![](/Portals/all/thumbs/102/102201.jpeg)
ChessBase 17 - Mega package - Edition 2024
It is the program of choice for anyone who loves the game and wants to know more about it. Start your personal success story with ChessBase and enjoy the game even more.
The classic Fritz 18 engine has a manually optimised rating function and achieves a playing strength of about 3300 ELO. In the new Fritz 18 Neuronal engine, the rating function has been replaced by a fast neuronal network (NNUE), enabling play 120 ELO points better, (i.e., achieving about 3,400+ ELO). NNUE stands for "Efficiently updatable neural network", the abbreviation being read from right to left because the idea was first published in Japanese in connection with shogi.
Just try out the new engine and compare the results with your standard engine. Open a game and run the standard engine (in the example Fritz 18). Then go to the Engine menu and here to Add Kibitzer:
In the selection list, click on Fritz 18 Neuronal. The second engine appears on the screen and looks something like this:
In the example, the previous standard engine, Fritz 18, is on the left. The new Fritz 18 neuronal engine is in the right-hand window. Both calculate in parallel and the same time at the same position. Just have a look and you will see that engines can be of different opinion.
By the way: If you have installed the new engine, the same possibility of analysing with two engines running in parallel is also available in ChessBase.
The special feature of NNUEs is that they are relatively small, and can be efficiently evaluated incrementally during the search. NNUEs do not require a GPU - the instruction set extensions of modern CPUs (SSE2, AVX2 or AVX-512) are sufficient. The fastest version of Fritz 18 Neuronal, which uses the AVX-512 instruction set, achieves approximately the same number of nodes (NPS) as Fritz 18; the AVX2 version (all modern CPUs) is only 5% slower.
A total of approx. 5 million games were played, and from these, 400 million rated positions were generated for training the net. The first net was trained based on evaluations from 200,000 self-play games of Fritz18, then in a further 80 training iterations, a total of about 300 intermediate versions of the net were tested and the best net was chosen step by step.
Advertising |