Author: Robert Hyatt
Date: 19:03:31 03/27/01
Go up one level in this thread
On March 27, 2001 at 18:12:42, Christophe Theron wrote: >On March 27, 2001 at 14:08:03, Robert Hyatt wrote: > >>On March 27, 2001 at 12:55:10, Christophe Theron wrote: >> >>> >>> >>>That's right. >>> >>>Actually as the title says, the message is directed to people who are >>>considering to buy a dual. >>> >>>As far as I know quads are so expensive that it would be ridiculous to buy one >>>just to play chess. >>> >>> >>> >> >> >> >>At the moment, perhaps. 6 years ago duals were just as expensive. Now they >>are dirt cheap. As quads become more common, their prices will continue to >>drop. 5 years ago a quad MB for pentium pro 200s would set you back almost >>$8,000. Today you can buy an Intel SC450NX for 2500 bucks, that includes >>three hot-swappable 400 wat power supplies, motherboard, 6-slot hot-swap raid >>disk cage, 3 on-board scsi controllers, 1 on-board video controller, etc. >> >>All you lack is cpus, memory and drives. >> >>That is a huge reduction. The curve is going downward each year. Now the >>quads are slowly reaching reasonable price points while the 8-way boxes are >>way expensive. In 5 years that too will change I'll bet... > > > >So I would advice people who are considering buying a dual right now to delay >their buy by several years... Using that logic you should advise everyone to _never_ buy a computer, because next year will _always_ have a faster processor. > > > > > >>>You are always thinking with unlimited resources in mind! >>> >>>I don't disagree with you here, but in real life there are people wondering if >>>it's worth it to buy a dual. >>> >>>And depending on how much money they can put on it, they will have to choose >>>between a single 1.GHz and a dual 1GHz. >> >>OK... but there the dual will perform like a 1.7ghz machine. Which will >>turn into around 60 rating points improvement. That is not trivially >>ignorable. >> >>Each time I teach a parallel programming course here, I will find around one >>out of every 10 students has a dual-processor machine already. And when I ask >>what they paid, they generally say 500-1000 US bucks... > > > >In the place I live, I cannot even buy a dual. I must order it overseas. > >Bob, there are people outside the United State of America, you know. I ordered mine. There is no "local store" that sells quads here. You can go to any local computer store and find a dual however... Gateway even sells them at the local outlet.. > >You are very lucky to live in a place where you can get all kind of high tech >stuffs for a fraction of your monthly salary, but in other countries a dual >represents a huge amount of money. A dual represents 500 bucks for 2x800 or so. I don't know what the exchange rate is, but that can't be huge... > >For example, a dual represents more than my average monthly salary. You _really_ make less than $500 US dollars per month? > > > > > >>>If you can afford to buy a dual 1.2GHz, then you just stop after reading the >>>first paragraph. >>> >>>If not, then I think the rest is worth reading... >>> >>> >>> >>> >>>>That is flawed. For multiple reasons. The shared hash table holds _most_ EGTB >>>>results after a single probe. The EGTB cache is threaded and shares data read >>>>between the two (or more threads). With the compression scheme Eugene uses, >>>>the reads are kept to a minimum. I have run extensive tests on my quad with >>>>one single 9-gig SCSI drive servicing 4 threads for EGTB reads. I don't see >>>>any severe strangulation due to disk backlogs. most threads are searching >>>>close enough to each other in the tree that they are probing the _same_ >>>>tablebases. The caching Eugene wrote handles this quite well. >>> >>> >>> >>>OK, I admit that I have not done any test on this issue, so your input is >>>appreciated. >>> >>>If my figures are wrong I will publish an update for this text. >>> >>>Do you have any measure of the slowdown expected when 2 thread are accesing >>>intensively the same EGTB files? That would help us to compute the corresponding >>>ELO loss. >> >>I generally don't notice any degradation at all. Mainly because of the large >>well-managed cache buffers, no doubt. But then the operating system also does >>a lot of file caching on top of what Eugene does, and all of this (on a 512mb >>machine) goes a long way toward controlling "disk buzz". > > > >Well on my computer when I set up an endgame position I have my hard disk >working really hard. > >I can hardly see how this poor hard disk could manage to serve two threads >instead of one without some performance penalty. First, the two threads are searching positions that are similar, so they will both probe the same tables for the most part. And since the probe code uses a shared cache for both threads, many reads are eliminated. And once a position is probed, it sticks in the hash table from that point forward anyway... I typically see very little slowdown when reaching 8-10 pieces on the board. In a few rare cases, it does slow way down but it isn't very common with the way I implemented the probe... > >On the other hand, if you need to have a high perf SCSI drive to satisfy the >needs of the dual, and such an amount of memory, this has to be added to the >invoice. SCSI and EIDE are not too far apart. Typically a couple of hundred bucks added for the same size. And EIDE is not horrible for database probes. I run on such machines all the time and don't see any serious performance problem when compared to SCSI. > >Remember that all this is about what you get for the money, what you need really >and is it worth it. Almost all new PCs have 256mb of RAM. 128mb DIMMS are selling for 50 bucks nowadays, so going to 512mb adds $100 to the price of the machine. And the machines I use are pretty much all 512mb boxes... This seems to work acceptably. > > > > > >>>I did not try to cover quads in the message because I don't think many people >>>could afford to buy one. >>> >> >>I realize that. But 5 years ago you wouldn't have found anyond considering >>buying a dual either. Quads will eventually reach the same pricing level, >>based on a curve over the last 5 years.. > > > >People, wait for 5 years before you buy a SMP machine. Don't buy that 1.5ghz machine this year either. Wait for next year's 2.5ghz processors. But wait, don't buy that either as the next year we will be at 4.0 ghz. But why waste the money on that machine when the next year will see 5+ghz processors? As Eugene said, based on _lots_ of experience, I would _much_ prefer a dual 800 to a single 1.2ghz machine. It responds better. It handles two compute- bound processes _far_ better due to less cache thrashing. And it will be cheaper to boot. > > > > > >>>So far I have seen a number of people on CCC asking for duals, but nobody ever >>>said he was considering to buy a quad. >>> >>> >>> >>> >>> >>>>>A difference in ELO points in real life turns into a winning percentage. >>>>>That's exactly what ELO means, and how it is computed. >>>>> >>>>>For winning percentages above 20% and under 80%, there is an approximated >>>>>formula that works pretty well: >>>>> >>>>> ELOdiff = ( WinPercentage - 50 ) * 7 >>>>> >>>>>From this you can deduce how to compute WinPercentage if you have the ELOdiff: >>>>> >>>>> WinPercentage = ELOdiff / 7 + 50 >>>>> >>>>>If ELOdiff=25, then WinPercentage = 53.57% (we are between 20% and 80% >>>>>so our above formula applies). >>>>> >>>>>So we are talking about a difference of 3.5 games each time you play 100. >>>>> >>>>> >>>>>**************************************************************************** >>>>>** When you play 100 games with your dual 1GHz against ** >>>>>** your single 1.2GHz, you can expect the dual to win typically ** >>>>>** by a 3.5 games margin. ** >>>>>**************************************************************************** >>>> >>>> >>>>I would change that to >>>> >>>>winpct=60/7+50 which is about 60%. Out of 100 games that turns into winning >>>>60 and losing 40. BTW in your above comment you need to double that 3.5. If >>>>I win 53.5 games out of 100, you win 46.5. The _difference_ is 7 games. Not >>>>3.5 >>> >>> >>>When you win a game, your opponent loses it. I don't count this as 2 games. >>> >>> >>> >>> >> >>Then maybe your term wasn't clear to me instead. You said "I win 53.5% of >>the games. Out of 100 games that is a difference of 3.5 games." If I win >>53.5% of the games, you win 46.5% of the games. That is a bit different >>since our scores are separated by 7, not 3.5... > > > >OK, OK. I don't want to split hairs. > >What counts is the difference between the winning percentage and 50%. Because it >is what you multiply by 7 to get an estimate of the ELO difference. > >If I run a match with even hardware, I win 50 out of 100 games. > >If I run a match with a dual I win 53.5 out of 100 games. > >With my dual I win 3.5 more games. OK? > > > > Christophe OK. but I will bet that you win more than 3.5 more games. Because the actual performance gain is still in the 1.7X range overall, maybe better in a full game (I get 1.7X over large problem sets. chess programs do better in a game as they typically search deeper due to pondering.
This page took 0.01 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.