Page 2 of 2
Re: How to make SF10 work for Arena 1.1 on the Raspberry Pi?
Posted: Sun Mar 24, 2019 10:24 pm
by Raphexon
Nordlandia wrote: ↑Sun Mar 24, 2019 8:59 pm
How about the first version of pi?
Version 1.0 - 1core
Still no chance?
Roughly 20 kn/s according the poster below.
Post was in August 2012, so sf_2.2.2 which has a CCRL of 3113 (SF10 is 3548), it was competitive with Anmon and SOS 5.1 on an i5-2405S CPU @ 2.50GHz. (Both engines can only use 1 core) Both are roughly 2550 CCRL elo (40/4).
https://www.raspberrypi.org/forums/view ... hp?t=15663
http://lefounumerique.clicforum.fr/t120 ... rry-Pi.htm
Fritz 8 was beating top humans on a Centrino 1.7 ghz.
https://en.wikipedia.org/wiki/Human%E2% ... 80%932005)
On passmark it has a single thread rating of 521, the i5 has a single thread rating of 1614.
https://www.cpubenchmark.net/cpu.php?cp ... Hz&id=1159
https://www.cpubenchmark.net/cpu.php?cp ... GHz&id=795
So I'd wager Anmon and Sos on the i5 are roughly equal to Fritz 8 on a 2004 laptop.
So I think Magnus could win against SF 2.2 on the first Raspberry Pi.
SF10 and the latest dev version should smash every human ever, even with just 20 kn/s.
Re: How to make SF10 work for Arena 1.1 on the Raspberry Pi?
Posted: Wed Mar 27, 2019 2:57 am
by MikeB
Nordlandia wrote: ↑Sun Mar 24, 2019 8:59 pm
How about the first version of pi?
Version 1.0 - 1core
Still no chance?
It would depend on the nps - current SF , 10 or higher is playing at solid GM strength at around 25000 NPS. How do I know this - well I don't "know" it - its conjecture in part on my 30 years of computer chess experience and with extensive testing with SF at different NPS rate. SF is different from a lot of engines is it scales much better art the lower NPS rate and then once it gets to Elo 2900 or 3000 , it behaves more like other engines. It's not a steady rise in Elo as speed increases ( say 50 Elo increase represents doubling of Elo - for SF at very low nps ( anywhere under under 10,000 nps) , it may be more like 75 Elo. GMs will struggle with SF mightily over 50,000 nps. Single Core 3B+ Pi's do around 250K nps, but with Pis it's fairly easy to lower the Mhz, so maybe an early Pi zero could be throttle down to 50K nps - I'm not sure - but that would be the target and then recalibrate as necessary. I'm not an expert- just casual anecdotal observations while I was implementing Play By Elo ( UCI Limit Strength in McCain). I have McCain play at ~55000 nps at the 2700 ELo level - my guess is that it's probably50 Elo harder than 2700 - but part of that is that I wanted to make beating a certain level a stretch and not easier than expected. My 2800 level is seems to be broken - will have to a look at it. Elo level 2750 works - its at about 75000 nps. My guess is that it at least 50 Elo underrated , maybe 100 Elo underrated.