[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[computer-go] Moore and computer games
I am no expert in Moore's law, but we all agree that it is of overwhelming
influence in computer go when looking forward a year or 30.
A friend of mine, doctor Eeuwe Sieds Zijlstra, said that electrons run
roughly at 1/3 of lightspeed and that this would be equal to 1.3Ghz, which
for that type of material might pose a problem.
They then started to use Cupper.
Processors get massively parallel now too.
Let's express Moore's law in gflops.
Fastest supercomputer now is 50 tflops.
In 2005 the fastest supercomputer might have 360 Tflop already.
Chrilly i heard murmuring already years ago about a dual core hydra.
Chrilly was ahead of his time. Next year there will be dual core processors.
They will find new tricks, don't worry.
The only real problem there is for computer go will be the hard fact that
it is real difficult to parallellize efficiently a program, as the
bandwidth might get faster, but the interconnects between the cpu's aren't
improving much in ping pong speeds, which is the program in MPI terms that
gets used to figure out the speed it takes to ship a ping and get back a pong.
Of course the clever hardware guys have found a free factor 2 by only
publishing a 'one way ping pong latency', which is simply a pingpong time
divided by 2.
With the current search depths a global hashtable might not work that
efficiently yet for go software. So this is a good time to enter a Go
Zugzwang.
Perhaps Chrilly wants to explain why.
Even at a 512 processor 1Mhz transputer system with a 10 mbit network they
just search at 200 nodes a second in total.
Others search at way cheaper 8 bits single chips 1Mhz H8 (which is in the
cheap chesscomputers in the store) about 1500 nodes a second (under 1000
cycles a node).
I'm not saying a factor 1000 is missing somewhere, but let's start with a
factor 100.
Yet it is this latency problem that will get the only real problem in the
future. Because you already saw the problem.
1Mhz transputers, 10 mbit network. Latency probably was real bad at that 10
mbit network. I am guessing it was like 200 us.
Now there is 2.4Ghz opterons, fritz no longer is at under 1000 cycles a
node, as he has worked for 6 years at its chessknowledge now.
It was 5 million nodes at a quad opteron. That was not optimized however
for opteron and it was a 32 bits executable.
Crafty is optimized for it and searched 7.74 million nodes a second at a
quad opteron (2.4Ghz) in world champs 2004.
The fastest affordable network cards are myrinet cards (there is faster
solutions but well...)
Their one way ping pong time is 5 us.
That's factor 40 faster network roughly.
This where the cpu's have become more than a 3000 times faster.
In reality way more than that.
I do know that for some encryption software i wrote, that the RSA goes 3
times faster at an opteron than it goes at a k7. The k7 and opteron run at
similar speeds. How much diep gets faster, well you can see that for
yourself online at www.aceshardware.com, where diep gets tested at each new
processor.
Let's do concreted math.
End of 1996 i bought a 200Mhz pentiumpro processor. That was real real fast.
Regrettably it wasn't in time for the dutch open international
computerchamps so i had to run on my old machine. A P5-100Mhz. Diep was 3.0
times faster at the pentiumpro.
The PII-450 a few years later was exactly 2.0 times faster than the PII.
I went from DOS to Windows
Better compilers then speeded it up 40%.
>From PII-450 i went to dual P3-800 in the year 2000.
The P3 at the same clockspeed was 20% faster than PII.
The machine was a dual which added a factor 2. And there is a clockspeed
difference of nearly a factor 2.
In 2001 i went from my dual P3-800 to a dual K7-1.2Ghz
Now *that* was a speedjump, despite the small clock difference.
Later i upgraded the cpu's to 1.6ghz and bought a second dual k7 as the
dual opteron was too expensive for me.
A k7 MP is 20% faster than a P3.
World champs 2004 i played at a quad opteron 2.0Ghz
So from 2 to 4 processors and a single cpu opteron is more than 43% faster
than an equal clocked K7 for diep.
Next year there is dual core opterons and end of 2006 there is dual core
pentium m's too. I hope intel can speed that up to resist AMD a little.
AMD is taking over all that market share intel has now otherwise.
Compiler improvements all together is another 20% sincethen.
I expect there a big boost for x86-64 somewhere next year when microsoft
releases a 64 bits compiler allowing pgo. At least 10% for the 8==>16
registers and i hope for another 5-7% from PGO.
Diep is a 32 bits program, not 64 bits unless you want to count my Zobrist
hashing and a few beancounters.
So concluding the total from this users viewpoint from 1996 until today:
1996 p5->p6 : 3.0 *
1998 p6->PII : 2.0 *
msvc compiler : 1.4 *
2000 PII=>P3 : 1.2 * 800/450 * 2 *
2001 p3=>MP : 1.2/0.8 * 1.2 *
2002 ==>2.1Ghz : 2.127 / 1.2 *
compilers : 1.2
2004 quad 2.0 : 2 * 1.43 * 2.0/2.127 =
= 3.0 * 2.0 * 1.4 * 1.2 * 16/9 * 2 * 1.5 * 1.2 * 2.0 * 1.2 * 2 * 1.43 * 1.2 =
= 10 * 16/9 * 17.28 * 1.43 * 1.2 = 527 times faster from running viewpoint.
So from Moore's viewpoint that's log(527) / log 2 = 9.0
Wow what a great number to calculate.
>From DIEP's viewpoint Moore happens each year.
Now some will argue that a factor 4 is because of multiprocessing.
However i must inform you that the US government is asking for a 1 petaflop
machine already. If i remember well next year a 65536 processor machine
will be there from IBM.
Right now a 12288 processor supercomputer gets constructed in Netherlands.
And so on.
Those machines have more and more processors too and they are all of them
soon dual core too. IBM power5 already is.
At 18:56 7-11-2004 -0500, Don Dailey wrote:
>
>> Also in go, in order to beat the pro's in a year or 30, you MUST use brute
>> force searching algorithms.
>>
>> If you don't use them, you *always* will keep playing at utmost beginners
>> level.
>
>There are some excellent arguments that Moores law won't hold forever
>due to limitations imposed by the laws of physics. There are other
>speculations that say science either finds ways around these limits or
>takes advantage of other physical laws (such as quantum computing.)
>
>It doesn't seem to me that brute force alpha/beta global searching HAS
>to be the only way, but it is a definite possibility if Moores law
>finds a way to keep working for a few more decades.
>
>- Don
>
>
>_______________________________________________
>computer-go mailing list
>computer-go@xxxxxxxxxxxxxxxxx
>http://www.computer-go.org/mailman/listinfo/computer-go/
>
>
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/