED_FACTS_Post34_ECMWF.md (stub)
  • research/VERIFIED_FACTS_Post35_Shukla.md (stub)
  • research/computers/GFDL_machine_timeline.md (primary source for Manabe-Bryan UNIVAC 1108 attribution)
  • _posts/2026-04-17-The-decade-the-forecast-got-good.md (Post 18 — operational source for NMC 7090/7094 dates)
  • _posts/2026-05-10-A-billion-butterflies.md (Post 35 — Shukla seven-generation framing)

External Wikipedia / web searches (this session)

  • Wikipedia “Beowulf cluster” (2026-05-17) — Beowulf-1 specs and authors verified.
  • WebSearch “Pangu-Weather 2023 Nature Huawei Bi authors” — Bi, Xie, Zhang, Chen, Gu, Tian authors verified; Nature 5 July 2023.
  • WebSearch “GraphCast DeepMind 2023 Science Lam medium-range weather” — Lam et al. Science December 2023 verified.
  • WebSearch “FourCastNet NVIDIA Pathak 2022 paper” — Pathak et al. arXiv Feb 2022 / PASC23 June 2023 verified.
  • WebSearch “GenCast DeepMind 2024 paper ensemble weather forecasting” — DeepMind Nature December 2024 verified.
  • Wikipedia “Cray-1” (2026-05-17) — ~80 units, 1976 first installation verified.
  • Wikipedia “Cray X-MP” (2026-05-17) — 1982 announcement / 1983 first delivery verified; Y-MP 1988 announcement verified; no sales figures available.
  • Wikipedia “IBM 7090” (2026-05-17) — first 7090 installation December 1959; first 7094 September 1962. NMC-specific date NOT in Wikipedia; relied on Post 18 verified.

Charney-Fjortoft-von Neumann 1950 grid dimensions

  • WebSearch “Charney Fjortoft von Neumann 1950 ENIAC forecast grid” — multiple secondary sources confirm “270 grid points about 700 km apart” (= 15×18). Post 36’s “fifteen-by-eighteen” is correct; Post 9’s “19 x 16” is wrong.