Share
Embed
The short life of the Cray-3 supercomputer
After the very successful Cray-1 and the shelf warmer Cray-2, the new Cray-3 was intended to be another cornerstone in supercomputing, developed once again by Seymour Cray. There was sold only one unit, though, and rumours are, that it wasn't even payed because of stability problems right from the beginning. Cray himself already worked on a Cray-4, which wasn't finished because of the sudden death of S. Cray. The talk spans the last ca. 10 years of Seymour Cray, when he was working on the Cray-3 and Cray-4 supers. It's the last part of a five parts sequel about life and work of Seymour Cray.
Wolfgang Stief
October 12, 2019
Transcript
35 slides
-
1 Wolfgang Stief VCFB 2019 – Berlin – October 2019
Cray-3 ff. English Version -
# whoami • Electrician, Dipl.-Ing. (FH) • freelancing, sys4 AG
Storage, E-Mail, project management, board member • Computermuseum München, Cray-Cyber.org “Everything below 30 A is leakage current.” • https://twitter.com/stiefkind/ [email protected] E-Mail: [email protected] 2 Pic: Wolfgang Stief, CC0 -
00_README.txt • work in progress last part • Cray and
Surroundings Companies People Technology Forks • 5-part sequel (VCFE/VCFB) • http://www.speakerdeck.com/stiefkind/ https://media.ccc.de/search?q=Wolfgang+Stief 3 -
What happened so far (Part 1, 1949-1959) • CSAW (US
Navy, WWII) ➛ ERA (1946) ➛ Remington-Rand (1951) ➛ Sperry-Rand (1951) ➛ Control Data (1957) • 1951: Seymour Cray (∗1925 †1996) starts at ERA B. Sc. Electrical Engineering (1949), M. Sc. Applied Mathematics (1951) • Control Data Corporation (Fall 1957) Plan: U$ 600.000 seed money, shares privately hold seed capital reached: U$ 1.2 Mio. • CDC Little Character (prototyping for transistor computers) • CDC 1604 (1959) 48bit, 0.2 MHz ➛ worlds fastest computer 4 -
What happened so far (Part 2, 1959-1972) • CDC6600: 100
ns ≙ 10 MHz, 1-3 MFLOPS (1959-1965) > Si transistors, liquid cooled (Freon) • CDC7600: 25 ns ≙ 40 MHz (1965-1969) > Pipelining; > U$ 8 Mio. for only ca. 4x Performance ➛ too expensive for most cust. • Control Data Corporation: Expansion weltweit, Zukäufe > stock price U$ 1 (1957) ➛ U$ 300 (1964) > lawsuit CDC ./. IBM because of marketing practice, monopoly, antitrust > main business oriented towards service instead of hardware • CDC8600: 8 ns ≙ 125 MHz, quad CPU (discreet) > reliability (!) ➛ way too much solder joints, heat evolution 5 -
What happened so far (part 3, 1972-1976) • foundation Cray
Research Inc. > Q1/1972, development CDC8600 is stuck, CDC becomes investor • design Cray-1: single CPU, 12.5ns clock cycle (=80Mhz) > vector CPU • mid of march 1976 IPO > 600.000 Shares, ca. 10 Mio U$, payment of outstanding debts, finish Cray-1 development • plan Cray-1: build/sell 2 Systems per year > sold ca. 80 systems during 6 years, price per unit U$ 5 … 8 Mio. • John Rollwagen becomes president > max. 10 years, stays until 1993 (Department of Commerce @ Clinton) 6 -
What happened so far (part 4, 1976-1986) • design Cray-2:
4-CPU, 4ns clock cycle (≙250 MHz) > development starts 1976 > first idea: 64-CPU MPP incl. Compiler ➛ gets warped after 2 years • challenge Cray-2: packing density, heat > VLSI development in concurrent development team (1981/1982) > immersed cooling • @ Cray Research Inc.: growth, pressure to succeed • parallel to Cray-2: X-MP, binary compatibility to Cray-1 > arises 100% without Seymour Cray • market launch Cray-2 in 1985 (9! years in development) > 27 sold systems, U$ 12 … 17 Mio./unit • at ∼ same time: 4-CPU X-MP > 126 sold systems, ca. U$ 16 Mio./unit 7 -
# ls -l • Cray Research Inc. 1987 ff. •
Cray Computer Corp. from 1989 • Cray-3, Cray-4 • SRC Computers LLC Cray-5/6 • Forks 9 -
• ca. 1985/86 first ideas and development > Seymour Cray
“jumps” on after development of Cray-2 > Cray-2 needs to become “production ready” • plan for Cray-3: 2 ns clock cycle (≙ 500 MHz) • max. 16 CPU • GaAs semiconductor for CPU > faster switching > already used in high frequency technology > DTL (see also CDC 1604) > less power dissipation then silicon > memory stays with silicon Cray-3 Development 10 -
“My primary mission today is to tell you everything I
know about building a computer out of Gallium Arsenide. Now, that won’t take a lot of time, because I don’t know that much.” 11 Seymour Cray, “What’s all this about Gallium Arsenide?” (Lecture, November 1988) -
• Cray MP, Steve Chen > “next generation” Cray X-MP
und Y-MP • Cray-3, Seymour Cray > “new development” from Cray-1 and Cray-2 • Steve Chen gets promotion to Senior Vice President > maximum for engineers @ Cray Research ➛ internal competition Situation @ Cray Research, ca. 1987 12 -
• Steve Chen gets responsible for “MP Project” > subordinated
to CEO John Rollwagen > CEO doesn’t know much about technology ➛ becomes problem later on • Les Davis finishes production of Y-MP Situation @ Cray Research, ca. 1987 13 • 3 projects, alle cost money, none makes money > Cray Y-MP (Les Davis) > Cray MP (Steve Chen) > Cray-3 (Seymour Cray) -
• Cray X-MP “is done” (customers ask for replacement) •
Cray Y-MP needs 1 more year for volume production • Cray-3 in development for 3 years already > lots of problems on different parts > further time schedule very unclear • Japanese pushing into US market > Fujitsu, NEC, Hitachi • Control Data plans reentry in supercomputing market > ETA-10 Situation @ Cray Research, ca. 1987 14 -
• Cray-3-Team should be quicken up • entire design team
incl. Seymour Cray > assembly there for 5 months already > design so far was exclusively in Chippewa Falls > decision was surprising and totally unexpected • Cray Research buys a former Inmos building Relocation nach Colorado Springs – Aug. 1988 15 • transportation chaos, equipment gets damaged • new hires @ Colorado Springs > only industrial engineering, no systems design • from Chippewa Falls: only ca. 10-15 systems designers > too less for a quick start > none of the “old ones” is willing explain “her baby” to the new ones -
• volume production of chip modules • CPU tests
> 2ns cycle time > testing equipment not freely available, needs to be customized ($$$) • no Les Davis available > X-MP + Y-MP @ Chippewa Problems, Problems, Problems 16 end of 1988 first layoffs @ Colorado Springs -
• foundation Cray Computer Corporation, ca. mid 1989 > already
ca. 4 years development Cray-3 • seed money from Cray Research Inc. > U$ 50 Mio. in assets (buildings, factories, etc.) > U$ 100 Mio. in debt obligations • chaos stays > barely hierarchy, rarely formal meetings > many engineers from HP, DEC (used to hierarchy) > S. Cray, now 65 years old, still acts as sole chief designer > criticism on S. Cray could lead to layoff The Cray Way – chaos stays 17 Cray Computer Corp. becomes competitor of Cray Research Inc. -
• only a few semiconductor manufacturers > all big manufacturers
$$$ invest in silicon fabrication > buys manufacturing lines from Gigabit Logic ./. shares • from August 1990 Cray Computer Corporation manufactures first own GaAs wafers > ca. 15% yield • production Cray-3 can finally start > until now 5 years of develpoment Problems, Problems, Problems – GaAs 18 -
• first shipment Cray-3 in 1993 > 8-9 years development
• price/performance gets more important in mean time than raw performance • no backward compatibility • no software applications • no system software > (UNICOS) • cold war ends > smaller budgets > customers are no more willing to write systems software on their own It’s done, when it’s done 19 Cray Computer Corp. won’t sell a single Cray-3 system. -
• Spring 1994: design of Cray-4 starts • new chips,
higher gate density > only 5 modules/CPU, instead of 40 @ Cray-3 • development runs very smooth from the beginning • very tight financial framework The Cray Way – once again 21 -
• tests of foreground processor, backend processor and memory only
ca. 3-4 weeks each > Cray-3: 18 months each • fall 1994: staging diagnostic tests for 1-CPU Cray-4 • tests show: 1 ns Cycle Time (≙1GHz) > once again fastest computer on earth (Pentium III 1GHz ca. 2000) Cray-4 on track 22 -
• plan: beginning of 1995 first systems with 4-/8-CPUs
onto market > ca. 2x performance of Cray-3 at only 20% cost • Cray Computer Corporation plans further on > Cray-5 @ 0,5 ns cycle time (≙ 2 GHz) GaAs too slow? Maybe new material once again? > Cray-6 @ 0,25 ns cycle time (≙ 4 GHz) • Cray-5: entire design by Seymour Cray • Cray-6: logic equations by Seymour Cray > then retirement Cray-4 ff. 23 -
• development Cray Y-MP > Computer Aided Design (logic equations,
wiring/routing) > X-MP: 1 CPU @ 250 Boards, each ca. DIN A4 size > Y-MP: 8 CPUs @ same volume as X-MP > completion “takes 6 more months” • Steve Chen “gets” Cray MP Projekt > should be finished in 1989, development costs planned for U$ 50 Mio., gets a quick raise to U$ 75 Mio. > 64 CPUs, multichip modules, multiple pipelines > optical clock, optical switching > 1 ns clock cycle (≙ 1 GHz) > team grows in only few months to >100 people > (!!) budget cut U$ 1 Mio. + 40 people less Meanwhile @ Cray Research (1987++) 24 -
• foundation Supercomputer Systems Inc. (Januar 1988) > Steve Chen,
Investor IBM, Cray Research denies invest > Eau Claire, Wisconsin (state offers U$ 42 Mio. for settlement) > Steve Chen + 43 engineers of Cray Research > Invest/seed money U$ 150 Mio. > money runs out before development of SS-1 is finished • development Cray C90 starts > 40 “saved” people of MP project > chief developer Steve Nelson > C90 will become the last ECL based machine > Computermuseum München: 4-CPU, 250 kW, 2/3 of the (big) footprint is power supply Meanwhile @ Cray Research (1987++) 25 -
Genealogy 26 CDC 6600 CDC 7600 (CDC 8600) Cray-1 Cray-2
(Cray-4) [Cray-5] [Cray-6] (Cray-3) X-MP Y-MP (MP) C90 T90 T3D T3E CS6400 Sun E10000 -
• March 24, 1995 further funding didn’t work out •
March 25, 1995 Chapter 11 • management tries further funding until July 1995 Cray-4 – But! 27 ⚡ Cray Computer Corporation won’t sell a single computer in 7 years of company history. The company is not at all relevant in the supercomputing industry. -
• foundation SRC Computers > Terry Willkom, head of production
@ CCC 1990-1992 > SRC – Seymour Robert Cray • “a typical Cray computer with Intel inside” > 512 CPUs > 1 TFLOP (12.000x Cray-1) Yet another “The Cray Way” 28 -
• traffic accident on September, 22 1996 > rescue service
need 90min, to cut Seymour Cray out of his Jeep Cherokee > broken neck, heavy whiplash injury 29 On October 5, 1996 Seymour Cray dies. -
Start fresh with each new project. Design simply using proven
technologies. Work in small groups with a single decision maker. Cray Principles – The Cray Way 30 -
Beyond this talk — reading • Charles J. Murray —
The Supermen The Story of Seymour Cray and the Technical Wizards behind the Supercomputer • Kay A. Robbins, Steven Robbins — The Cray X-MP/Model 24 A Case Study in Pipelined Architecture and Vector Processing 31 -
Beyond this talk — reading • Bitsavers PDF Archive (Mirror
Uni Stuttgart) http://bitsavers.informatik.uni-stuttgart.de/pdf/cray/CRAY-3/ 32 -
• Oral History of Les Davis (May 2010) http://archive.computerhistory.org/resources/access/text/2012/04/102657915-05-01-acc.pdf •
Gordon Bell — A Seymour Cray Perspective (presentation) http://research.microsoft.com/en-us/um/people/gbell/craytalk/ • Cray-3 Supercomputer Systems brochure 33 Beyond this talk — reading -
Beyond this talk — watching • Cray Research — A
Story of the Supercomputer (documentation) https://www.youtube.com/watch?v=wn03wn3k47Y • What's All This About Gallium Arsenide? Lecture by Seymour Cray (November 1988) https://www.youtube.com/watch?v=8Z9VStbhplQ 34