I don't know where you are getting the 1000x factor from - I'm looking at 1994 and the pentium with the 200 MHZ processors were available then. It's a lot slower than a single core of an i7, but it's nothing like 1000x slower. You have to go back a lot farther than that to see 1000x.
But I'm hoping someone here actually still has a old 1994 vintage computer (the best of the day if we compare it to i7) and we can run chess genius or crafty or something on it and test it.
Meanwhile, I'm going to run 1 or 2 games per day with chess genius vs Robbolitto.
I'm going to adjust the time control for each match in an attempt to find the rough break even point. I should be able to get within 100 ELO this way without having to play millions of games.
Don
move count based pruning
Moderator: Ras
-
- Posts: 20943
- Joined: Mon Feb 27, 2006 7:30 pm
- Location: Birmingham, AL
Re: move count based pruning
I am talking pure NPS. I will see if I can find any printed information. I do remember that when the P6/200 came out, we were able to hit around 70K nodes per second. This was available in early 1996 and we used it in the 1996 WMCCC event. The P5/200 was a pure dog compared to the OOE P6 chip. The first P5 we bought was early in 1994, and this is the chip I was discussing as representative of that time frame. In looking back, by the middle of 1994 we had the p5/133 boxes here, having started with a 60mhz and then 90mhz version.Don wrote:I don't know where you are getting the 1000x factor from - I'm looking at 1994 and the pentium with the 200 MHZ processors were available then. It's a lot slower than a single core of an i7, but it's nothing like 1000x slower. You have to go back a lot farther than that to see 1000x.
But I'm hoping someone here actually still has a old 1994 vintage computer (the best of the day if we compare it to i7) and we can run chess genius or crafty or something on it and test it.
Meanwhile, I'm going to run 1 or 2 games per day with chess genius vs Robbolitto.
I'm going to adjust the time control for each match in an attempt to find the rough break even point. I should be able to get within 100 ELO this way without having to play millions of games.
Don
I believe Crafty peaked around 30K nodes per second on the fastest P5 I ever ran on. Compared to 100M two years ago. So 1000x is not completely unrealistic, IMHO.
-
- Posts: 20943
- Joined: Mon Feb 27, 2006 7:30 pm
- Location: Birmingham, AL
Re: move count based pruning
I believe the best I have is a 400mhz pentium something in an old sony vaio laptop. But it is certainly post-OOE introduction so it is not going to compare well to a pure in-order p5 processor. I'm looking thru the junk pile as we speak.Don wrote:Does anyone have a 200 MHZ Pentium machine from 1994? Would like to do some timing tests on chess computers. One that runs DOS would be perfect!
Don
bob wrote:About the only data I could provide would be to run a match between Crafty and Cray Blitz on current hardware, then figure out how to slow the hardware down by something like 1000X (in 1995 Crafty did 5K, today 20M so maybe more like 4000X for the hardware speed change in the PC from the original Pentium-whatever in late 1994 to early 1995, compared to today's speeds (the best I have hit is 100M on an 8x4 prototype server-type box about 2 years ago, we maybe my 4,000x is still too low).Don wrote:The issue is how much different are they in software, then we compare the hardware too. We don't need to debate it, we just need to run the experiment if we can.bob wrote: So? Nobody says there has been _no_ software improvements. But what happens with the same two programs on current hardware? Would not be surprised for them to be about the same distance apart, yet worlds stronger...
In that case, which provides more, hardware or software? I suspect this is a lot clearer than most want to admit.
In 1996 Crafty finished in 3rd and 4th places (Crafty and the gunda-1 clone) at the WMCCC event, so it was pretty representative of computer chess engines of the time. Only problem is that the Cray Blitz version I have is not going to work with multple-cpus, it was built around some specific stuff Cray used (Task common for one). So maybe we can just figure out single-cpu comparisons. I know we did 5K on an early pentium. I am not sure if was a -66, -90, or -133 however. I know that I can hit 5M on a single CPU I7 today. So that is back to the 1,000x hardware.
So for fun, let's take a fast I7 today, single-cpu, against a single-cpu from the past. How far back to go? I can't go back beyond 1994-1995 for Crafty since it played on ICS (now ICC) in December of 1994 and didn't exist prior to that. Therefore no guess about what 486 and earlier CPUs would have done.
Given a hardware factor, I should at least be able to compare CB vs Crafty on old, and CB vs Crafty on new, to see if hardware speed helps one more than the other. My suspicion would be that CB would gain more, since it's branching factor was much higer, as was its accuracy. But I'm not certain. I'd be concerned about crafty on old and CB on new, or vice-versa, as that is a _huge_ time handicap, perhaps enough to produce 30000-0 results, or something real close to it, which wouldn't provice any information...
I am not even sure you can take an old commercial program like fritz and compare its old rating to the same program on new hardware, since the comparison would involve some inflation in ratings produced over the years...
I think measuring Elo of an old engine on old hardware is going to be more than just problematic.

-
- Posts: 13447
- Joined: Wed Mar 08, 2006 9:02 pm
- Location: Dallas, Texas
- Full name: Matthew Hull
Re: move count based pruning
If two programs could be slowed (nps-wise) on modern hardware by introducing do-lots-of-nothing loops at similar logic points, and limiting memory for each one, then a comparison on "slow hardware" could be attempted perhaps.bob wrote:I am talking pure NPS. I will see if I can find any printed information. I do remember that when the P6/200 came out, we were able to hit around 70K nodes per second. This was available in early 1996 and we used it in the 1996 WMCCC event. The P5/200 was a pure dog compared to the OOE P6 chip. The first P5 we bought was early in 1994, and this is the chip I was discussing as representative of that time frame. In looking back, by the middle of 1994 we had the p5/133 boxes here, having started with a 60mhz and then 90mhz version.Don wrote:I don't know where you are getting the 1000x factor from - I'm looking at 1994 and the pentium with the 200 MHZ processors were available then. It's a lot slower than a single core of an i7, but it's nothing like 1000x slower. You have to go back a lot farther than that to see 1000x.
But I'm hoping someone here actually still has a old 1994 vintage computer (the best of the day if we compare it to i7) and we can run chess genius or crafty or something on it and test it.
Meanwhile, I'm going to run 1 or 2 games per day with chess genius vs Robbolitto.
I'm going to adjust the time control for each match in an attempt to find the rough break even point. I should be able to get within 100 ELO this way without having to play millions of games.
Don
I believe Crafty peaked around 30K nodes per second on the fastest P5 I ever ran on. Compared to 100M two years ago. So 1000x is not completely unrealistic, IMHO.
Matthew Hull
-
- Posts: 5106
- Joined: Tue Apr 29, 2008 4:27 pm
Re: move count based pruning
100 million nodes per second 2 years ago? Stockfish does something like only 2 million so Crafty is about 50x faster in nodes per second?bob wrote:I am talking pure NPS. I will see if I can find any printed information. I do remember that when the P6/200 came out, we were able to hit around 70K nodes per second. This was available in early 1996 and we used it in the 1996 WMCCC event. The P5/200 was a pure dog compared to the OOE P6 chip. The first P5 we bought was early in 1994, and this is the chip I was discussing as representative of that time frame. In looking back, by the middle of 1994 we had the p5/133 boxes here, having started with a 60mhz and then 90mhz version.Don wrote:I don't know where you are getting the 1000x factor from - I'm looking at 1994 and the pentium with the 200 MHZ processors were available then. It's a lot slower than a single core of an i7, but it's nothing like 1000x slower. You have to go back a lot farther than that to see 1000x.
But I'm hoping someone here actually still has a old 1994 vintage computer (the best of the day if we compare it to i7) and we can run chess genius or crafty or something on it and test it.
Meanwhile, I'm going to run 1 or 2 games per day with chess genius vs Robbolitto.
I'm going to adjust the time control for each match in an attempt to find the rough break even point. I should be able to get within 100 ELO this way without having to play millions of games.
Don
I believe Crafty peaked around 30K nodes per second on the fastest P5 I ever ran on. Compared to 100M two years ago. So 1000x is not completely unrealistic, IMHO.
-
- Posts: 900
- Joined: Tue Apr 27, 2010 3:48 pm
Re: move count based pruning
I think 1000x is probably about right... I recently did a test on a Pentium 4 1.6 GHz and it was about 100x slower than a Phenom II X6 2.8 GHz. A Pentium should be at least 10x slower than that Pentium 4.
-
- Posts: 900
- Joined: Tue Apr 27, 2010 3:48 pm
Re: move count based pruning
2 million with one core, right? He was talking about a 32-core server for the 100M nps number.Don wrote:
100 million nodes per second 2 years ago? Stockfish does something like only 2 million so Crafty is about 50x faster in nodes per second?
We should definitely not compare a octa-cpu systerm of today with a single-cpu of 15 years ago though. That's not a fair comparison.
-
- Posts: 5106
- Joined: Tue Apr 29, 2008 4:27 pm
Re: move count based pruning
I'm looking only for round numbers anyway. I believe software and hardware are very roughly equally responsible for the advances over the years. Until we figure this out, let's give 1994 and 2010 about 1000 to 1 for single core performance. We are probably not off by more than a couple of years anyway.rbarreira wrote:I think 1000x is probably about right... I recently did a test on a Pentium 4 1.6 GHz and it was about 100x slower than a Phenom II X6 2.8 GHz. A Pentium should be at least 10x slower than that Pentium 4.
-
- Posts: 20943
- Joined: Mon Feb 27, 2006 7:30 pm
- Location: Birmingham, AL
Re: move count based pruning
Crafty is doing 20M-30M nodes per second on my 3-year old dual quad box I have been using. The 100M number was on a machine with 32 cores, about 2 years ago or so. It was an 8-cpu quad-core server box prototype. Nothing particularly fancy or pricey... I have seen 30M-40M on a dual-quad Nehalem box we had here for a while for evaluation, this about 6-8 months ago.Don wrote:100 million nodes per second 2 years ago? Stockfish does something like only 2 million so Crafty is about 50x faster in nodes per second?bob wrote:I am talking pure NPS. I will see if I can find any printed information. I do remember that when the P6/200 came out, we were able to hit around 70K nodes per second. This was available in early 1996 and we used it in the 1996 WMCCC event. The P5/200 was a pure dog compared to the OOE P6 chip. The first P5 we bought was early in 1994, and this is the chip I was discussing as representative of that time frame. In looking back, by the middle of 1994 we had the p5/133 boxes here, having started with a 60mhz and then 90mhz version.Don wrote:I don't know where you are getting the 1000x factor from - I'm looking at 1994 and the pentium with the 200 MHZ processors were available then. It's a lot slower than a single core of an i7, but it's nothing like 1000x slower. You have to go back a lot farther than that to see 1000x.
But I'm hoping someone here actually still has a old 1994 vintage computer (the best of the day if we compare it to i7) and we can run chess genius or crafty or something on it and test it.
Meanwhile, I'm going to run 1 or 2 games per day with chess genius vs Robbolitto.
I'm going to adjust the time control for each match in an attempt to find the rough break even point. I should be able to get within 100 ELO this way without having to play millions of games.
Don
I believe Crafty peaked around 30K nodes per second on the fastest P5 I ever ran on. Compared to 100M two years ago. So 1000x is not completely unrealistic, IMHO.
For this "comparison" do we ignore the most significant hardware advances over the past 5-6 years, which has mainly been multi-core???
-
- Posts: 20943
- Joined: Mon Feb 27, 2006 7:30 pm
- Location: Birmingham, AL
Re: move count based pruning
I disagree. We are talking about comparing technology vs software, correct? 14 years ago there were no dual-core boxes. In 1995 there were no significant dual-cpu boxes. I bought a quad pentium-pro 200 in 1997, and it was way beyond $10,000 in price. Over the past 4-5-6 years the primary advancements have been in multiple-cores, so I am not sure why if we are going to measure hardware improvements, we don't count _all_ improvements...rbarreira wrote:2 million with one core, right? He was talking about a 32-core server for the 100M nps number.Don wrote:
100 million nodes per second 2 years ago? Stockfish does something like only 2 million so Crafty is about 50x faster in nodes per second?
We should definitely not compare a octa-cpu systerm of today with a single-cpu of 15 years ago though. That's not a fair comparison.