看了下，版本没差。
config只改了线程数，缺省6，现在的55是测试软件跑出来的。如果用缺省值，只有几百搜索/秒
改了这个之后空盘b60权重1xxx 搜索/秒，用b40权重是 2xxx 搜索/秒
threads调降到32，差别不大。
联想的刃9000k性能比兼容机差这么多吗
TRT_60B_202205> name
TRT_60B_202205> version
TRT_60B_202205> list_commands
TRT_60B_202205> komi 7.5
TRT_60B_202205> boardsize 19
2022-05-25 18:59:09+0800: GTP Engine starting...
2022-05-25 18:59:09+0800: KataGo v1.11.0
2022-05-25 18:59:09+0800: Using Chinese rules initially, unless GTP/GUI overrides this
2022-05-25 18:59:09+0800: Using 55 CPU thread(s) for search
2022-05-25 18:59:09+0800: nnRandSeed0 = 5038078417846332057
2022-05-25 18:59:09+0800: After dedups: nnModelFile0 = weights\kata1-b60c320-s5967310848-d2859394052.bin.gz useFP16 auto useNHWC auto
2022-05-25 18:59:09+0800: Initializing neural net buffer to be size 19 * 19 exactly
2022-05-25 18:59:12+0800: TensorRT backend thread 0: Found GPU NVIDIA GeForce RTX 3080 Ti memory 12884377600 compute capability major 8 minor 6
2022-05-25 18:59:12+0800: TensorRT backend thread 0: Initializing (may take a long time)
2022-05-25 18:59:13+0800: Creating new cache file: KataGoData/trtcache/kata1-b60c320-s5967310848-d2859394052_gpu_86_19x19-exact_batch56_fp16
2022-05-25 19:03:14+0800: TensorRT backend thread 0: Model version 10 useFP16 = true
2022-05-25 19:03:14+0800: TensorRT backend thread 0: Model name: kata1-b60c320-s5967310848-d2859394052
2022-05-25 19:03:14+0800: Loaded neural net with nnXLen 19 nnYLen 19
2022-05-25 19:03:14+0800: Initializing board with boardXSize 19 boardYSize 19
2022-05-25 19:03:14+0800: Loaded config katago_configs\default_gtp.cfg
2022-05-25 19:03:14+0800: Loaded model weights\kata1-b60c320-s5967310848-d2859394052.bin.gz
2022-05-25 19:03:14+0800: Model name: kata1-b60c320-s5967310848-d2859394052
2022-05-25 19:03:14+0800: GTP ready, beginning main protocol loop
= KataGoCustom
TRT_60B_202205> kata-get-param playoutDoublingAdvantage
TRT_60B_202205> kata-get-param analysisWideRootNoise
TRT_60B_202205> kata-get-rules
= 1.11.0
= protocol_version
【 在 zszqzzzf 的大作中提到: 】
: katago111
:
--
修改:AlexandriaI FROM 114.240.247.*
FROM 114.240.247.*