Life In 19x19

Thanks for the report!
OK, so lingering questions I still have:

1. is there a source for the statement that AMD's latest drivers have fixed the issue? Which release did the trick?

2. IIRC, the problem with these GPUs first manifested not as an overt failure to finish some computation, but as hidden arithmetic errors which contributed bad data to crowd-sourced scientific computing projects:
https://www.techpowerup.com/261603/amd- ... -seti-home
Error behavior for 5700 XT + OpenCL was reported as recently as June:
https://community.amd.com/thread/248056

So, how do we have confidence that the neural net is not only outputting moves, but doing so according to the intended computation? Conceivably the performance could be somewhat robust to errors, to the extent that amateur or even 9p players couldn't notice the dip in strength.

There should be a way to cross-validate behavior across PCs from a given position + random seed, but this hasn't been touched on so far and I don't understand the above issues well enough to know if it's necessary.

dojo_b wrote: ****************************
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)

Thread OP here. Finally got it working with a clean install of 20.04 and the 2020-09-29 amdgpu-pro drivers from AMD.

Here are my benchmark stats:

numSearchThreads = 5: 10 / 10 positions, visits/s = 325.34 nnEvals/s = 290.18 nnBatches/s = 116.66 avgBatchSize = 2.49 (24.7 secs) (EloDiff baseline)
numSearchThreads = 8: 10 / 10 positions, visits/s = 416.75 nnEvals/s = 370.84 nnBatches/s = 93.57 avgBatchSize = 3.96 (19.4 secs) (EloDiff +80)
numSearchThreads = 10: 10 / 10 positions, visits/s = 421.40 nnEvals/s = 372.96 nnBatches/s = 75.58 avgBatchSize = 4.93 (19.2 secs) (EloDiff +76)
numSearchThreads = 12: 10 / 10 positions, visits/s = 452.13 nnEvals/s = 407.42 nnBatches/s = 69.19 avgBatchSize = 5.89 (17.9 secs) (EloDiff +95)
numSearchThreads = 16: 10 / 10 positions, visits/s = 506.45 nnEvals/s = 450.22 nnBatches/s = 57.42 avgBatchSize = 7.84 (16.1 secs) (EloDiff +124)
numSearchThreads = 20: 10 / 10 positions, visits/s = 376.52 nnEvals/s = 342.27 nnBatches/s = 35.08 avgBatchSize = 9.76 (21.8 secs) (EloDiff -13)

and my PC hardware: AMD Ryzen 7 3700X, 64GB RAM, ATI Radeon RX 5700 XT

Hopefully, we'll soon get some info on KataGo's performance with new AMD Radeon cards (such as the RX 6800XT and 6900XT)!
When someone gets one, please share your test results with us !

Friday9i wrote:Hopefully, we'll soon get some info on KataGo's performance with new AMD Radeon cards (such as the RX 6800XT and 6900XT)!
When someone gets one, please share your test results with us !

Anyone gets info on Radeon cards for KataGo: is it working well now, and if yes, what is the performance? Thanks a lot