LC0 finally?

[megamau]

It is actually quite simple.

It show how different nets (column C) behave with time odds against Stockfish 9.

Column D shows the results when both Leela and Stockfish use only 1 node per move.
Column E shows the results when they both search 10 nodes per moves etc.

The green columns use same number of nodes (which means Leela has much more time), the yellow columns use 10 times more nodes for Stockfish (still time advantage to Leela), ....., all the way to the blue one at the end where Stockfish uses 10 millions more nodes (and thus also more time).

The results are "relative elo". When the values in column N, O, P , Q (1000 time less nodes) will reach 0, it will mean that Leela is on par with Alphazero.

[Milos]

The results are "relative elo". When the values in column N, O, P , Q (1000 time less nodes) will reach 0, it will mean that Leela is on par with Alphazero.

That is not actually true. With more nodes LC0 is actually scaling worse, i.e. LC0 performance in 1:1000 nodes is much, much better than in case of 80k:80M. However, 80k:80M could be worse (due to SF's "bad" scaling on many cores) than 10k:10M, where roughly the worst LC0:SF performance would be.