Hello, I was wondering about the GPU setup of the NGD teachers and how they use AI.
I remember Mikkgo saying in a video from last year that he has a connection to Antti's graphics card.
So do you all have laptops or desktops with powerful GPUs, or do you use ssh to connect to one especially good GPU? Or maybe you use a cloud GPU on AWS or Google?
What do you use when you're on the road? (Not so much now, I suppose, but this kind of expensive hardware is a long-term decision.)
On a day trip during the EGD in Brussels last year Antti used an iPad Air from 2019 to analyze a game; using the app "A Master of Go" with a Leela Zero network, it got quite a lot of playouts; about ten times as many as on my iPad from 2018. The app is very nice but given a choice, I prefer Lizzie or Sabaki.
I'm asking because I'm not entirely happy with my current setup and need to make some decisions.
Mikkgo having a connection to my GPU is news to me! Did he somehow hack my computer without my knowing? 😃 More seriously, I assume either Mikkgo misspoke in the video or you misunderstood him?
I used to have a Google Cloud account set up as per Benjamin Teuber’s guide (https://github.com/bsteuber/lizzie-gcloud-setup), but stopped using that already more than a year ago. The reason to this was that my new desktop computer and iPad Air provided more than good enough an alternative. The Google Cloud setting had an annoying lag when using it, and it was also unnecessarily difficult to update the network used.
The iPad Air is my go-to device whenever I am on the move. Apple’s A12 chips and newer include a neural engine, which for example the A Master of Go app can use to power a considerably strong go AI. Furthermore, A Master of Go nowadays includes KataGo networks, which I think is indispensable for studying – I have found that simply looking at the winrate is not good enough for me.
On my desktop computer I am running the 40-block KataGo network with Lizzie on a Vega 56 GPU – not exactly monster hardware anymore by today’s standards, but certainly strong enough for go study. As far as I know, a GeForce 1060 is already strong enough for go AI use if you want to save on hardware costs. When you look at the price of the whole package, however, a new iPad Air or iPad mini (which also has an A12) is definitely the most cost-efficient option.
Have you checked out ZBaduk (https://www.zbaduk.com/ )? These days you may not need a dedicated device at all in order to study with an AI.
It's possible I misheard it; it's at https://youtu.be/L0V67Nt-aUY?t=131 .
At the moment I use a 2019 iMac with a Radeon Pro 580X, which I believe is a bit better than a GTX 1060. I run KataGo locally and can also ssh to it from my ancient MacBook Air.
I know of ZBaduk and consider it as a backup solurion for when the ssh connection doesn't work.
I also tried Benjamin's Google Cloud setup but at some point it got too annoying with constant disconnections (because of the preemptible GPUs I suppose); I also had to always make sure that the instance was really stopped after using it.
When Apple releases the new generation of MacBooks with Apple Silicon processors that include a Neural Engine starting next year, it will be interesting to see whether KataGo runs on that. I'm pretty sure that "A Master of Go" can run on it (if the developer releases a macOS version), but even better would be a CoreML backend for KataGo so it's possible to use it with Lizzie or Sabaki.
Huh – checking the video, I can see how you came to think that!
I think Mikkgo meant that he and Jeff have been able to double-check the validity of some sequences and joseki by asking me, who will in turn ask my Vega 56. This nuance is of course not very clear in the video.
I seem to recall reading that the new Apple ARM notebooks could run iPad programs out of the box (I couldn’t verify this with a quick googling). If this (hopefully) is the case, then it would be possible to run A Master of Go out of the box without even having to depend on the developer to provide a Mac version.
I got a laptop with a GTX 1050 TI which I thought would give me enough playouts for my purposes. In general I'm pretty happy with the results in terms of the strength of the AI, and the added benefit is that such a laptop can be taken to tournaments relatively comfortably.
These days... an iPad with an A12 chip is a good solution, as Antti pointed out. I would perhaps have invested less in a computer and gotten an iPad air or mini with an A12 had the option been available in 2018.
All times are in Helsinki time (eet with summer time).
Check the current review credits balance
Public lectures on Twitch every 2nd and 4th Saturday of the month at 1 pm.
Jeff and Mikko stream on Twitch on Fridays at 6 pm.