Nuclear will probe its endgame databases in RAM. For example, imagine a R+P vs. R ending. Hit "at a leaf node" from a great distance, the evaluation routine cannot search it. The program makes basically a "0-ply guess" at the value of the position, then reports this score back down the tree.


R+P = +600, R = -500, so it will score the position as +100.

The positions might be a dead draw, so the "real score" should be 0. The program, if it is up a pawn many moves from this leaf node, should avoid swapping into it. Instead, the program will play for it, "thinking" it is up one pawn.

Now imagine your Pawn can be on any of 60 squares, your king on any of the 79 remaining, and your Rook on the 78 left from there. The opponent's Rook can be on any of the other 77 squares, and the king will be on no more than any of 76 squares, with about 64 on average due to all the checks that can take place. That would be about 1,821,980,160 positions, x 2 since either side can move gives you 3,643,960,320 positions with just R + P vs. R!

How many of those 3.6 billion positions do you think Vortex can "guess" right with a 0-ply search? Not many.

Now imagine instead of +100 for being up one pawn, Vortex could return +6320 for mate in 80 moves, or 0 for a dead draw, or -6285 for being mated in 115 moves!

This information is PERFECT, no EVALUATION has to be "guessed", and this will greatly effect the way the program will play, even from a great distance!

There are roughly another 3.6 billion Q+P vs. Q positions, and Arch + pawn vs. Arch, Arch + pawn vs. Chancellor, etc, etc.

What you have is about 300 billion positions in a huge database probed in RAM being delivered to the search engine.

There is no way to create an effective buffer with less than 8 GB of RAM, and I would even recommend 16 GB. Macintosh systems now ship with 8 GB on some of their boxes. 64-bit 8x opterons are at the same stage I believe.

This type of knowledge needs RAM, RAM, and more RAM. But it would be soooo cool.

This software is MONTHS away from completion, not YEARS. The databases are already computed and are disk resident. Writing routines to probe it in RAM has never been done, and this is the last step.