I hope you don't mind me posting here for a little advice:
I'm going to by a new PC and I'm really hoping that someone can help me to understand to what degree computing power and graphics are essential in computer go.

I'm really interested in experimenting with Leela zero, AQ, and whatever incredible strength AI is soon to be available for modest home computer. What kind of performance can I expect with what kind of spec? is this likely to change? is there a minimum spec that will deliver good performance that you recommend i should hit?

As an example I'm looking at:
an ASUS, i7-7500U, 8GB RAM, Nvidia GTX940MX
(£690 amazon.co.uk/dp/B073WH686H)

Would this graphics card offer a significant increase in the speed and performance of go programs over just a good processor? do I need to go for something more in the area of a gaming laptop to get good quick AI performance?

If you are serious about Go AI go for a desktop pc.

Thanks very much for your input.

Unfortunately I'm very short on space so a desktop would be tricky. Would you say that I just can't get into decent strength go programs with a laptop set up?

Its all about the gpu cores/speed. A fast cpu will only do a few playouts per second. A modest gpu will be 10x the speed of the cpu doing the playouts.

You probably dont need an intel i7 processor, and i5 would do just fine.

This one is cheaper and may be comparable gpu
https://www.amazon.co.uk/dp/B0765CBZYN/

its got a gtx 1050 gpu which is a pretty decent gaming gpu.

myself i'd probably buy second hand but thats a pretty big discount and good price for the asus laptops at amazon.

What do you consider a "decent strength go program" ?

the gtx 1050 im guessing would give something like 10000 playouts in 10 seconds.

i dont have one but its somewhere in that ball park.

For the current leela zero this would be at pro level. mybe not very high pro. i'm really not sure.

Thanks Tryss, that's a valid point - I'm not really qualifying what I'm looking for very clearly!

I suppose I'd really like to be able to do things like:
-review games at at a fast pace, but have some confidence that the Ai is really recommending well read out, high quality variations
-play faster games against a very strong Ai. I'm only 4k European, so pro strength is not necessary at all, but I'd like to be able to give the bot less thinking time and still be playing a strong Ai
-maybe have a play with some basic scripts and the APIs for these bots - see if I can do some automated game analysis

Thanks nasdaq.

You're insights on gtx 1050 vs i7 processor was exactly the kind of tip I was hoping for.

I will probably go for something with the 1050 as there seem to be a few options. The one you pointed out does look like a good pick.

Hi Mike,

if you dont know already, Go Review partner is good for automated reviewing.

homepage:
http://ns1297.imingo.net/

sabaki also has a leela zero addon to add variations to the sgf file.

Number of visits is not a good indicator for strength. Bigger networks can be stronger at a lower number of visits.

Be careful with the notebook CPUs. I use an I7 notebook and love it for the things it has to do. But performance wise it is no match by far for my 7 year old I5 2500k desktop CPU. It is nice for internet, word processing and viewing some pdf files. I can run Zenith 7 (CPU only AI) on the notebook, but I have to be very very patient (I use it sometimes to analyse games with friends, but it is no fun for everyday AI supported reviewing)

Reviewing in realtime with Lizzie and a 1080Ti class graphics card has to be experienced (You can move through the game and view your mistakes and best alternatives in an instant no waiting at all, not 40 sec per move). Only Zenith offers a similar good user experience at a little bit lower strength. It is a joy to review your games and only takes a few minutes to do so on a capable GPU. (Notebook variants are significantly slower due to temperature throtteling). If you want to improve fast, reviewing every game with Lizzie is a very good way to do so.

Thanks for the tips on tools. I'd heard some of these mentioned around the forum, but never had the spec to make them worth using.

Thanks Marcel, and also thanks for your post about the Zhao Baolong game - very interesting.

Looks like I'll have to compromise a little compered to a good desktop set up, but it looks like a fairly good result still.

I thought I'd tag onto this post as I'm also thinking about buying a new desktop PC, mainly for running Leela Zero etc. I've been out of the PC hardware game for ages (last bought a Pentium 4 in 2004!), so am acquainting myself with the current lingo. My use case is running LeelaZero, maybe AQ, PhoenixGo, Elf if possible, and maybe maybe doing some neural network development of my own. I don't care about playing PC games, so if a regular desktop without a gaming super duper CPU but with a fat GPU is the way to go that's fine.

- Processor:
1) Intel or AMD? (If Intel probably want i5+?)
2) Any compatibility things to think about?
3) Does it really matter much or is it just GPU power that's important?

- GPU:
4) Nvidia or AMD or is there anything else?
5) any compatibility issues? I think LZ works with both, but didn't Facebook's ELF only work on Nvidia?
6) If I were to play around with neural networks are there any commonly used frameworks which means I should use one or other?
7) Is a general GPU benchmark like "G3D Mark" here a good metric to judge go bot performance or is there some more specific benchmark I should look at which basically scales linearly with the answer to "How many playouts will a 15-block LZ network give me in 1 second?"
7) Could anyone please give some sample numbers e.g. I have a GTX 1050 and get 20k in ~10 seconds.

Other things like RAM don't seem so important and not much difference between boxes, though I might get an SSD for funzies.

Thanks

The important thing here is whether the applications use OpenCL or CUDA to run code on the GPU. OpenCL is an open standard that is supported by both AMD and Nvidia, whereas CUDA is Nvidia's own framework that only they support. Unfortunately CUDA is really the dominant option, supported strongly by Nvidia, which is why quite a lot of things are Nvidia-only. Also, although Nvidia's drivers support OpenCL, I think I've heard they tend to be some way behind the latest versions.

Leela Zero uses OpenCL, so it should work anywhere (but I don't know if it's more optimised for some devices).

Edit: Of course, this doesn't help much with actually choosing hardware. I'm quite interested in the answer myself, for exactly the same reasons!

I have a GTX 1080 and get about 1k LZ visits (192x15 network) a second.


Especially these days when programs are memory hogs, you will notice it if you don't have enough RAM. I wouldn't skimp on it. Also, having an SSD makes a real qualitative difference, and now that I have one I really wouldn't want to go back.

ELF is a nightmare to get up and running, however the weights were converted to the Leela Zero format, and they run very well on the Leela Zero engine. (This combination is by far the strongest thing you can easily get working. It is outrageously strong.)

If you're interested in machine learning these days, you really want a nvidia GPU. So, as you say you are possibly interested to try fiddle with neural network & deep learning, I suggest this option.

Too bad AMD seems to have missed the Deep Learning rush.

Thanks, my work laptop (i3 3120M, couldn't get the GPU version to work and probably not so great) gets ~10 a second.


My initial browsing showed both a bog standard office PC and low-mid gaming came with 8GB of DDR4 RAM so thought that'd do. Or is the "8GB DDR4" hiding that one is faster or the higher end motherboard will make it faster?

8GB should be fine; if you had said 4GB I would have tried to talk you up. My laptop has 16GB and I am glad that it does, but I also do things like run a virtual Windows machine inside macOS and run chess engines that use an arbitrary amount of RAM.

I don't think you have to worry much about exactly how fast the RAM is, just that you are using it instead of swapping to disk.

To feedback my experience

I couldn't really get a desktop setup as I'm particularly short on space. I got a laptop - knowing, thanks to this thread that it would be a significant compromise compared to a desktop set up. I found a HP gaming machine (not that I play any other games) with an Nvidia 1050 card, an i7, 8gb, hdd(no funzies).

I find I can get around 150-300 play outs a second when running the ELF network in lizzie (this seems to vary hugely at different times.) I find that for most purposes I can review quite happily and Elf won't change it's mind much as it adds more reading. If I really want it to take a closer look at a position, I can be patient and wait for the extra play-outs.

Thanks very much to the contributors on this thread. I felt I made an informed decision and I'm really quite happy with the results.

So the difference between a GeForce GTX 1050 and 1080 is quite a lot, more than (1080-1050)/1050, 10 is the series and 50 or 80 the model number . So for my and others benefit, here are some benchmarks of various Nvidia graphics cards (from https://www.videocardbenchmark.net/gpu_list.php) which I am assuming for now is a decent proxy of LeelaZero performance.

So the 1060 looks around the sweet spot where the marginal performance / cost starts to go down. But do I wait a bit for the bitcoin fad to fade and prices drop more... Also as LZ keeps improving she'll get better with fewer visits.

P.S. Am I right in thinking it's GPU speed rather than video RAM that's most important for Go AIs, so a 3GB 1060 isn't much worse than the 6GB version?