cache sizes
Tom Kerrigan
I think it might depend on the program. For example, I hear that MChess
has a great big eval function. My guess is that it uses a lot of tables
and stuff that might fit well in cache, in which case you might be able
to expect over 10%. Fritz, on the other hand, pretty much doesn't need
any memory at all (grin) so I doubt you would see any speedup with it.
Actually, a lot of programs also have tiny computer counterparts, or at
least the programmers have programmed dedicated machines, so I doubt
most of the programs would really thrash with 256k cache. In other words,
I doubt the upgrade would be worth the money/trouble.
Cheers,
Tom
James Long (wzrd...@fia.net) wrote:
: What kind of performance increase could I expect if I upgraded
: my 256k cache to 512k? Is +10% nps a reasonable estimate?
: James
Vincent Diepeveen
>I think it might depend on the program. For example, I hear that MChess
This is terribly right: it depends heavily on the program used.
For example:
Diep at a P166+ at 60ns is considerable faster than the same P166+
at 70ns. Same for all intel processors except PP200.
Kallisto at the same P166+ at 60ns however didn't run faster than
it ran at 70ns.
This P166+ test was done by Bart Weststrate, and he explained
the reason: the total datastructure of Kallisto probably (for sure
is perhaps better word) including code fits into 256kb code.
Of course hashtables are not counted.
Didn't ask the size of the level2 cache on bart's P166+,
but the Diep datastructure cannot even fit in 512kb, that's for sure.
Only my move generator already is easting about 180kb, just because
i use 'int' in my program, which translates to 4 bytes, where if
i would have written my program in assembly, i would use 1 byte for
every move, which gives 180/4 = 45kb, which bye the way can be reduced
to less than 20kb, without loosing speed, if i would invest time in
it.
So in this case using a higher level language is a disadvantage.
Vincent.
>has a great big eval function. My guess is that it uses a lot of tables
>and stuff that might fit well in cache, in which case you might be able
>to expect over 10%. Fritz, on the other hand, pretty much doesn't need
>any memory at all (grin) so I doubt you would see any speedup with it.
>Actually, a lot of programs also have tiny computer counterparts, or at
>least the programmers have programmed dedicated machines, so I doubt
>most of the programs would really thrash with 256k cache. In other words,
>I doubt the upgrade would be worth the money/trouble.
Buy a K6 with a cheap socket-7 mainboard, from for example antron.
See tom's hardware page.
If you are really out of money, buy a K6-166 and overclock it to 200Mhz.
Don't forget to cool it well. Buy an old Big tower too (or preferred new
one). They cool better. ATX even better (but if you can afford ATX,
then you can buy a K6-233 with SDRAM and some extra cool blocks and coolers
and and overclock it without problems to 3 x 100 = 300Mhz. Should
be a beast for relatively little money compared to PentiumII.
Anyone knows how well SDRAM at 12 ns (don't confuse this with the
60ns DIMMs most want you to sell) perform for chessprograms?
>Cheers,
>Tom
>
>James Long (wzrd...@fia.net) wrote:
>: What kind of performance increase could I expect if I upgraded
>: my 256k cache to 512k? Is +10% nps a reasonable estimate?
>
>: James
--
+----------------------------------------------------+
| Vincent Diepeveen email: vdie...@cs.ruu.nl |
| http://www.students.cs.ruu.nl/~vdiepeve/ |
+----------------------------------------------------+
Howard Exner
Vincent Diepeveen <vdie...@cs.ruu.nl> wrote in article
<5sv575$fsl$1...@krant.cs.ruu.nl>...
> In <33EA66...@fia.net> James Long <wzrd...@fia.net> writes:
>
> >What kind of performance increase could I expect if I upgraded
> >my 256k cache to 512k? Is +10% nps a reasonable estimate?
> >
> >James
>
>
> If you use Diep, then it gives you a lot, especially if you have
> a INTEL Pentium.
>
> Previous versions of Diep were reported faster on P200MMX than on
> the PP200. Lately i added some knowledge, improved my q-search and
> move ordering and voila see the next results i this time did
MYSELF.
> So i checked that it used the same amount of hashtables, the same
> keyboard speed, and the same boot-sequence (starting with F5, not
> even loading a mouse for the test).
Would a motherboard with 1MB L2 Cache increase chess performance over
512 KB ? I'm considering the purchase of the new FIC PA-2007 MO and
it has the 1 MB upgrade option. It is a TX board with the new AMD-640
chipset.
Does chess performance get a boost from higher bus speeds or is it
just raw internal processor speed that matters?
brucemo
Howard Exner wrote:
> Would a motherboard with 1MB L2 Cache increase chess performance over
> 512 KB ? I'm considering the purchase of the new FIC PA-2007 MO and
> it has the 1 MB upgrade option. It is a TX board with the new AMD-640
> chipset.
>
> Does chess performance get a boost from higher bus speeds or is it
> just raw internal processor speed that matters?
It depends upon what the program does. Mine hits memory very hard so it
benefits from improvements in the components that surround the processor.
I bet most other programs are this way to greater or lesser degree.
A chess program doesn't usually need a monstrously fast hard disk, but
processor, cache, memory, improvements in these help a lot.
bruce
Komputer Korner
Howard Exner wrote:
> Would a motherboard with 1MB L2 Cache increase chess performance over
> 512 KB ? I'm considering the purchase of the new FIC PA-2007 MO and
> it has the 1 MB upgrade option. It is a TX board with the new AMD-640
> chipset.
>
> Does chess performance get a boost from higher bus speeds or is it
> just raw internal processor speed that matters?
Any time you use hash tables, bus speed has to be very important.
--
Best regards
Komputer Korner
The inkompetent komputer
If you see a 1 in my email address, take it out.
Note that my true email is still kor...@netcom.ca
I don't often check the email of the sympatico address.
Vincent Diepeveen
>What kind of performance increase could I expect if I upgraded
>my 256k cache to 512k? Is +10% nps a reasonable estimate?
>
>James
If you use Diep, then it gives you a lot, especially if you have
a INTEL Pentium.
Previous versions of Diep were reported faster on P200MMX than on
the PP200. Lately i added some knowledge, improved my q-search and
move ordering and voila see the next results i this time did MYSELF.
So i checked that it used the same amount of hashtables, the same
keyboard speed, and the same boot-sequence (starting with F5, not
even loading a mouse for the test).
The speed index of Diep 1.57.20:
All tests done with 60NS edoram, except for the P133 and P100,
there 70NS memory was used.
P100 100
P133 133
P200 200
P200-MMX 260
K6-200Mhz (3 x 66 bus) 315.5
PP200-256kb cache (3x 66Mhz bus) 300
K6-233Mhz (3.5 x 66 bus) 344.5
PP200 running on 240mhz: (4x 60bus) 345
So MMX is slower now, and considerable slower.
Probably the speed and the size of the level 2 cache
(i'd estimate that the speed of the level 2 cache decreases
the prestation of Diep).
The level2 cache of the Pro dramatically improves Diep.
The K6 gives an awfull lot for its money. At higher speeds however,
the K6 resigns because the bus is not running that fast.
Perhaps a K6-233 running on 3x 100Mhz bus = 300Mhz improves things?
Also there is a difference between 70ns and 60ns, even when
the level 2 cache 512kb.
On a P133, Ernst Walet showed that the difference between
70ns and 60ns was considerable, but less than 5% (about 3-4).
A PP200 set at 3.5 x 66 Mhz was 2% slower than
my PP200 set at 4 x 60 Mhz = 240Mhz
My PP200 set at 4x60Mhz (ASUS mainboard these asus guys are really
making good stuff, although expensive!) is giving in fact 233 Mhz,
when looking after the speed of Diep.
Setting a K6-200 to 233 (3.5x instead of 3x, bus still at 66Mhz),
only improved node count with 10%.
So a wonderful performance of the K6, considering that
the K6-200 with a good cooler can run on 233Mhz, giving the same speed
a PP200-S with an even better and terribly more expensive cooler at 240Mhz.
The K6-200 chip with mainboard is about 3 times cheaper than
a PP200-S with asus mainboard (but i bought the PP200 9 months ago,
just after DCCC96 started).
Vincent
Martin Borriss
In article <33F387D8...@netcom.ca>,
Komputer Korner <kor...@netcom.ca> writes:
>Howard Exner wrote:
>
>> Would a motherboard with 1MB L2 Cache increase chess performance over
>> 512 KB ? I'm considering the purchase of the new FIC PA-2007 MO and
>> it has the 1 MB upgrade option. It is a TX board with the new AMD-640
>> chipset.
>>
>> Does chess performance get a boost from higher bus speeds or is it
>> just raw internal processor speed that matters?
>
> Any time you use hash tables, bus speed has to be very important.
Perhaps it is different in Canada, but if you ask for bus speed you are
usually told IO bus speed (e.g. 132 MByte/s max. for normal PCI). From the
context it is clear that memory bandwidth is meant. So if Korner sends his
customers to the next dealer they likely ask about the wrong number.
Martin
Robert Hyatt
Martin Borriss (bor...@inf.tu-dresden.de) wrote:
: In article <5sv5u6$g9h$1...@krant.cs.ruu.nl>,
: vdie...@cs.ruu.nl (Vincent Diepeveen) writes:
: >
: >Anyone knows how well SDRAM at 12 ns (don't confuse this with the
: >60ns DIMMs most want you to sell) perform for chessprograms?
: No significant speedup for my program, but this probably doesn't say much.
: Martin
: --
: Martin....@inf.tu-dresden.de
: .
The problem with SDRAM is that it is no faster than normal DRAM. It can
"stream" data to the CPU faster, *after* the first word is fetched, but
fetching that first word is just as slow as normal DRAM memory. And for
programs that are doing lots of ramdom accessing of small blocks of
data, it doesn't help. It would be quite a bit better forprograms that
move blocks of data around, like the very old crafty versions, for
example..
Robert Hyatt
Komputer Korner (kor...@netcom.ca) wrote:
: Robert Hyatt wrote:
: > :
: >
: > Actually, hashing represents only a small part of the total memory
: > bandwidth. IE, we probe once at the start of a ply, but then do a
: > lot of other things many times. The most memory-intensive part of
: > a chess program is typically move generation, where it tries to
: > shove 38+ moves to memory at one time. Next is probably evaluation
: > where lots of data structures are read/compared...
: >
: > In short, memory bandwidth is very low, but the requirement is
: > very high. Faster is almost always better... :)
: Bob wouldn't the on- chip cache hold most of the information a chess
: program needs instead of having to fetch it from memory?
Depends. remember that the 256kb P6/200 cache is only 64K words. That
is not a lot of data. And that cache is direct-mapped, which makes it
very sensitive to where the program is loaded into memory (many have
noticed this problem and reported it frequently here.)
However, for a move table, on a single square, you would have 27 queen
moves, 14 rook moves, 13 bishop moves, 8 knight moves, 8 king moves and
4 pawn moves. total= 74 words. X64 squares and it is getting bigger.
And when you factor in all the other data structures needed (the large
Zobrist random number table for hashing, (8 bytes x 64 squares x 13 pieces)
plus the instructions in the search engine (which goes way up into the
thousands of course) everything doesn't quite tuck away in cache and stick
there. And then there are interrupts that have to be handled fairly
frequently (clock is but one) that causes things to be zapped in cache
as well. It's just not big enough. The alphas (1 meg and up) are
"getting there" but even they are not large enough until you get to the
4 meg level...
Robert Hyatt
Komputer Korner (kor...@netcom.ca) wrote:
: Howard Exner wrote:
: > Would a motherboard with 1MB L2 Cache increase chess performance over
: > 512 KB ? I'm considering the purchase of the new FIC PA-2007 MO and
: > it has the 1 MB upgrade option. It is a TX board with the new AMD-640
: > chipset.
: >
: > Does chess performance get a boost from higher bus speeds or is it
: > just raw internal processor speed that matters?
: Any time you use hash tables, bus speed has to be very important.
: --
: Best regards
: Komputer Korner
Komputer Korner
Robert Hyatt wrote:
> :
>
> Actually, hashing represents only a small part of the total memory
> bandwidth. IE, we probe once at the start of a ply, but then do a
> lot of other things many times. The most memory-intensive part of
> a chess program is typically move generation, where it tries to
> shove 38+ moves to memory at one time. Next is probably evaluation
> where lots of data structures are read/compared...
>
> In short, memory bandwidth is very low, but the requirement is
> very high. Faster is almost always better... :)
Bob wouldn't the on- chip cache hold most of the information a chess
program needs instead of having to fetch it from memory?
--
Best regards
Komputer Korner
The inkompetent komputer
Martin Borriss
Vincent Diepeveen
In <01bca8d5$b52f2f00$4125...@dlcwest.dlcwest.com> "Howard Exner" <hex...@nospam.dlcwest.com> writes:
>
>Vincent Diepeveen <vdie...@cs.ruu.nl> wrote in article
><5sv575$fsl$1...@krant.cs.ruu.nl>...
>> In <33EA66...@fia.net> James Long <wzrd...@fia.net> writes:
>>
>> >What kind of performance increase could I expect if I upgraded
>> >my 256k cache to 512k? Is +10% nps a reasonable estimate?
>> >
>> >James
>>
>>
>> If you use Diep, then it gives you a lot, especially if you have
>> a INTEL Pentium.
>>
>> Previous versions of Diep were reported faster on P200MMX than on
>> the PP200. Lately i added some knowledge, improved my q-search and
>> move ordering and voila see the next results i this time did
>MYSELF.
>> So i checked that it used the same amount of hashtables, the same
>> keyboard speed, and the same boot-sequence (starting with F5, not
>> even loading a mouse for the test).
>
>Would a motherboard with 1MB L2 Cache increase chess performance over
>512 KB ? I'm considering the purchase of the new FIC PA-2007 MO and
>it has the 1 MB upgrade option. It is a TX board with the new AMD-640
>chipset.
>
>Does chess performance get a boost from higher bus speeds or is it
>just raw internal processor speed that matters?
1 MB L2 cache should considerably improve the performance of Diep.
I consider buying a 1MB L2 cache motherboard myself in september.
Should be optimized for K6 too.
I'm wondering how much M2 processor running at 233 or 266 Mhz
would give Diep. Are there tests done with M2 processors on
chessprograms? On Hiarcs for example (that compares a little with
Diep, as Hiarcs seems also faster on Pro than on MMX, which pleas
for Hiarcs)?
Vincent Diepeveen
>Howard Exner wrote:
>
>> Would a motherboard with 1MB L2 Cache increase chess performance over
>> 512 KB ? I'm considering the purchase of the new FIC PA-2007 MO and
>> it has the 1 MB upgrade option. It is a TX board with the new AMD-640
>> chipset.
>>
>> Does chess performance get a boost from higher bus speeds or is it
>> just raw internal processor speed that matters?
>
> Any time you use hash tables, bus speed has to be very important.
Yes, for processors depending for speed of level 2 cache on busspeed,
this is very important.
For Pro processors bus speed is not that important when compared to K6.
For processors depending on bus speed so much also the speed of the
memory becomes important (but in the less than 5% range) for Diep.
>--
>
>Best regards
>Komputer Korner
>
>The inkompetent komputer
>
>If you see a 1 in my email address, take it out.
>Note that my true email is still kor...@netcom.ca
>I don't often check the email of the sympatico address.
>
>