Life In 19x19

Anyone can provide a reference desktop computer specs for go AI review purpose? My computer vendor asks me to provide one based on which we can decide the specs, analysis speed, and cost.

I have a notebook with Intel i5 CPU and Windows 10 64 bit OS, and a note 8 phone. I want to run go AI for reviewing games and find bad moves and good moves, in terms of winrate delta or score delta. Typical use cases are:

Ah-Q app on my phone that connects to LZ engine (with different weights) my notebook via SSH, but this is limited to LZ but not Katago, since the latter requires GPU.

Run go review partner on my notebook, but this way I have to set with my notebook.

The analysis speed of LZ264 (40b) is about 3.5 visits / secs., if I want to analyze a game with 250 moves with 10k visits / move, it takes about 177 minutes, if 1k, 17 minutes. I wonder if I have a GPU desktop, I can run it even faster.

Anyone any advice? (Computer specs, and analysis speed) Thanks in advance.

Use Zbaduk for now.
All other advice is related to your budget. How much do you want to spend. Desktop GPUS are much faster.

dino1019 wrote:The analysis speed of LZ264 (40b) is about 3.5 visits / secs., if I want to analyze a game with 250 moves with 10k visits / move, it takes about 177 minutes, if 1k, 17 minutes. I wonder if I have a GPU desktop, I can run it even faster.

One thing you might do is to run the analysis for 15 minutes. Then, based on the analysis, play moves that you are interested in understanding better to get LZ's opinion at 10k. It is quite possible that at 10k for a particular move, LZ will change its mind about a move you made or about the move it would choose.

How does Apple A13 compare to Desktop GPUs?

dino1019 wrote:Ah-Q app on my phone that connects to LZ engine (with different weights) my notebook via SSH, but this is limited to LZ but not Katago, since the latter requires GPU.

Katago can run on the Intel Integrated Graphics, such as the Intel Iris Graphics 6100 found in my MacBook Pro (early 2015 model). I've read something about Katago having issues with it, but I haven't run into this.

Gomoto wrote:Use Zbaduk for now.

I agree with this. Forum member spook is doing an amazing job with ZBaduk. It works very well, it's easy (no need to install anything) and relatively cheap!

Yes, zbaduk is a potentially useful go review tool, but it's not fully customized for mobile as ogs does yet, also I cannot see list of bad moves as AI sensei does. ogs only shows the first 3 bad moves before paying, and AI sensei provides free analysis for 3 games only. If there's a fast, mobile and low cost way of knowing bad moves, then I can use Ah-Q + LZ@PC remotely for live analysis on bad moves.

You can also consider setting a cloud instance and running KataGo/LZ on there. If you don't play computer games very much, it can be a reasonable choice, even in terms of budget.

dino1019 wrote:The analysis speed of LZ264 (40b) is about 3.5 visits / secs., if I want to analyze a game with 250 moves with 10k visits / move, it takes about 177 minutes, if 1k, 17 minutes. I wonder if I have a GPU desktop, I can run it even faster.

Anyone any advice? (Computer specs, and analysis speed) Thanks in advance.

I don't know how much you can/want to invest. On my GTX 1660 Ti (around 300€), I get ~300-350 visits per second on LZ265 using Lizzie.

dino1019 wrote: The analysis speed of LZ264 (40b) is about 3.5 visits / secs., if I want to analyze a game with 250 moves with 10k visits / move, it takes about 177 minutes, if 1k, 17 minutes.

How did you make that calculation?

I got 250*10000/3.5/60 = 11905 minutes = 198 hours.

If your computer is too slow, use 15-block networks instead, they are strong enough.

dino1019 wrote:The analysis speed of LZ264 (40b) is about 3.5 visits / secs., if I want to analyze a game with 250 moves with 10k visits / move, it takes about 177 minutes, if 1k, 17 minutes. I wonder if I have a GPU desktop, I can run it even faster.

Anyone any advice? (Computer specs, and analysis speed) Thanks in advance.

3.5 visits/sec, are you sure you don't mean 35? My Ryzen 5 (without using GPU) gives me 25-30 visits/sec for a 40b network.

Any mid price GPU will take you up to 200-300 visits/sec, no need to fuss too much over the specs. I get a little over 200 on my GTX 1070.

Attached is screenshot of Ah-Q app with remote connection to Windows 10 (Intel i5) and run LZ264 (40b). You see the 500 visits are received in 2 or 2.5 min., so the speed is 3.5 to 5 vps. (the default threads is 4)

You understand that the win rate estimates based upon fewer than 100 visits are basically meaningless. In fact, the Elf commentaries on pro games do not report any winrate estimates based on fewer than 500 visits.

For a kyu player, 100 visits is fine if you use Katago (otherwise LeelaZero may occasionally misread a ladder). Not that the bot will never make mistakes, but it will not make more mistakes than a high-dan player anyway.

P.S. My notebook computer doesn't have a GPU, and I get about 50 visits/second with 15-block Katago.

xela wrote:

dino1019 wrote:The analysis speed of LZ264 (40b) is about 3.5 visits / secs., if I want to analyze a game with 250 moves with 10k visits / move, it takes about 177 minutes, if 1k, 17 minutes. I wonder if I have a GPU desktop, I can run it even faster.

Anyone any advice? (Computer specs, and analysis speed) Thanks in advance.

3.5 visits/sec, are you sure you don't mean 35? My Ryzen 5 (without using GPU) gives me 25-30 visits/sec for a 40b network.

Any mid price GPU will take you up to 200-300 visits/sec, no need to fuss too much over the specs. I get a little over 200 on my GTX 1070.

The Ryzen 5 has vega graphics, much faster than intel hd graphics. Ryzen 5 desktop might have vega 11, which is faster than a gt 1030 for opencl! https://browser.geekbench.com/opencl-benchmarks

jlt wrote:For a kyu player, 100 visits is fine if you use Katago (otherwise LeelaZero may occasionally misread a ladder). Not that the bot will never make mistakes, but it will not make more mistakes than a high-dan player anyway.

Incorrect winrate estimates do not mean that the order of correct winrates would be any different. But it's not unusual that the order will change. That's why I recommend doing a quick first run and then comparing interesting options by making those plays. Doing so will produce high enough visits to be reasonably confident about the comparisons.

In the example given, for instance, the keima might be interesting, even though it has a low winrate estimate and only 16 visits. Making that play could easily change the winrate estimate by 10% or more, with 500 visits, or 1,000 visits, or 10,000 visits. It's not like you do that for every option, just for the ones that look interesting after the first run.