Checking the backends with the new lc0 binary

[corres]

...
All seems to be working fine with 0.26.3-rc2. And now getting much faster speed with cuda 11.1 with the big nets 384x30. Game average right now is 38.2 Knps on a 2080ti with default settings.

Really, as I also found earlier with the newer run-time dll (ver.11.0 and 11.1) and better cudnn dll plus software enhancement the Leela ver.0.26.3 (rc2) is faster about +50 %.
But starting from the results of Laskos the speed of only one RTX 2080 TI on a 384x30 (big) net may be
about 20-25 knps, and not ~38 knps as mwyoung stated.

[mwyoung]

...
All seems to be working fine with 0.26.3-rc2. And now getting much faster speed with cuda 11.1 with the big nets 384x30. Game average right now is 38.2 Knps on a 2080ti with default settings.

Really, as I also found earlier with the newer run-time dll (ver.11.0 and 11.1) and better cudnn dll plus software enhancement the Leela ver.0.26.3 (rc2) is faster about +50 %.
But starting from the results of Laskos the speed of only one RTX 2080 TI on a 384x30 (big) net may be
about 20-25 knps, and not ~38 knps as mwyoung stated.

They are not inconsistant moron. I gave a "game average" speed and clearly said "game average". This was reported by Banksia. Lasko gave a position speed. I have hit speeds over 50 Knps depending on the position

[MMarco]

...
All seems to be working fine with 0.26.3-rc2. And now getting much faster speed with cuda 11.1 with the big nets 384x30. Game average right now is 38.2 Knps on a 2080ti with default settings.

Really, as I also found earlier with the newer run-time dll (ver.11.0 and 11.1) and better cudnn dll plus software enhancement the Leela ver.0.26.3 (rc2) is faster about +50 %.
But starting from the results of Laskos the speed of only one RTX 2080 TI on a 384x30 (big) net may be
about 20-25 knps, and not ~38 knps as mwyoung stated.

I also get close to a 50% improvement on my mobile 2060: 6800 nps vs 4600 nps for J92-190. That is a tremendous speed-up!

[mwyoung]

...
All seems to be working fine with 0.26.3-rc2. And now getting much faster speed with cuda 11.1 with the big nets 384x30. Game average right now is 38.2 Knps on a 2080ti with default settings.

Really, as I also found earlier with the newer run-time dll (ver.11.0 and 11.1) and better cudnn dll plus software enhancement the Leela ver.0.26.3 (rc2) is faster about +50 %.
But starting from the results of Laskos the speed of only one RTX 2080 TI on a 384x30 (big) net may be
about 20-25 knps, and not ~38 knps as mwyoung stated.

I also get close to a 50% improvement on my mobile 2060: 6800 nps vs 4600 nps for J92-190. That is a tremendous speed-up!

Yes and that is the point. Here is a 2 game average speed of what I was quoting from testing. I just ran this in the last hour, for reference for this post.

Code: Select all

Result:
------------------------------------------------------------------------------
  #  name              games    wins   draws  losses   score    los%  elo+/-
  1. Lc0 v0.26.3-rc2       2       0       2       0     1.0-nan(ind)    -0.0
  2. Stockfish 051020      2       0       2       0     1.0-nan(ind)    -0.0

Cross table:
------------------------------------------------------------------------------
  #  name                 score   games         1         2
  1. Lc0 v0.26.3-rc2        1.0       2         x        ==
  2. Stockfish 051020       1.0       2        ==         x

Tech:
------------------------------------------------------------------------------

Tech (average nodes, depths, time/m per move, others per game), counted for computing moves only, ignored moves with zero nodes:
  #  name                nodes/m         NPS  depth/m   time/m    moves     time
  1. Lc0 v0.26.3-rc2        347K       46632     10.7      7.4     66.0    491.7
  2. Stockfish 051020    278840K    38136243     55.0      7.3     66.0    482.6
     all ---             136322K    18913286     32.9      7.4     66.0    487.1

Lc0 in 2 games had a average speed of 46.6 Knps. In 66 positions. This is for NN (J92-240). Very nice speed up.

[Laskos]

With the new CUDA backend, the speeds in the benchmark (34 positions each 10 seconds) with JH J92-190 net

Code: Select all

lc0_v263rc2_cuda.exe benchmark --backend=cuda-fp16 --minibatch-size=240 --weights=C:\Users\Kai\LC0_CUDA\J92-190

seem to be (some speculation involved)

3090: 32 knps --- $1600
3080: 25 knps --- $800
2080 Ti: 18 knps --- $700 second hand
2080: 13 knps --- $450 second hand
My 2070: 11 knps --- $350 second hand
2060: 8 knps --- $250 second hand

3070: ? 17-18 knps ? $600
A100: ? 50-60 knps? mucho denaros $$$

[corres]

...
All seems to be working fine with 0.26.3-rc2. And now getting much faster speed with cuda 11.1 with the big nets 384x30. Game average right now is 38.2 Knps on a 2080ti with default settings.

Really, as I also found earlier with the newer run-time dll (ver.11.0 and 11.1) and better cudnn dll plus software enhancement the Leela ver.0.26.3 (rc2) is faster about +50 %.
But starting from the results of Laskos the speed of only one RTX 2080 TI on a 384x30 (big) net may be
about 20-25 knps, and not ~38 knps as mwyoung stated.

They are not inconsistant moron. I gave a "game average" speed and clearly said "game average". This was reported by Banksia. Lasko gave a position speed. I have hit speeds over 50 Knps depending on the position

You are a real technical moron, who give a speed number without any explanation about the spring of the number.
I do not know "Banksia" but I suppose he has more connection to technical thing than you and he reported about the method how they got the 39.2 Knps result. So you have problem not only with the view of the technical issue but your memory also, moron.

[mwyoung]

...
All seems to be working fine with 0.26.3-rc2. And now getting much faster speed with cuda 11.1 with the big nets 384x30. Game average right now is 38.2 Knps on a 2080ti with default settings.

Really, as I also found earlier with the newer run-time dll (ver.11.0 and 11.1) and better cudnn dll plus software enhancement the Leela ver.0.26.3 (rc2) is faster about +50 %.
But starting from the results of Laskos the speed of only one RTX 2080 TI on a 384x30 (big) net may be
about 20-25 knps, and not ~38 knps as mwyoung stated.

They are not inconsistant moron. I gave a "game average" speed and clearly said "game average". This was reported by Banksia. Lasko gave a position speed. I have hit speeds over 50 Knps depending on the position

You are a real technical moron, who give a speed number without any explanation about the spring of the number.
I do not know "Banksia" but I suppose he has more connection to technical thing than you and he reported about the method how they got the 39.2 Knps result. So you have problem not only with the view of the technical issue but your memory also, moron.

Sorry you can not read or can you. You just want to play gotcha games. And like a poor marksman. You keep shooting yourself as usual.

[corres]

...
All seems to be working fine with 0.26.3-rc2. And now getting much faster speed with cuda 11.1 with the big nets 384x30. Game average right now is 38.2 Knps on a 2080ti with default settings.

Really, as I also found earlier with the newer run-time dll (ver.11.0 and 11.1) and better cudnn dll plus software enhancement the Leela ver.0.26.3 (rc2) is faster about +50 %.
But starting from the results of Laskos the speed of only one RTX 2080 TI on a 384x30 (big) net may be
about 20-25 knps, and not ~38 knps as mwyoung stated.

They are not inconsistant moron. I gave a "game average" speed and clearly said "game average". This was reported by Banksia. Lasko gave a position speed. I have hit speeds over 50 Knps depending on the position

You are a real technical moron, who give a speed number without any explanation about the spring of the number.
I do not know "Banksia" but I suppose he has more connection to technical thing than you and he reported about the method how they got the 39.2 Knps result. So you have problem not only with the view of the technical issue but your memory also, moron.

Sorry you can not read or can you. You just want to play gotcha games. And like a poor marksman. You keep shooting yourself as usual.

As a matter of fact, you are who "usually" write on this forum silly things. Obviously you have no any lilac steam about technical things you only imitate the technical knowledge what you got from others.

[mwyoung]

...
All seems to be working fine with 0.26.3-rc2. And now getting much faster speed with cuda 11.1 with the big nets 384x30. Game average right now is 38.2 Knps on a 2080ti with default settings.

Really, as I also found earlier with the newer run-time dll (ver.11.0 and 11.1) and better cudnn dll plus software enhancement the Leela ver.0.26.3 (rc2) is faster about +50 %.
But starting from the results of Laskos the speed of only one RTX 2080 TI on a 384x30 (big) net may be
about 20-25 knps, and not ~38 knps as mwyoung stated.

They are not inconsistant moron. I gave a "game average" speed and clearly said "game average". This was reported by Banksia. Lasko gave a position speed. I have hit speeds over 50 Knps depending on the position

You are a real technical moron, who give a speed number without any explanation about the spring of the number.
I do not know "Banksia" but I suppose he has more connection to technical thing than you and he reported about the method how they got the 39.2 Knps result. So you have problem not only with the view of the technical issue but your memory also, moron.

Sorry you can not read or can you. You just want to play gotcha games. And like a poor marksman. You keep shooting yourself as usual.

As a matter of fact, you are who "usually" write on this forum silly things. Obviously you have no any lilac steam about technical things you only imitate the technical knowledge what you got from others.

I know like predicting the results of TCEC. Because of my testing. I see no one saying how SF will crush Lc0 anymore.

Where are you guys.