Life In 19x19

Golaxy vs KataGo 4

http://eidogo.com/#t3TBt434

The Nvidia GeForce RTX 4090 is the most powerful GPU available, faster than A100.

B: 8 4090 KataGo s719 network

W: Commercial Golaxy 80x 50b

Rules: Japanese Go rules
Komi: 6.5
W+R

Visits: Commercial Golaxy 80x : 20000 v/s
8 4090 KataGo: About 100000~530000~540000 v/s

https://discord.gg/92zXjuna9D
Join Golaxy's discord for any further discussions.I will not reply to any messages here.

0714952377 wrote:The Nvidia GeForce RTX 4090 is the most powerful GPU available, faster than A100.

B: 8 4090 KataGo s719 network [...]

8 4090 KataGo: 530000~540000 v/s

This means 1 4090 gives 66875 v/s.

However, such a number is unrealistic. "Up to" makes more sense. Even then, such a number has little meaning. For comparison, my 4070 gives 3400 v/s ~ 353,000 v/s depending on the position and current exploration; 3400 v/s without reversion (creation of the same situation by different sequences), up to 15000 with little reversion and rarely 353,000 v/s with the heaviest reversion. An RTX 4090 has about 2.1~2.2x the speed of a 4070 so I expect 7310 v/s ~ 758,950 v/s. So 8 4090 might give 58,480 v/s ~ 6,071,600 v/s if installation of Nvidia libraries and tuning are rather optimal. Divide by 4.89 for the worst TensorRT installation and tuning. See https://home.snafu.de/jasiek/AI_Computer.html

Congratulations on having access to 8 4090 though!:) Which mainboard and cooling does the computer use, please? In Germany, I would pay roughly €3,600 / year for electricity for using such a system for several hours per day... The computer must have cost €22,000 (do it yourself) ~ €29,500 (prebuilt), I guess.

Thank you for sharing games and some speed numbers! However, the more relevant estimated numbers are the number of visits of the top candidate move per turn and the number of visits of the second candidate move per turn. What are they roughly?

EDIT: According to Benjamin Teuber, A100 in an online service gives 15,000 v/s for €3/h but I do not know details. If this should be the lower bound, A100 would be significantly faster than 4090.