Life In 19x19

Thanks very much, Marcel. Hope I don't get the weird error message this time.

Bill Spight wrote:I have a 2014 MacBook Pro, and this year is when I would usually get a new computer. However, looking around the web for what is available, it seems like a lot of things are changing about hardware to run neural nets, so it may be best to wait a while. One thing I am thinking of is getting an NVIDIA RTX 2080 Ti, which seems to be the thing to get these days, and using it as an external GPU. The Macbook has Thunderbolt, but not Thunderbolt 3. Apparently that can be handled with an adapter. If I go that route I have two concerns. First, I have only 16 GB of RAM, which might not be enough for good performance. I have seen 64 GB recommended. Second, one possible problem may be overheating the MacBook. OC, the external GPU will have its own cooling. But the one guy I found using an external GPU with a 2014 MacBook Pro had a silent video about his setup, which showed a temperature monitoring program on his machine.

I welcome any ideas and suggestions. Many thanks.

I tried using 16 GB with KataGo, but with normal-ish settings you never hit even close to that. I bought a new computer and it's 16GB for this reason.

I'm going to use an RTX 2060 for KataGo and Leela Zero. That should be enough, about half of the performance of 2080Ti, but for only $300.

FWIW, I built a new system completely for it instead of going the eGPU route. It only cost $900 total in a 5.5L case (okay, I still got a wireless keyboard and mouse, putting it to $1000, but that's going to be useful anyway). I'm still waiting for components, but my whole computer is the size of an eGPU enclosure and can run independently (I plan to SSH into it to run KataGo). I can connect it to the TV to install the OS.

With this solution it won't interfere with other programs (RAM-wise, CPU-wise, etc.), since KataGo uses the CPU heavily as well. I got an AMD 3600 for its ability to use 12 threads (should be good enough to feed even a V100) and a B450 motherboard since I don't need extra features from the newer motherboards (I hope it's flashed with a new BIOS, lol)

You can build a bigger one as well for more, since you probably will need better air flow in your case for beefier components

Bill Spight wrote:I have a 2014 MacBook Pro, and this year is when I would usually get a new computer. However, looking around the web for what is available, it seems like a lot of things are changing about hardware to run neural nets, so it may be best to wait a while. One thing I am thinking of is getting an NVIDIA RTX 2080 Ti, which seems to be the thing to get these days

iopq wrote:I'm going to use an RTX 2060 for KataGo and Leela Zero. That should be enough, about half of the performance of 2080Ti, but for only $300.

FWIW, I built a new system completely for it instead of going the eGPU route. It only cost $900 total in a 5.5L case (okay, I still got a wireless keyboard and mouse, putting it to $1000, but that's going to be useful anyway). I'm still waiting for components, but my whole computer is the size of an eGPU enclosure and can run independently (I plan to SSH into it to run KataGo). I can connect it to the TV to install the OS.

Many thanks, iopq! Very helpful.

One recommendation I have seen for an independent computer is to have 4 GPU slots, even if you only install one GPU, for later expansion. But instead of getting the pricey 2080 Ti, maybe three 2060 GPUs (plus fans) would provide better performance for comparable cost. What do you think? Thanks.

It would work, but it may not scale perfectly on all programs. I personally haven't tried it, but there's been some reports on LZ issue tracker:

https://github.com/leela-zero/leela-zero/issues/2525

YMMV, it may still be better

@ iopq

Many thanks.