Faqv4 bazel master : add benchmark Tesla V100 + minor readme changes

[wonderingabout]

@wodesuck

general look here : https://github.com/wonderingabout/PhoenixGo/tree/faqv4-bazel-master

2 main changes in this PR :

1) add benchmark Tesla V100 :

this benchmark showed a few interesting conclusions :

• batch size 16 is a very efficient batch size on Tesla V100, providing an average 900 simulations per move, and a 130% speed boost as compared to batch size 4
• As of February 2019, PhoenixGo does not benefit from CPU thread number higher than 2 cores / 4 threads on Tesla V100

These numbers may change if PhoenixGo supports newer Tensorflow versions, as well as newer tensorRT versions too of course (latest is tensorRT 5.0 which brings native support for RTX and tesla V100)

latest stable version is tensorflow 1.13 close to final release : https://github.com/tensorflow/tensorflow/releases

tensorflow 1.13 changelog says this in : https://github.com/tensorflow/tensorflow/releases/tag/v1.13.0-rc2

TensorFlow GPU binaries are now built against CUDA 10 and TensorRT 5.0.

so i think it would be very great to support tensorflow 1.13 with PhoenixGo

benchmark gtx1060 75W has been slightly updated as well

2) add "undo" on readme + add FAQ question + other minor readme changes

other minor readme updates (i checked on list_commands and undo appears at the bottom of the list now) another new FAQ question A8.

conclusion : after merging this, i think it PhoenixGo may want to look at tensorflow support of newer versions like 1.12 or 1.3, as well as tensorRT 5.0 (to add support for RTX cards, and native support for Volta (no need to build model anymore))

[tencent-adm]

All committers have signed the CLA.

[wonderingabout]

@wodesuck

i will do again all the benchmarks when PhoenixGo supports tensorflow 1.13, this will be a major speed improvement, as well as native support for nvidia RTX and Volta

i will also update the bazel and readme documentation, and i am willing to help for the testing if needed

[wonderingabout]

@zxkyjimmy

you release the PhoenixGo update for tensorflow 1.8 in https://github.com/Tencent/PhoenixGo/commit/6b17598bd4b249d78112f76f5d35272b64d31605

can we do something for tensorflow 1.13 too, for support of cuda 10.0 and tensorrt 5.0 in RTX cards and Volta ? big thanks

see tensorflow 1.13 changelog : https://github.com/tensorflow/tensorflow/releases/tag/v1.13.0-rc2

TensorFlow GPU binaries are now built against CUDA 10 and TensorRT 5.0.

[zxkyjimmy]

It's my pleasure. But I have a lot of deadlines to handle. I will start this work next week.

[wonderingabout]

@zxkyjimmy

it is also my pleasure :+1:

whenever you can, the potential speed improvements that we can expect from tensorRT computation are arround 40% or more, as compared to notensorrt on Volta and RTX

i will do all the testing when your work is ready, big thanks :+1:

[wonderingabout]

today again, phoenixgo proved to play very smart and complex game, very satisfying :+1: let's keep improve phoenixgo further ! :+1: https://online-go.com/game/16674632

[alreadydone]

Yeah surely phoenixgo needs improvement as it's losing 14 games in a row to LZ on Fox! (LZ has 4xV100; not sure about phoenixgo but feel free to improve it with hardware!)

[wonderingabout]

@alreadydone

4 V100 for a 40b, vs we dont know what, if i remember correctly it's a Tesla P40, so it should not give more than 0.5 V100, right @wodesuck ?

of course the comparison would be unfair then :+1:

but it is always interesting to have diversity, we gain knowledge from that

[wonderingabout]

@zxkyjimmy small update just to notify you that tensorflow 1.13.1 final version just got released yesterday as @l1t1 told me https://github.com/tensorflow/tensorflow/releases

i know many things can take time, so whenever you can for PhoenixGo i will be always happy :+1:

[zxkyjimmy]

I see. (((o(ﾟ▽ﾟ)o)))

[wonderingabout]

@zxkyjimmy

✺◟(∗❛ᴗ❛ั∗)◞✺

[wonderingabout]

big thanks @wodesuck :+1: