Stockfish 8 scalability (1 core testing)
Time control |
Hash |
Result (200
games) |
Elo gain |
10+0.1 vs 5+0.05 |
4 MB |
82-12-106 |
127.0 |
20+0.2 vs 10+0.1 |
8 MB |
76-11-113 |
117.2 |
40+0.4 vs 20+0.2 |
16 MB |
70-8-122 |
111.4 |
80+0.8 vs 40+0.4 |
32 MB |
57-10-133 |
83.2 |
160+1.6 vs 80+0.8 |
64 MB |
56-11-133 |
79.5 |
320+3.2 vs 160+1.6 |
128 MB |
41-5-154 |
63.2 |
640+6.4 vs 320+3.2 |
256 MB |
31-6-163 |
43.7 |
1280+12.8 vs 640+6.4 |
512 MB |
29-5-166 |
42.0 |
On these data I applied the least squares method for
my theoretical model (ELO = A * ln(depth) + C),
and the model explained 99.87% of the variability of the measured data (see
the image below).
The absolute value of the Elo is provisional.
Conclusions:
1) Somewhat lower Elo gain is expected under hash pressure;
2) Somewhat lower Elo gain is expected on multicore systems;
3) Scalability is almost perfect up to depth 20;
4) Some possible improvements are expected on higher depths.