-
Notifications
You must be signed in to change notification settings - Fork 530
TCEC
zz4032 edited this page Sep 29, 2020
·
82 revisions
TCEC (Top Chess Engine Championship) is a computer chess tournament organized and maintained by Chessdom at https://tcec-chess.com/. This page documents Leela configuration in each of the "seasons" of this tournament.
Division | Executable | Network | Placement | Result |
---|---|---|---|---|
SuFi | v0.26.3-rc1 | J92-190 | tbd. | tbd. |
DivP | v0.26.2-rc1 | J92-100 | 2/8 | 32.5/56 |
- Hardware: 4x V100 + Xeon 8163 CPU @ 2.50GHz, 32 vcores
- Non-default parameters:
- MoveOverheadMs=1000
- Backend=demux
- BackendOptions=backend=cudnn-fp16,custom_winograd=true,(gpu=0),(gpu=1),(gpu=2),(gpu=3)
- NNCacheSize=20000000
- MinibatchSize=320
- MaxPrefetch=160
- MaxCollisionEvents=917
- MaxOutOfOrderEvalsFactor=2.4
- SmartPruningMinimumBatches=600
- StrictUciTiming=true
- MovesLeftMaxEffect=0.2
- MovesLeftThreshold=0.862976
- MovesLeftSlope=0.0004
- Added/Changed for SuFi:
- BackendOptions=backend=cudnn-fp16,custom_winograd=true,res_block_fusing=true,(gpu=0),(gpu=1),(gpu=2),(gpu=3)
- MovesLeftThreshold=0.0
- MovesLeftSlope=0.004
- MovesLeftQuadraticFactor=0.0
- MovesLeftScaledFactor=1.0
Division | Executable | Network | Placement | Result |
---|---|---|---|---|
SuFi | v0.25.1 | svjio-t60-3972-mlh | 2/2 | 46.5/100 |
DivP | v0.25.1 | sv-t60-3010 | 2/8 | 24.5/42 |
- Hardware: 4x V100 + Xeon 8163 CPU @ 2.50GHz, 32 vcores
- Non-default parameters:
- MoveOverheadMs=1000
- Backend=demux
- BackendOptions=backend=cudnn-fp16,custom_winograd=true,(gpu=0),(gpu=1),(gpu=2),(gpu=3)
- NNCacheSize=20000000
- MinibatchSize=320
- MaxPrefetch=160
- MaxCollisionEvents=917
- MaxOutOfOrderEvalsFactor=2.4
- SmartPruningMinimumBatches=600
- StrictUciTiming=true
- Added for SuFi:
- MovesLeftMaxEffect=0.2
- MovesLeftThreshold=0.0
- MovesLeftSlope=0.007
- MovesLeftQuadraticFactor=0.0
- MovesLeftScaledFactor=1.0
- MovesLeftConstantFactor=0.0
Division | Executable | Network | Placement | Result |
---|---|---|---|---|
Div4 | lc0 "v0.16.1" (b1301aa9) | 10161 | 1st place | 22.0/28 |
Div3 | lc0 "v0.16.1" (b1301aa9) | 10520 | 3rd place | 16.0/28 |
- Hardware: TCEC GPU server has 2x GTX 1080 ti. Full specs here
- Command line:
lc0 -w 10161 -t 4 --minibatch-size=512 --backend=multiplexing "--backend-opts=(backend=cudnn,gpu=0),(backend=cudnn,gpu=1)" --nncache=2000000 -l lc0debug.txt
- UCI:
Move time overhead in milliseconds=2000
- Div3 added:
Time weight curve peak ply=37; Aggressive smart pruning threshold=0.604; Scale thinking time=3.05;
- Div3 added:
TCEC "v0.16.1" binary from appveyor
- Net IDs: kb3 and id481
- Hardware: TheAnswer volunteered 2x1080 ti via ssh access.
- Command line:
./lc0 -w [kb3|id481] --backend=multiplexing '--backend-opts=a(backend=cudnn,threads=2,gpu=0),b(backend=cudnn,threads=2,gpu=1)' -t 4
(kb1 and default respectively with increased move overhead to help with latency).
- Versions
- Div4: lczero v0.10, ID125 - +0 -25 =2 (and 1 win by forfeit)
- Hardware: TCEC CPU server 2 x 8 core Intel Xeon E5-2689 @ 3300 MHz
- Command line:
lczero -w id125 ....?