How fast was the Cray?

by Frederic Friedel
9/23/2016 – The Cray-1 was the most successful supercomputer in history. It was built in 1975 and cost eight million dollars. Soon a chess program was ported to it, and "Cray Blitz" proceeded to win the Computer Chess World Championship twice in a row. So how fast exactly was the Cray, and how much progress we have made in four decades. Are today’s consumer computers faster, and if yes by what factor? Ten? You are not going to believe this.

ChessBase 14 Download ChessBase 14 Download

Everyone uses ChessBase, from the World Champion to the amateur next door. Start your personal success story with ChessBase 14 and enjoy your chess even more!


Along with the ChessBase 14 program you can access the Live Database of 8 million games, and receive three months of free ChesssBase Account Premium membership and all of our online apps! Have a look today!

More...

Cray-1 – the eight million dollar super-computer

By Frederic Friedel

As a young TV science journalist at the beginning of the 1980s I travelled to the Max Planck Institute for Astrophysics in Garching, near Munich, Germany. They had the fastest computer in the world, a Cray-1, and I was there to do a story about it. I remember entering Max Planck with my film crew, sitting down and telling some scientists why we had come. Then I asked “where is the computer, the Cray?” – “You are sitting on it,” they replied. Oops, right they were. This is what the Cray-1 looked like:

The Max Planck Cray-1, with storage units in the background. Each could hold three gigabyte!

Of course when you took off the panels of the tower and the base things looked different

This super-computer had been developed by CDC engineer Seymour Cray, who had found backers on Wall Street for the project. It took four years to build, and in 1975 when the first 80 MHz Cray-1 was announced, interest was so high that a bidding war broke out between the Lawrence Livermore and the Los Alamos National Laboratories. The latter eventually won and took delivery of a trial machine with the serial number 001. The first regular customer was the National Center for Atmospheric Research (NCAR). They payed $8 million for a Cray-1 (serial number 3) and a million for storage disks. Over the years Cray Research sold over eighty Cray-1s, making it one of the most successful supercomputers in history. You will find a lot more information on this at Wiki.

Anyway, I was absolutely thrilled to see the Max Planck super-machine, and my six-minute piece on it for German TV was positively effusive. A couple of years later I got to see another Cray, this one in the basement of the Bell Laboratories, where my friend Ken Thompson had access to it. It was faster and more advanced than the one in Germany, and we dutifully drooled over it.

Robert Hyatt and Albert Gower (on the right) who have defeated Ken Thompson
and Joe Condon (left) who held the title with their dedicated Belle machine

Around 1980 Cray Research chose to sponsor the program "Blitz", written by Robert Hyatt, Harry L. Nelson and Albert Gower. It was ported to the Cray and participated in computer chess events until the mid 1990s. "Cray Blitz" won several ACM computer chess events, and two consecutive World Computer Chess Championships, the first in 1983 in New York City, and the second in 1986 in Cologne, Germany.

Cray Blitz and the wrong bishop

Recently I wrote an article on an endgame involving a “wrong bishop”. For those of you who missed it here is a summary. The endgame king, bishop and pawn vs king is always a draw when

  1. the pawn is on the a or h-file and the bishop does not control the promotion square;
  2. the defending king can reach the promotion square in time to defend it.

For a long time this simple piece of chess knowledge eluded chess playing computers. They simply couldn't calculate all the way to the promotion, and without specific instructions on how to handle the position they would often take a wrong decision. During the 1980s, when the first programs were being told about the wrong bishop draw, I devised a test position to find out if they could handle the endgame correctly.

[Event "Theoretical draw"] [Site "?"] [Date "2016.05.13"] [Round "?"] [White "The wrong bishop - test"] [Black "White to play"] [Result "1/2-1/2"] [SetUp "1"] [FEN "k7/7p/8/2nKb3/8/8/8/8 w - - 0 1"] [PlyCount "11"] [SourceDate "2016.05.13"] {As an astute chess player you will immediately know what to do:} 1. Kxc5 $1 ({ But for the computers at the time it was too difficult. They would inevitably take the bishop and lose the game, e.g.} 1. Kxe5 $4 h5 2. Kf4 Ne6+ 3. Kg3 Kb7 4. Kh4 Ng7 {and Black wins.}) 1... h5 2. Kd5 h4 3. Ke4 h3 4. Kf3 h2 5. Kg2 Kb7 6. Kh1 {theoretical draw.} 1/2-1/2

The temptation for early day computers was simply too great. They went astray for the following reasons:

  1. a bishop is valued higher than a knight, especially in open endgame positions;
  2. Kxe5 centralises the king, while Kxc5 doesn't
  3. if it takes the knight the black pawn advances all the way to h2, where it gets a lot of bonus points for being one square away from promotion. After taking the bishop it can be kept on the top half of the board.

So any sensible computer would take the bishop, unless of course it was told to do otherwise. Which seemed to be the case in the following game.

[Event "Mississippi State Championship"] [Site "?"] [Date "1982.??.??"] [Round "?"] [White "Human"] [Black "Cray Blitz"] [Result "0-1"] [WhiteElo "2265"] [SetUp "1"] [FEN "R7/p4r2/1p2b3/4k3/P7/4KB2/8/8 b - - 0 1"] [PlyCount "17"] [SourceDate "2016.05.13"] {Naturally the computer immediately saw the trick and played} 1... Rxf3+ $1 2. Kxf3 Bd5+ 3. Ke3 Bxa8 {With a piece and a pawn down the human player was reduced to trying a swindle:} 4. a5 $6 {Could it work, would the program take the free pawn offer? Nope, Cray Blitz avoided the pitfall and played} Be4 $1 { winning the game comfortably after} (4... bxa5 $4 5. Kd3 Bd5 6. Kc3 Ba2 7. Kb2 Be6 8. Ka1 {with the theoretical wrong bishop draw.}) 5. Kd2 Kd4 6. Kc1 b5 7. Kb2 Kc4 8. a6 b4 9. Ka2 Kc3 {with a mate announcement in six.} 0-1

At the time this was celebrated as the first instance of practical chess knowledge, implemented into a computer (the “intelligent method”), leading to a victory when otherwise a draw would be expected. But some time later I obtained the computer logs and discussed the position with Harry Nelson of the Cray team. It turned out that Cray Blitz had simply searched deep enough to see the promotion after 4...Be4, – by pure brute force –, while it found no queen on the board after 4...bxa5. It was that simple.

In part two of my article I told you how I had shown a subtle wrong bishop position to strong grandmasters, including former World Champions Mikhail Botvinnik and Mikhail Tal – and to the Computer Chess World Champion Cray Blitz. It spent 13 seconds considering a wrong key move – with a logical +4.032 score. Then, at ten ply, it switched to the correct move and displayed +10.878. Mikhail Botvinnik saw this all happen and he was not at all happy that a computer was able solve the position so effectively by pure brute force, without the chess knowledge he, a proponent of the "intelligent method", so ardently advocated.

How fast was the Cray-1?

While I was writing these historical articles I started to wonder: how fast, exactly, was the Cray-1 , and how did it compare to today’s machines. It was expedient that our friend John Nunn had just bought his son Michael a good mid-range graphics card for his 18th birthday. John did some calculations for me and wrote back: “The Cray-1 could do 130 Megaflops" [million floating point operations per second]. "The NVidia Graphics card in Michael’s computer can do 2258 Gigaflops. So it is about 17,000 times as powerful by this measure.” John conceded that, of course, the architectures are very different.

Seventeen thousand times more powerful? A card you can hold in the palm of your hand, and which costs around $300?

I consulted Ken Thompson on this – he has been work on the forefront of computing for over fifty years. Ken, who never capitalizes anything, wrote:

the cray had 2–5ns cycle time. (depending on model) in that time, it could get up to 7 arithmetic units executing an instruction. the vector length was 64 and it took a few instructions to start and a few to shut down. some models had up to 8 processors. so, peak rate is about 0.5G (2ns) * 7 * 8 or about 25G. now to get to reality – not everything can be vectored. and when you can vector, usually only a few of the units are used. in fact, most instruction are simply run at the clock rate. some instructions take multiple units (divide, square root). but the cray had huge bandwidth to memory. the i/o was staggering.

You figure this out. Do you believe that the factor 17,000x is possible, or are we making a logical mistake? Experts, please tell us in our feedback section below.

Source: Frederic Friedel at Medium.com


Topics History

Editor-in-Chief of the ChessBase News Page. Studied Philosophy and Linguistics at the University of Hamburg and Oxford, graduating with a thesis on speech act theory and moral language. He started a university career but switched to science journalism, producing documentaries for German TV. In 1986 he co-founded ChessBase.
Discussion and Feedback Join the public discussion or submit your feedback to the editors


Discuss

Rules for reader comments

 
 

Not registered yet? Register

BenRedic BenRedic 9/29/2016 10:50
Can't believe nobody mentioned that Robert Hyatt continued his work on chess programming in the form of the open source winboard engine Crafty, which has been around for as long as I can remember and is still being developed. The latest and greatest (Crafty 25.0 64-bit 4CPU) currently holds a 2959 rating on CCRL 40/40.

Also interesting is the fact that the fortran source code for Cray Blitz is available for download. I have also come over a version that is supposedly compiled by Jim Ablett for PCs (http://kirr.homeunix.org/chess/engines/Jim%20Ablett/CRAY%20BLITZ/), but unfortunately it does not seem to work properly when I test it. The program starts just fine, but when it is the computers turn to move I get an access violation error. Has anybody had any luck getting this program up and running? Could be interesting to see what Cray Blitz actually played like :-)
zookid zookid 9/28/2016 07:50
...not capitalizing, apparently comes from a "flaw" in the dna. scientists have not even begun to break it down exactly how or why this happens. many people (often passive-aggressive) have this condition, although not many are as well known as e.e. cummings. i too carry this erratic gene, as well as my father, first cousin, and one of my sons. therapy and medication have proven fruitless to this point, but there is still hope and time. thank you.
starso starso 9/28/2016 06:18
" Ken, who never capitalizes anything, wrote: "

Anybody knows if there´s any special reason for not capitalizing? Or just for the sake of breaking ortography reules? Just curious.
Force of Nature Force of Nature 9/25/2016 05:49
So when looking for a commercially affordable Laptop to use primarily with chess engines, which as you say don't utilize GPU's very efficiently, then please explain to the layman what are the most important specs to go for?
mistery mistery 9/24/2016 08:16
Again apples to oranges, but I would be interested to know how a CRAY compares to an iPhone.
dhochee dhochee 9/24/2016 06:09
Moore's law, which has held reasonably true since 1975, is that transistor density, which is closely tied to overall processing power, would double every two years (or 18 months, by another estimate). It was a bit faster than that early on and has tapered off in recent years, but it's still useful for getting a ballpark figure for the increase in computing power over the years.

Between 1980 and 2016, that would be an increase of 2^18, or ~262,000. Of course, as others have already commented, comparing a supercomputer to a graphics card is apples to oranges due to completely different architectures, but regardless, the increase in calculation power isn't really shocking to those who have followed computing through the years.
Samuel Siltanen Samuel Siltanen 9/24/2016 09:01
The problem with the comparison is that there is no chess engine running on a GPU today. The algorithms used in chess engines cannot easily utilize the massive parallel computing power in them. The alpha-beta search and its variants are inherently serial. In addition, GPUs have other limitations that prevent efficient implementation of chess algorithms on them: very limited local fast memory (global memory accesses take hundreds of clock cycles), slow branching instructions, and no easy communication between processors. They are not yet fully general purpose processors.

Of course, the supercomputer Cray also had parallel processors, but there were orders of magnitude less of them. There are algorithms that scale to a few processors, but not to thousands of them. The scaling will not be linear, so much of the computing power will be wasted.
jsaldea12 jsaldea12 9/24/2016 03:22
I believe you will agree that super-computers nowadays, like Komodo, Fritz, Stockfish, etc. are much advanced than Cray. Super-computers, the more programmed with as much all openings and positions, middle and end games, will outstrip chess grandmasters. But one biggest advantage of computers is that it moves with the speed of light, and encountering stated programmed positions, billions, is simple to machines. Though gifted, Grandmasters are no match in comparison but given the proper time to analyze, GMs can reach the level of programming of computers, plus creativeness of new moves not progremmed. Thus,k to be fair, have a match between computer, even without odds, on condition like this: allot to the super-computer,say, 10 seconds each move, and 2 hours for a grandmaster or master to equalize level of analysis. thus, of equal strength.. After all, computers are the making of human mind. What do you think?
Queeg Queeg 9/24/2016 12:36
Comparing a supercomputer with a graphics card is like comparing a lorry with a racing go-kart. The latter is faster but the former is much more powerful.

The comparison is misleading for the following reasons:
metric: FLOPS are irrelevant because chess engines utilise integer arithmetic almost exclusively.
SIMD: GPUs are optimised for applying instructions to a huge set of data. This pattern is virtually absent from chess engines.
latency: graphics cards come with too few fast local memory/cache memory. Accessing their huge global memory is penalised with high latency times. Chess engines would be stalled by this slow RAM access.
branching: GPUs have long instruction pipelines but lack branch prediction. If a program takes a different path than expected, a lot of processing time will be wasted. Chess engines would suffer big time.

Chess engines are a fine example for a computing problem which is completely unfit for graphics cards.
gmwdim gmwdim 9/23/2016 09:40
17000 times faster in doing floating point operations is not surprising in the least, if you consider Moore's Law which is the empirical observation that chip performance doubles every 2 years. Over 40 years that would be 20 doublings for a factor of 2^20 or about a factor of a million. Of course, floating point operations per second does not translate directly to speed in searching chess positions, but it's no coincidence that a chess app running on a smartphone such as Play Magnus is significantly stronger than Cray Blitz, due to a combination of faster hardware and smarter software.
Catastrophe Catastrophe 9/23/2016 07:48
Cray Blitz searched about 3000 nodes/sec (3 kn/sec). Source: http://www.aaai.org/Papers/Workshops/1997/WS-97-04/WS97-04-013

High end hardware on modern engines is reaching 25000 kn/s, and top of the line close to 50000 kn/s. These are very expensive machines, but nowhere near the ~$8 million cost of the Cray-1. The only benchmark we have on a modern machine that comes remotely close to costing this much, is Jonny running on ~2000+ cores at the recent WCCC. Supposedly this put it at or above 1 Gn/sec (1000000 kn/s), but the scaling on that many cores was probably dreadful.
FGMP FGMP 9/23/2016 07:45
TLDR...just kidding
rgorn rgorn 9/23/2016 07:43
Here are some historic Linpack numbers:

http://www.netlib.org/utk/people/JackDongarra/faq-linpack.html#_Toc27885750

It gives the cycle time as 12.5 nsec and 3.4 MFlop/s for the 1979 model and 12 MFlop/s for the 1983 model. Both for one CPU.
Aighearach Aighearach 9/23/2016 07:07
You're making a mistake, because outside of benchmark fake problems it is that IO bandwidth that will be your bottleneck, usually just the bandwidth between the RAM and the CPU will limit the speed.

A modern computer can do thousands of things at the same time that the Cray would have had to do one at a time, and so for the type of things people use a personal computer for, it is faster overall. But the sort of hard problems that a supercomputer is useful for? The Cray would still spank a desktop computer! Modern computers have faster-sounding IO than the Cray had, but not very much bandwidth.

Even an early 90s desktop workstation often has more memory bandwidth available to the CPU to do actual calculations than a modern CPU, and for many tasks makes a faster server. But again it is slow at task switching and has low total memory so it can't run a modern OS and do a modern workload. But if you run the old software on the new computer, often the old one is faster at the tasks it was put to. Of course, that's comparing a computer that was $50k+ then to one that is <$1k now.
Werewolf Werewolf 9/23/2016 04:15
Of course for chess the subject of floating points (GFLOPS) is not relevant. So while it may be true the graphics card is much faster at processing mathematical / graphical instructions, we don't care about that for chess purposes. Integer performance matters more, and, of course, there are very good reasons why chess programs don't run on graphics cards....
I'm rather surprised the article doesn't mention Cray Blitz's nps. That would at least be a starting point for working out its speed.
enava enava 9/23/2016 01:23
Sorry, in 1946
enava enava 9/23/2016 01:22
That's not a fair comparison, if you compare processors with 36 years between. Just for comparison, ENIAC computer was 'only' capable to make a few thousands of additions or a few hundreds of multiplications in a second and it was considered a huge achievement in 146
Sito Dekker Sito Dekker 9/23/2016 12:03
It's simple: 2258G / (0.5G * 7 * 8) = 90
So the NVidia card is "just" 90 times faster than the Cray-1
Giddy Giddy 9/23/2016 11:49
@TMMM: I agree that a factor 17000 would not be surprising. The "You are not going to believe this" in the front page text is simple clickbait.
TMMM TMMM 6/26/2016 06:37
Why would a factor 17000 be unrealistic? Surely it is possible and I would not even call it surprising. As a simple example, consider video games, which were pixelly and laggy decades ago, and now are astonishingly detailed running at 60FPS on modern machines with ease.
1