Life In 19x19

Well im having the same problem as you. The release didnt happen yet, right?

Fellow Go fan and chagrined new 5700 XT owner here. I knew when I bought it that AMD wasn't exactly the leader in deep-learning support, but I didn't realize that this specific model was so broken. I am joining the forum and thread to keep posted about any developments around this so I can hopefully play KataGo and Leela Zero one of these days. Thanks for all updates.

I just found this thread and I’m now nervous. I bought a new iMac with Radeon Pro 5700 xt. It’ll be delivered in a couple of weeks.

One of the reasons was I figured in the likely event that Apple fully dropped support for Open CL in its next OS (already deprecated), I could still use boot camp to run katago in Windows.

Is there any news on this front lately? Have recent updates worked? Is it just a windows problem?

Well, at least Apple has a generous return policy if it fails.

Thanks.

Hi, let me just say that I actually managed to get Katago running on Win 10 with 5700 XT seemingly without a hitch or any special steps. Runs fast, plays strong. (Note that I am not doing any neural-net training so far, just using the existing weights.)

dojo_b wrote:Hi, let me just say that I actually managed to get Katago running on Win 10 with 5700 XT seemingly without a hitch or any special steps. Runs fast, plays strong. (Note that I am not doing any neural-net training so far, just using the existing weights.)

AMD has patched the bug on Windows with latest drivers. Allegedly also for Ubuntu 20.X, but I haven't been able to install the amdgpu-pro driver package successfully as of yet.

Thanks both. I feel better now!

dojo_b wrote:Hi, let me just say that I actually managed to get Katago running on Win 10 with 5700 XT seemingly without a hitch or any special steps. Runs fast, plays strong. (Note that I am not doing any neural-net training so far, just using the existing weights.)

Runs fast is great news! How many playouts per second are you getting with the main 40b engine? Let us know which processor and how much RAM you have too, if you don’t mind sharing.

I hope the new Radeon Pro 5700 XT works well too.

I'll try to give playout stats if you can explain how (I couldn't find it).

AMD Radeon 5700 XT gpu
AMD Ryzen 7 3700X 8-core cpu
32 GB DDR4 SDRAM
AORUS B450 Pro WiFi mobo
Intel 660p Series M.2 2280 1TB PCIe NVMe SSD

(aside, this thing is amazing for 3D gaming and general computing)

dojo_b wrote:Hi, let me just say that I actually managed to get Katago running on Win 10 with 5700 XT seemingly without a hitch or any special steps. Runs fast, plays strong. (Note that I am not doing any neural-net training so far, just using the existing weights.)

Cool, excellent news. Could you give some more details on the performance you get on the most recent 40b KataGo net?
To get visits/second, you can make a command with "katago benchmark -model <NEURALNET>.gz -config <GTP_CONFIG>.cfg"

dojo_b wrote:I'll try to give playout stats if you can explain how (I couldn't find it).

AMD Radeon 5700 XT gpu
AMD Ryzen 7 3700X 8-core cpu
32 GB DDR4 SDRAM
AORUS B450 Pro WiFi mobo
Intel 660p Series M.2 2280 1TB PCIe NVMe SSD

(aside, this thing is amazing for 3D gaming and general computing)

Looks like a great machine!
For benchmarking, try to follow Friday9i's guidance.
If that’s tricky, ask here and I’m sure you’ll get help.

In the meantime, put in a few board positions, see how many playouts it does in a minute and divide by 60.

I’m sure we'll ask questions about your katago config next.

I’m just delighted that it’s working at all.

Thanks for the info on the finally working AMD cards. Much appreciated.


After watching the Nvidia launch and the infos available on the next AMD GPUs for consoles there are two (related) details that will stop me from buying a Nvidia offer at launch this time.

Nvidia increases the power consumption to north of 300W. I am not concerned about the enviromental perspective here, although it is still an argument as well

Nvidia is afraid of the AMD performance, that is why the power consumption is so high.

Why is that? Nvidia is still on the Samsung 10nm chip manufactioring process although it is called 8nm. AMD uses the advanced TSMC 7nm process that already helped to turn their CPUs into a very capable competitor for Intel.

Nvidia has to increase the power consumption to keep up with AMD.

Perhaps they will still be faster a little bit, but the AMD high PC end card will be probably very attractive.

I for one can easily wait till october to confirm my thoughts. I still love my 1080Ti.

dojo_b wrote:Hi, let me just say that I actually managed to get Katago running on Win 10 with 5700 XT seemingly without a hitch or any special steps. Runs fast, plays strong. (Note that I am not doing any neural-net training so far, just using the existing weights.)

Playouts per second?

Thanks for your patience, I had to find time to figure out some of the command-line details. Here is my process, followed by what I think are the relevant results. Any comments/interpretation is welcome.

1. I downloaded the current relevant version of katago, katago-v1.6.1-gpu-opencl-windows-x64.zip

2. Downloaded this neural net [331M] g170-b40c256x2-s5095420928-d1229425124.zip
recommended as the strongest 40-block, from https://d3dndmfyhecmj0.cloudfront.net/g ... index.html

3. after putting the net in a folder alongside katago, ran the configuration process:
.\katago.exe genconfig -model nn/model.bin.gz -output gtp_custom.cfg

-made a few arbitrary decisions: Japanese rules default; 12gb cache budget (3gb is default)
-as I understand it, the process also tested my PC and returned the config file, gtp_custom.cfg

4. ran the command
.\katago benchmark -model 'nn/model.bin.gz' -config 'gtp_custom.cfg'
with results below.

Note, I could hear the PC working extra hard during the benchmark testing.

****************************
Ordered summary of results:

numSearchThreads = 5: 10 / 10 positions, visits/s = 390.12 nnEvals/s = 341.84 nnBatches/s = 137.46 avgBatchSize = 2.49 (20.6 secs) (EloDiff baseline)
numSearchThreads = 8: 10 / 10 positions, visits/s = 483.20 nnEvals/s = 426.08 nnBatches/s = 107.48 avgBatchSize = 3.96 (16.7 secs) (EloDiff +68)
numSearchThreads = 10: 10 / 10 positions, visits/s = 475.83 nnEvals/s = 418.13 nnBatches/s = 84.81 avgBatchSize = 4.93 (17.0 secs) (EloDiff +55)
numSearchThreads = 12: 10 / 10 positions, visits/s = 532.33 nnEvals/s = 471.48 nnBatches/s = 79.88 avgBatchSize = 5.90 (15.2 secs) (EloDiff +91)
numSearchThreads = 16: 10 / 10 positions, visits/s = 583.43 nnEvals/s = 518.51 nnBatches/s = 66.15 avgBatchSize = 7.84 (14.0 secs) (EloDiff +113)
numSearchThreads = 20: 10 / 10 positions, visits/s = 420.69 nnEvals/s = 381.50 nnBatches/s = 39.19 avgBatchSize = 9.73 (19.5 secs) (EloDiff -35)

************************

Also, I did a casual sample of KaTrain 1.4.0 (separate download), in particular its self-play speed for five games, Japanese rules, default settings and after running its own self-configuration by assessing my PC. Times are +-1 sec.

44 s, 301 moves B+0.5

40 s, 239 moves B+0.5

77 s, 300 moves W+31.5

80 s, 310 moves, W+33.5

59 s, 320 moves, W+3.5

dojo_b wrote: numSearchThreads = 16: 10 / 10 positions, visits/s = 583.43 nnEvals/s = 518.51 nnBatches/s = 66.15 avgBatchSize = 7.84 (14.0 secs) (EloDiff +113)

That’s 25x faster than my iPad.
My new computer arrives tomorrow. Fingers crossed I can get comparable performance.

Thanks, dojo_b!

Good news, meh news, bad news, and a bit of optimism for 2020 iMac users with AMD Radeon Pro 5700XT cards (and I would assume the whole line of AMD Radeon Pro 5xxx cards).

Good news - Leela Zero, KataGo, Lizzie, and Katrain all work with the GPU in MacOS.

Meh news - Current performance is far below the Radeon RX 5700 XT on Windows, far worse than the Geekbench OpenCL score implies it should be. Using g170-b40c256x2-s509, my performance was as follows:

numSearchThreads = 5: 10 / 10 positions, visits/s = 118.55 nnEvals/s = 110.16 nnBatches/s = 44.38 avgBatchSize = 2.48 (34.1 secs) (EloDiff baseline)
numSearchThreads = 6: 10 / 10 positions, visits/s = 147.74 nnEvals/s = 133.44 nnBatches/s = 44.94 avgBatchSize = 2.97 (27.4 secs) (EloDiff +75)
numSearchThreads = 8: 10 / 10 positions, visits/s = 175.55 nnEvals/s = 160.67 nnBatches/s = 40.89 avgBatchSize = 3.93 (23.2 secs) (EloDiff +127)
numSearchThreads = 10: 10 / 10 positions, visits/s = 138.49 nnEvals/s = 128.36 nnBatches/s = 26.28 avgBatchSize = 4.89 (29.5 secs) (EloDiff +21)
numSearchThreads = 12: 10 / 10 positions, visits/s = 176.73 nnEvals/s = 164.47 nnBatches/s = 28.04 avgBatchSize = 5.87 (23.3 secs) (EloDiff +103)
numSearchThreads = 20: 10 / 10 positions, visits/s = 127.83 nnEvals/s = 120.36 nnBatches/s = 12.63 avgBatchSize = 9.53 (32.8 secs) (EloDiff -87)

Lifein19x19 user dojo_b got 580 nnEvals/s in 16 threads with a comparable card and RAM on a Windows machine. If anyone has ideas on how to improve this performance, I'm all ears!

Here are my specs:

AMD Radeon Pro 5700 XT gpu
3.8GHz 8‑core 10th-generation Intel Core i7, Turbo Boost up to 5.0GHz
64 GB 2666MHz DDR4 memory
APPLE SSD AP1024N

Bad News - I believe the original corruption that affected Radeon RX 5700 XT in Windows is still present in Windows 10 boot camp. I installed Lizzie, Leela Zero, and katago in Windows 10 boot camp. Leela Zero would run for a few seconds then quit, and katago wouldn't run at all. Because the symptoms match the original disease, I'm not going to explore this further for now.

Reason for Optimism - Drivers get better over time. Just as AMD fixed the bug in Windows Radeon RX 5700 XT drivers, hopefully AMD and Apple will improve the Apple drivers, and fix the boot camp drivers.

Thanks to L19 user yoyoma for his help getting the builds to work in MacOS. Dude turned a whole weekend struggle into a two-hour collaboration.