Woaw! So you're being modest, but in reality, your TB beats the pants out of everything out there. And yes, probing speed will surely be linked to size (disk access much slower than RAM). The size of your TB is more comparable to the size of a bitbase than a TB, so it's really a big improvement over Nalimov and Gaviota.syzygy wrote:The 5-piece Nalimov TB are over 7 GB. I believe the 6-piece are over 1.2 TB (and those lack the 5v1 tables).lucasart wrote:How does the size of 5-men table size base compare with: the Nalimov TB ? the Gaviota TB ? And in terms of "average probing speed" ?
The 5-piece Gaviota TB are 6.5GB with LZMA compression, see here.
Average probing speed of my tables should be considerably better than that of Nalimov tables, but to be honest I have not measured it. But probing 1.2 TB Nalimov 6-piece tables will of course lead to much more disk trashing than probing 68.3 GB of tables. And 68.3 GB on SSD is much more affordable than 1.2 TB on SSD.
Would it be possible to modify your code to generate a bitbase only:
- one bit per position (disregarding 50-move counter)
- bit just says if it's a draw or not.
if not a draw, good engines will (in the vast majority of cases) be able to find the win on their own. the real difficulty is to understand that certain positions are a "no progress situation", despite a material advantage, or an apparent advantage in terms of eval score.
A very small bitbase with only draw info, that could fit in RAM for 5-men, would be really nice.