Life In 19x19

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.

dfan wrote:

Uberdude wrote: 7) Could anyone please give some sample numbers e.g. I have a GTX 1050 and get 20k in ~10 seconds.

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

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

dfan wrote:

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

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.

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?

Uberdude wrote:

dfan wrote: 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.

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?

I got this system with a GeForce 1060 and am happy with it: get about 500 playouts per second with the 15-block LeelaZero.

https://www.scan.co.uk/products/scan-ga ... 1tb-win-10