Topic: WildRig Multi 0.35.1 beta 2 multi-algo miner for AMD & NVIDIA - page 63. (Read 92264 times)

Thank you Majika, I took the liberty to link your post about the HEX hashrate to our XDNA Ann thread!
https://bitcointalksearch.org/topic/m.46554102

Intensity 23 is too high for RX560 card, and that's why you get n/a on it. And worksize 256 is not optimal as I know. What about hashrate - seems you get worst hashorder. Today I will make some benchmarks with worst and best situations.

my test with something hex
 
GPU0 ASUS ROGSTRIX RX560 4GB
GPU1 MSI ARMOR RX570 8GB
GPU2 ASUS ROGSTRIX RX570 4GB
GPU3 MSI ARMOR RX480 4GB
GPU4 MSI GAMING RX580 4GB
GPU5 MSI ARMOR RX470 4GB

wildrig.exe --algo hex --opencl-threads 3 --opencl-launch 23x256 --strategy=0 --multiple-instance --print-full --url=stratum+tcp://xdna.suprnova.cc:4919 --user=USER.WORKER --pass=x

Test pool(s)
supranova

stratum+tcp://xdna.suprnova.cc:4919     0% fee atm [Primary]
stratum+tcp://rvn.suprnova.cc:6667       1.0 % fee [Primary]
stratum+tcp://eu.ravenminer.com:1111  0.9% fee [backup]

yeah, as I start mining my shares fly up the mining console screen but as the diff adapts these occurrences slow down and then stops as pool diff is normalized.

i'm going through every setting no matter how silly it looks I want to be methodical as possible and go through every switch/option to find the sweet spot for each algo. tbh, whats already on the OP example is pretty much spot on. but on the HEX algo. Its the fastest miner I've used period!.

Regards the X16r I need to finish going over (repeating) tests with the --strategy=1 option enabled.

The Strategy experimental setting is an interesting concept to make the (*multi-gpu*) mining rig act as a single device. I think that's cool.
I have never came across something like this before. 

True. It all depends on the pool.
Can the above member say which pool he is using for conducting his test using the miner on HEX?

About 2-3 shares per GPU - that depends on difficulty set by pool, lower difficulty and higher intensity - higher chance to get more than 1 share to be solved What about x16r algo - this is hard to find faster parameters because of random hashorder right now, without bench mode.

Test RIG x10 Sapphire RX580 Special Editions 1430Mhz/2250Mhz BIOS modded from 1750 timing straps

AMD TEST RIG - Platform & Drivers.
Radeon Software Version - 17.11.2
Radeon Software Edition - Crimson ReLive
Windows Version - Windows 10 (64 bit)
System Memory - 8 GB
CPU Type - Intel(R) Core(TM) i7-7700K CPU @ 4.20GHz
GPU #0-9 Radeon RX 580 Series
kernel: Ellesmere, intensity: 23, worksize: 256/256, cu: 36


 * VERSIONS:     WildRig/0.12.0 beta libuv/1.21.0 OpenCL/2.0 MSVC/2015
 * CPU:          Intel(R) Core(TM) i7-7700K CPU @ 4.20GHz x64 AES-NI
 * ALGO:         x16r
[16:01:42] donation: 2%
[16:01:42] strategy: 1
[16:01:42] send-stale: no
[16:01:42] watchdog: enabled
[16:01:42] good-risers: default  <== Will give this option a test under a few different configurations and report back.

Indeed, I like to try every combination to squeeze maximum performance out of my mining setup.
I try to be as methodical in my testing as possible. Ill keep posting my results and once complete I will post my tests here asap!
 with --opencl-threads 4 I tested lowering the Intensity and worksize but --opencl-threads 3 results are best for my setup. However the remaining variable which I need to duplicate some tests is with the  --strategy=1 changes (from --strategy=0).


I noticed with thread count set to --opencl-threads 4 (Worksize 256 and -I 23) makes it kind of choke up with stale shares, my rig seems much happier when it's set to --opencl-threads 3 (as per what was in the OP) in the examples.

As for the --strategy=0 | --strategy=1

..interesting

So if i understand how it works it correct. Setting the --strategy switch to "=1"  the workload will be "attacked" and/or "Hashed" by all 10 GPU's in my rig working in unison as opposed to
--strategy=0 where each of the 10 GPU's process the workload independently as an individual, ungrouped device split among each of my 10 gpus . Excuse my crude description about how the --strategy switch is setup to work.. -- TY:)


Could this be why I see so many 2 shares found at once messages (whilst in  --strategy=1 mode:-


============
S1t4i21w256
wildrig.exe --algo x16r --strategy=1 --opencl-platform=1 --opencl-threads 4 --opencl-launch 21x256

 
S1t4i22w256
wildrig.exe --algo x16r --strategy=1 --opencl-platform=1 --opencl-threads 4 --opencl-launch 22x256


S1t4i23w256
wildrig.exe --algo x16r --strategy=1 --opencl-platform=1 --opencl-threads 4 --opencl-launch 23x256


S1t3i23w256
wildrig.exe --algo x16r --strategy=1 --opencl-platform=1 --opencl-threads 3 --opencl-launch 23x256
My Hashrate: 92.91 MH/s



S1t3i22w256
wildrig.exe --algo x16r --strategy=1 --opencl-platform=1 --opencl-threads 3 --opencl-launch 22x256
My Hashrate: 182.20 MH/s


S1t3i21w256
wildrig.exe --algo x16r --strategy=1 --opencl-platform=1 --opencl-threads 3 --opencl-launch 21x256

=============

ALGO: HEX  More settings to test and report back on. (ASAP!)


Wildrig-multi is a BEAST at mining HEX Algo !!!

This is twice as fast as SGMiner-HEX (AMD)-v04.R Wildrig miner is a beast for HEX algo.
Im getting about 7.9~8.6MH/s per GPU -- POOL side reporting My Hashrate: 88.85 MH/s

Look at this also. Im finding LOADS of these 2 shares & some found 3 shares at once.



--edit

I find " --strategy=0 " is faster than  "--strategy=1" (***For HEX**) But this could be coincidental and a lot more testing many more possible combinations before im 100%
Fastest HEX & X16r algos miner I have found for AMD rigs so far. Excellent work Dev

I was going to write that 4 threads is too big amount, but seems you came with the same conclusion

Yeah, there still some places to fix that thing with few threads per gpu, because it shows thread id instead of gpu id.

What about strategy - that's experimental parameter came from original wildrig. 0 is default and divide nonce range(0..2^32) for all GPU's, so every GPU(or thread of it, if use --opencl-threads) work in its own range. Using strategy 1 all GPU's will work as one device - starting calculation from 0 to 2^32. In some cases that can give some advantage for amount of solved shares, but I didn't make any experiment, and for x16r this will be too hard

Ok to answer your question a little overview on my test rig
Test rig#1  ASUS B250 19pcie, 1600w + 1200 w PSU(s) + x10 Sapphire RX580 8gb NITRO+ Special Editions 1430Mhz/2250Mhz -0.50mV

I produce on Raven X16r algo ~ 9 - 11MH/s per card

accepted(123/0) diff 3.68G avg. effort 98.43% (24 ms)
[21:02:27] GPU #0: 10s/60s/15m 9412 10069 7783 kH/s
[21:02:27] GPU #1: 10s/60s/15m 9350 9878 7839 kH/s
[21:02:27] GPU #2: 10s/60s/15m 9290 9870 7833 kH/s
[21:02:27] GPU #3: 10s/60s/15m 9318 9889 8042 kH/s
[21:02:27] GPU #4: 10s/60s/15m 9424 10001 7868 kH/s
[21:02:27] GPU #5: 10s/60s/15m 9284 9885 7800 kH/s
[21:02:27] GPU #6: 10s/60s/15m 9420 10008 7926 kH/s
[21:02:27] GPU #7: 10s/60s/15m 9427 9901 7694 kH/s
[21:02:27] GPU #8: 10s/60s/15m 9427 9931 7729 kH/s
[21:02:27] GPU #9: 10s/60s/15m 9295 9882 7776 kH/s
[21:02:27] speed 10s/60s/15m 93651 99318 78294 kH/s max: 150644 kH/s

I plan to test different intensity:worksize combinations/settings and will report back.
--opencl-threads > 2,3,4
and play around with  --strategy=1 (if this will do anything or not) currently with those hashrates above "strategy: 0" was set.
Idk how many different --Strategy options there are 0,1,2,n...

I may split the devices over like 3 miner instances using the "--multiple-instance" option and changing the threads this way to also check on faster hashrates.

FYI:

Im seeing lots of these:-
**checkout the GPU count #26 ??*
Im only running 10 gpu's in this rig. it could be due to the of threads (4) I have set. idk..  

This is far from optimal settings and lots more tweaks to report/log. BUT HTH

Thanks for this great miner andrucrypt; its faster than Avermore 1.4.1 and is a whole lot more stable miner for me anyway


==UPDATE==


./wildrig.exe --algo x16r --opencl-platform=1 --opencl-threads 3 --opencl-launch 22x256 --strategy=0 --multiple-instance  --url=stratum+tcp://: --user=XXXXXXXXXX.tRIG011 --pass=x --print-full --watchdog

My Hashrate: 152.70 MH/s
148,782.83

What does "stripped kernels" mean?

Released 0.12.0 beta
- slight speedup for bcd
- fixed API for showing correct hashrate when use --opencl-threads
- stripped kernels

Question:
Has anyone figured out any specific arguments that can be used to increase x16r hashrate?

Hm, seems I need to improve my scripts for HiveOS a bit. But you can configure everything using "Extra config arguments". They are command-line, not json So just put there:

What about --opencl-launch - it can be different, check optimal parameters table and maybe experiment for your cards. Also on moment - currently HiveOS will show wrong hashrate in GUI because of bug in my API. Today I will release a fix.

Currently have no plans to work on them.

Can I ask for help with setup wildrig-multi in HiveOS

When I set "Hash algorithm: " to bcd and leave "Extra config arguments: " empty - I'm getting rejected shares "Invalid share" (seems wrong algo selected - it select's aergo by default)

When I set "Hash algorithm: " to --- and put "algo": "bcd" into "Extra config arguments:" field - i get error " ./wildrig-multi: unsupported non-option argument '"algo":'"

What parameters should I set for   "Hash algorithm: " or / and for "Extra config arguments:" to mine BCD  ?

when is it planned to add cpu algo?

Not sure... I mostly automated the test:
start wildrig.exe ... -l algo.log
timeout /t 120
taskkill
start wildrig

Will check logs and post

PS: it seems no accepted shares.
I will rerun the tests using settings from post1 in the topic

Thanks for testing! But I have a question - was there any accepted shares? Because I have a report that there was no shares at all ( And that probably because of big endian architecture, which is not tested in my kernels and I need sample of card to fix that...

Adding some results for Pitcairn and Tahiti
everything worked out-of-the-box - great job!
Obviously, if you ever want to, speed can be optimised - cards were not very hot. Everywhere used "safe settings" - 2 threads instead of 3, no attempt to choose --opencl-launch parameter - any hint if trying different settings (and what...) is worth the try?

All results are after running for 2 minutes. All cards run @ stock clocks eg, 265 is @900/1250, 270 is @1000/1350, 7950 were @900/1250

there is still no benchmark mode unfortunately, so use real pools and wallets 2-5minutes should be ok, but to be sure that it works - there should be accepted shares.