Author: Vincent Diepeveen
Date: 09:41:59 10/17/02
Go up one level in this thread
On October 16, 2002 at 11:03:33, emerson tan wrote:
Nodes a second is not important. I hope you realize that
if you create a special program to go as fast as possible,
that getting around 40 million nodes a second is easily
possible at a dual K7.
Do not ask how it plays though or how efficient it searches.
Important factors are
- he needs a new very good book. He will not even get
10th at the world championship when his book is from 1997,
and i do not know a single GM in the world who could do the
job for him. You need very special guys in this world to do
a book job. They are unique people, usually with many talents.
Just hiring a GM is not going to be a success in advance.
If you look what time it took for Alterman to contribute something
to the junior team, then you will start crying directly.
- the evaluation needs to get improved bigtime
- To get a billion nodes a second chip he needs around 100 million
dollar. Of course more cpu's doing around 40 MLN nodes a second
at say 500Mhz, he could do with just 10 million dollar.
But if you can afford 10 million dollar for 40MLN nps chips,
you can afford a big parallel machine too. Note that for a single
cpu chip doing about 4 million nodes a second, all he needs is
a cheap 3000 dollar FPGA thing. If you calculate well, then
you will see that deep blue got not so many nodes a second in
chip. it had 480 chips, and deep blue searched around 126 million
nodes a second on average against kasparov. So that's 265k nodes
a second at each chip.
So a single chip getting 4 million nodes a second is very efficient
compared to that.
- He needs more like a trillion nodes a second to compensate for
the inefficiency in hardware. No killermoves. No hashtables etcetera.
Of course the argument that it is possible to make hashtables in
hardware is not relevant as there is a price to that which is too
big to pay simply.
Even for IBM it was too expensive to pay for
hashtables in hardware, despite that Hsu had created possibilities
for it, the RAM wasn't put on the chips and wasn't connected to the
cpu's. Something that improves the chips of course do get used when
they work somehow. Only price could have been the reason? Don't you
think that too? If not what could be the reason to not use hashtables,
knowing they improve efficiency?
the important thing to remember is that if i want to drive to
Paris with 2 cars and i just ship cars in all directions without
looking on a map or roadboard (representing the inefficiency), then
the chance is they land everywhere except on the highway to Paris.
Even a trillion nodes a second isn't going to work if it is using
inefficient forms of search.
It is not very nice from Hsu to focus upon how many nodes a second
he plans to get. For IBM that was important in 1997 to make marketing
with. It is not a fair comparision.
If i go play at world champs 2003 with like 500 processors, i
do not talk about "this program uses up to a terabyte bandwidth
a second (1000000 MB/s) to outpower the other programs, whereas
the poor PC programs only have up to 0.000600 terabyte bandwidth
a second (600MB/s).
That is not a fair comparision. Do you see why it is not a fair
comparision?
He should say what search depth he plans to reach using such
chips.
However he quotes: "search depth is not so relevant". If it is not
so relevant then, why talk about nodes a second then anyway if
the usual goal of more nps (getting a bigger search depth) is
not considered important.
>EeEk(* DM) kibitzes: kib question from Frantic: According to what was
>published DB was evaluating 200 million positions per second (vs 2.5
>to 5 million for the 8-way Simmons server running Deep Fritz). How
>fast would be Beep Blue today if the project continued?
>CrazyBird(DM) kibitzes: it contains a few reference at the end of the
>book for the more technically inclined.
>CrazyBird(DM) kibitzes: if we redo the chip in say, 0.13 micron, and
>with a improved architecture, it should be possible to do one billion
>nodes/sec on a single chip.
>CrazyBird(DM) kibitzes: so a trillion nodes/sec machine is actually
>possible today.
>
>If the cost is not that high maybe Hsu should make ala chessmachine that can be
>plug into computers (assuming that he has no legal obligation from ibm) The
>desktop pc is a long way from hiting 1billion nodes/sec. I think most of the
>professional chessplayers and serious chess hobbyist will buy. He can easily get
>1 million orders. 1 billion nodes/sec, mmm....:)
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.