Fat Fritz Update

[Albert Silver]

Seems like it supports Cuda to me:

Fat Fritz (in lc0)
Detected 6 core(s) and 12 thread(s) in 1 group(s).
Group 0 has 6 core(s) and 12 thread(s).
isready
readyok
ucinewgame
Found pb network file: C:\Users\Jason\AppData\Roaming\ChessBase\Engines.uci\Fat Fritz/FatFritzCpu.weights
Creating backend [cuda-auto]...
Switching to [cuda-fp16]...
CUDA Runtime version: 11.1.0
Latest version of CUDA supported by the driver: 11.2.0
GPU: GeForce RTX 2060
GPU memory: 6 Gb
GPU clock frequency: 1680 MHz
GPU compute capability: 7.5

Strange, I used the Fat Fritz 20x256 net and on my RTX 2060, Fat Fritz 1.2 did not run with new CUDA 11 libraries. But without releasing a new FatFritz-net, the Fat Fritz 1.2 binary alone is complete useless - so who cares? The old Fat Fritz net can be used with Lc0 0.26.3 binary...

It should come with the CUDA 11 libraries itself. The update was only because there were complaints it would not work with the new 30xx cards. The Lc0 it uses is faster, but it was not the reason for the update. I stopped development of the NN itself back in February.

[Nordlandia]

Albert Silver: is there any plans to resume NN improvement at some point later on ?

[corres]

...
I stopped development of the NN itself back in February.

So the new(.2)net = the old(.1) one?
The -70 Elo difference is not a motive for development?

[carldaman]

So the new(.2)net = the old(.1) one?
The -70 Elo difference is not a motive for development?

Why continue development, just to catch up to freeware? It's only a commercial product...

[Albert Silver]

...
I stopped development of the NN itself back in February.

So the new(.2)net = the old(.1) one?
The -70 Elo difference is not a motive for development?

Sure, but I have other projects in the works.

[AndrewGrant]

Boggles my mind why the dedicated chess fans on these forums care at all about Fat Fritz.
Go just get the latest Leela version, or try someone's fork / spinoff for free if you want some flavour.

[Collingwood]

Boggles my mind why the dedicated chess fans on these forums care at all about Fat Fritz.
Go just get the latest Leela version, or try someone's fork / spinoff for free if you want some flavour.

Yes, it seems a little irrational.

[Collingwood]

The -70 Elo difference is not a motive for development?

Why continue development, just to catch up to freeware? It's only a commercial product...

And it may be very hard to catch up anyway. It may have hit its ceiling.

[corres]

The -70 Elo difference is not a motive for development?

Why continue development, just to catch up to freeware? It's only a commercial product...

And it may be very hard to catch up anyway. It may have hit its ceiling.

Maybe Fat Fritz reach its ceiling.
But there are newer method to make stronger net and only the FatFritz was made using (mainly) human games.
So we can hope the style of FatFritz is nearer to human style than of Leela.
This is the cause why I interested in FatFritz.

[Stephen Ham]

Hi Guys,

I too value Fat Fritz. In addition to its interesting and aggressive style which is useful for analysis, I like its evaluation function. While I'm critical of all engine evaluations, FF is superior to SF NNUE's tiny net evaluations. Given sufficient time, I think its evaluations are the best I've seen from an engine.

I also think FF hasn't been given a fair chance to demonstrate its performance. I want to see it in TCEC...although I know that won't happen, which is a shame. My perception is that it's roughly equal in strength/performance to SF NNUE...until reaching the endgame. Then the performance drops off sharply. However, I'll bet that Albert can fix that if given a chance to play in TCEC.

-Steve-