PioSOLVER 3.10 - PioSolver is FASTER, saves are SMALLER.
2x FASTER with save files 10x SMALLER.
This release brings two major updates that will benefit every PioSolver user.
New algorithm
In this version of PioSolver, we have replaced the solving algorithm.
The new algorithm has different characteristics and is not a simple speed improvement over the existing one.
- V4 Algorithm converges on more trees to very high accuracies.
- The bigger and more complex the tree, and the higher the desired accuracy - the higher is the advantage of the V4 algorithm.
- On average the new algorithm is 1.5x - 2x faster.
Check out visualisation of more detailed benchmark on our benchmark page
New save format
We have added a mechanism to largely reduce the size of saved files.
The actual size gain depends on the actual tree, the accuracy and the compression settings.
On average - comparing to the old saves using pio_cfr algorithm the size gains are:
- 12-17x for full saves
- 10-15x for small saves with no rivers
- 4-7x for micro saves with no turns
Technical appendix on new algorithm performance
Let's look deeper into details of new algorithm performance.
Hardware considerations
The new algorithm has different performance profile than the previous algorithms and will vary from machine to machine depending on the actual CPU and RAM.
The new algorithm also comes in two flavors - small and big. They are two implementations of the same algorithm, but one of them uses 2x more memory (v4_large), while the other does more CPU work (v4_small). V4_small is the default, but you will be prompted to check which one is actually faster on your machine.
- If you are RAM bound then v4_small will be faster.
- if you are CPU bound then v4_large will be faster.
In particular - if you have a CPU with a lot of cores (16+) and only dual channel RAM (instead of quad channel) then your PioSolver solving speed is going to be limited by RAM bandwidth more than CPU speed.
Case by case
It's hard to figure out when and how much will the V4 algorithm be faster to certain accuracies than pio_cfr. Our tests show general improvement in vast majority of trees and accuracies to make it a new default and hide algorithm selection in the GUI.
