Jouni wrote: ↑Tue Aug 07, 2018 4:24 pm
Even if Leela were at Stockfish dev level I am not interested yet! Because 1. You need additional hardware - no thanks and 2. it can't solve any difficult test position.
The additional hardware issue is a bogus one. We are too much into CPU paradigm, while the things in the last 2-3 years changed seriously. Now GPUs are much trendier in frontlines like cryptocurrencies and Deep Learning. Many people in these fields spend 70-80% on GPUs and 10-20% on CPUs. I compared apples to apples price-wise: my GPU GTX 1060 6GB is very close in price to my CPU i7-4790, maybe 10% cheaper. I am already considering buying a second GTX 1060 GPU, probably when the next Nvidia line will appear on the market, so price-wise, my PC will become GPU dominated. I personally know folks in bitcoin mining having arrays of 10-20 GPUs and only one Xeon CPU.
Your second point is referring to tactical test suites, very popular on Talkchess. I am not sure why such a plethora of people here are fascinated by these puzzles, most of the puzzles posted here almost never occur in real games. It's probably an inertia from magazines of the distant past with "solve mate in 3", and people of Talkchess cannot adjust to present times. On my positional test-suite, some of the nets of Leela are performing very well, just behind top 3 engines.
Karol Majewski wrote: ↑Tue Aug 07, 2018 5:24 pm
OK, so I have this CPU version of LC0 v 0.16 (blas). How strong is it on a single CPU with the latest network (573)?
Maybe Fruit 2.1 level in CCRL 40/4 conditions. Haven't checked on CPU long time. Also, always specify TC.
A comparison is easy to make and the unit of measurement is, of course, the Wattage. It may not be in favor of CPU's but that is a measure of computing effort
Personally, I do not use the engines as "horses" but only as analysis tools. LC0 has a much better position understanding than any other engine by a mile.
It still has a lot of tactical issues but those will be ironed out as training continues. It has also a lot to learn in endgames, especially what is drawn and what is not, but I am sure it will generalize greatly in that aspect too.
An extra benefit is that the solution scales perfectly with hardware increase.
My phones' battery holds 13Wh. SF 010718 on my phone can run on 8 cores for 4 hours straight with approx 5000knps until battery is dead, so it's eating ~3 watts per hour/5000knps. PC running SF at 10x5000knps=50000knps should consume 30 watts per hour than. Does it make any sense regarding wattage comparisons, or I missed something?
flither wrote: ↑Tue Aug 07, 2018 10:23 pm
...and Leela is not 3497 Elo. People, stop spreading fabricated ratings
No, last time I checked, it was 6200+ Elo on main server nets. So, it's more than twice as strong as Magnus Carlsen, who is barely 2850 Elo. But if you add Carlsen + Caruana + Aronian, they will have more Elo than Leela.
Karol Majewski wrote: ↑Tue Aug 07, 2018 5:24 pm
OK, so I have this CPU version of LC0 v 0.16 (blas). How strong is it on a single CPU with the latest network (573)?
Maybe Fruit 2.1 level in CCRL 40/4 conditions. Haven't checked on CPU long time. Also, always specify TC.
Lc0 0.16 on 1060 (net 570) that gets 7200nps is around 29% vs. latest SFdev on single core on 40/4 TC. That is roughly 250Elo below SFdev.
Lc0 0.16 blas on single core is around 24nps. That is like 300x or like 8.3 doublings weaker than on 1060. If we assume a doubling is like 60 Elo (conservative) that is at least another 500Elo in addition to those 250Elo. Assuming latest SFdev on 40/4 single core is around 3520Elo, Lc0 0.16 single core is not stronger than 2750Elo, which is around Fruit 2.1.