I verified perft(15)

[dangi12012]

Today in the night on 17.08.2026 - 00:49:52 my verification and calculation of perft(15) completed.
This calculation effort ran mostly on a single 8x 5090 machine, with 512gb of host memory

Raw GPU time (in terms of 1x gpu hours, including completely idle time, test runs, iterations)

Code: Select all

RTX 5090 x8: 203h 20m 9s
RTX 5090 x4: 182h 15m 41s
RTX 5090: 3h 16m 31s

The last leg yesterday - (resume mechanism so machines can be switched, so it starts at 63%)
Start - End

Code: Select all

[2026-08-16 13:18:48:453] - [GPU]  63.7% 168495518731/~264432813441 roots 14.0s elapsed 14687155.5 GN/s 50046.37 raw GN/s pool 347/378 ~ETA 40739.2s (60s 147546 raw)
[2026-08-17 00:49:36:442] - [GPU] 100.0% 208523003998/~208523003998 roots 41462.0s elapsed 13615389.4 GN/s 58960.15 raw GN/s pool 0/1 ETA     -- (60s 51348 raw)

The node speed hovers around 50-60 TN/s, which the on device TTs speeds up to around 13.62 - 21.2 PN/s

Interesting in that effort is the dedup tree, which is never fully materialized to fit withing constrained memory.

Code: Select all

[2026-08-17 00:49:52:220] - dedup tree (host roots):
[2026-08-17 00:49:52:220] - |- ply  1:           20 unique              20 nodes    1.00x dedup
[2026-08-17 00:49:52:220] - |- ply  2:          400 unique             400 nodes    1.00x dedup
[2026-08-17 00:49:52:220] - |- ply  3:         5362 unique            8902 nodes    1.66x dedup
[2026-08-17 00:49:52:220] - |- ply  4:        72078 unique          197281 nodes    2.74x dedup
[2026-08-17 00:49:52:220] - |- ply  5:       822518 unique         4865609 nodes    5.92x dedup
[2026-08-17 00:49:52:220] - |- ply  6:      9417683 unique       119060324 nodes   12.64x dedup
[2026-08-17 00:49:52:220] - |- ply  7:     96400335 unique      3195901860 nodes   33.15x dedup
[2026-08-17 00:49:52:220] - |- ply  8:    988192872 unique     84998978956 nodes   86.01x dedup
[2026-08-17 00:49:52:220] - |- ply  9:        fused   2439530234167 nodes  (streamed without materializing)
[2026-08-17 00:49:52:220] - \- ply 10:  75670835200 staged  69352859712417 nodes  916.51x dedup (streamed)

How and why - There I also posted worlds first kiwipete P11 results, and first Steven Edwards' symmetric first P10 run as well. I added this runs log too.
viewtopic.php?t=86588

This is the first independent verification of 2,015,099,950,053,364,471,960 and I can confirm correctness.
Signed - Daniel Infuehr

Also - P16 is totally reachable by more than 20x this effort. 20x+ since we will start one ply deep.
If you have access to a gpu cluster PM me.

[dangi12012]

And to make this totally clear - if someone wants to sponsor machines, I am serious.
With 20 machines parallel (each 8x 5090 or better), P16 could be calculated in 2-3 days.