How to use Lc0 0.28.0 with GTX 1650 card?

[Jouni]

With large net bench is 1385 only. With 0.27.0 bench is 1970. So 30% slowdown! They have obviously not tested GTX at all? No help with disabling multigather. BTW small=default net has no slowdown. But here is no sense to use it ever.

[yurikvelo]

I use GTX1650. Tell which net exactly are you testing, I will test to confirm your observations

[Jouni]

Slowdown with all 30*384 nets. Test with 610199.

[yurikvelo]

GTX 1650 @ 1665 MHz
Windows 11, nVIdia driver 496.49
cudnn-fp16
position startpos

384x30-610199.pb.gz
go depth 12

v0.27.0: 2593 nps
v0.28.0: 2381 nps (91.8% of 0.27)

256x20-42850.pb.gz
go depth 26

v0.27.0: 7689 nps
v0.28.0: 7175 nps (93.3% of 0.27)

[Jouni]

Thanks. The actual slowdown is in my card with 34 bench position:

0.27.0
===========================
Total time (ms) : 346204
Nodes searched : 618935
Nodes/second : 1788

0.28.0
===========================
Total time (ms) : 350847
Nodes searched : 485836
Nodes/second : 1385 = 78%

No sense to use 0.28.0. In all ratinglists (RTX card) difference to 0.27.0 is 1-2 ELO only.

[Krzysztof Grzelak]

And what graphics card to buy today for testing LcO.

[yurikvelo]

And what graphics card to buy today for testing LcO.

SELECT TOP 1 FROM list WHERE price<=acceptedLimit AND vendor='nVIdia'
ORDER BY price DESC

[Krzysztof Grzelak]

Thank you for your answer yurikvelo.

[Ferdy]

I got this comparison.



Peaking at depth 9, what does it mean?

[Jouni]

Lc0 default is 2 CPU threads. But in my tests with GTX card 1 thread is faster always!