TODO: Test Silent army v5 of a wide variety of devices

[maztheman]

Discuss your results here

[kruisdraad]

Windows build please

[maztheman]

it is failing, i have to fix it up...

[maztheman]

If you have a 1070 or 1080 please test it:

https://github.com/maztheman/nheqminer/releases/tag/v0.4h

[drigger]

[maztheman]

try:
https://github.com/maztheman/nheqminer/releases/download/v0.4h/v0.4h_1070_plus.7z

i remove all the thread waiting..

[maztheman]

the v5 silent army doesnt seem to be too great when converted to CUDA :(, however It might just be the way its looping...i might be able to fix it

[drigger]

Hmm, w/ 0.4h_1070_plus this happens: ;)
Just for the sake of comparison - this is what I get w/ the same 1080 on one of the original SA5 python ports for windows:

[maztheman]

yeah there is some major issue with cuda and that code that was posted for opencl. cuda keeps "timing" out when I try to do same kind of calls...I guess itll have to be a work in progress for now.

Ill post again when I see improvement with my GTX 650, which should then indicate some progress with 1080s, etc.

[maztheman]

I see from the author or silentarmy:

mbevand commented 2 days ago • edited
@tupieurods You are right. Didn't know shared atomics were not hardware implemented pre-Maxwell. That' s definitively the cause of the slowdown then, because this commit makes heavy use of shared atomics. I see no solution other than maintaining a 2nd separate version of input.cl specifically for pre-Maxwell Nvidia GPUs then.

[maztheman]

Well based off that guys message I reverted back to a more direct conversion, maybe itll work better:

https://github.com/maztheman/nheqminer/releases/download/v0.4h/v0.4h_MAXWEL_PLUS.7z

It was super slow on GTX 650 but looks like it will be because of the shared atomics

[drigger]

It gives me much higher I/s numbers + full GPU load, but no Sols, unfortunately =)

[maztheman]

hmm wierd, i get 78 sols/s with my R9 290 with the v5 code in open cl. 0 sols probably because some timeout is causing a crash

[kruisdraad]

I ran both versions.

Setup: 6x GTX1070 with Windows 10 latest drivers, etc (anno vers)

0.4h plain: 241 sol/s power usage 51%
0.4h maxplus: 2.4 sol/s and a near 400 I/s. Also the power usage is above 70%

previsous version got about 250 so its a little less, power usage is much more stable though.

[maztheman]

hmm 241 is really not that competitive to the SA linux version, is it?

[kruisdraad]

Honestly i havent tried the linux version yet, The zcminer-dev windows version gets about 300 so its not that bad.

You have Linux example on actual speeds?

[chronosek]

Compiled. For me crashing after 20 sec (i see constantly increase use of gpu memory till all 4GB is fillup) and then nvidia is in P5 locked state..

[23:00:05][0x000011dc] Using SSE2: YES
[23:00:05][0x000011dc] Using AVX: YES
[23:00:05][0x000011dc] Using AVX2: YES
[23:00:05][0x000011a8] stratum | Starting miner
[23:00:05][0x000011a8] stratum | Connecting to stratum server eu1-zcash.flypool.org:3333
[23:00:05][0x00001194] miner#0 | Starting thread #0 (CUDA-SILENTARMY) GeForce GTX 980 (#0) BLOCKS=64, THREADS=512
[23:00:05][0x000011a8] stratum | Connected!
[23:00:05][0x000011a8] stratum | Subscribed to stratum server
[23:00:05][0x000011a8] miner | Extranonce is fa613f0f55
[23:00:05][0x000011a8] stratum | ←[35mTarget set to 004189374bc6a7ef9db22d0e5604189374bc6a7ef9db22d0e5604189374bc6a7←[0m
[23:00:06][0x000011a8] stratum | ←[36mReceived new job #11ee30962342110c9785←[0m
[23:00:20][0x000011dc] ←[33mSpeed [300 sec]: 8.72074 I/s, 15.646 Sols/s←[0m
[23:00:36][0x000011dc] ←[33mSpeed [300 sec]: 8.94374 I/s, 16.6693 Sols/s←[0m
CUDA error 'out of memory' in func 'sa_cuda_context::solve' line 1029
CUDA error 'out of memory' in func 'sa_cuda_context::solve' line 1030
CUDA error 'unspecified launch failure' in func 'sa_cuda_context::solve' line 1073
CUDA error 'unspecified launch failure' in func 'sa_cuda_context::solve' line 1024
CUDA error 'unspecified launch failure' in func 'sa_cuda_context::solve' line 1025
CUDA error 'unspecified launch failure' in func 'sa_cuda_context::solve' line 1029
CUDA error 'unspecified launch failure' in func 'sa_cuda_context::solve' line 1030
CUDA error 'unspecified launch failure' in func 'sa_cuda_context::solve' line 1073
CUDA error 'unspecified launch failure' in func 'sa_cuda_context::solve' line 1024
[ ... ]

[maztheman]

yes, sorry there is a memory "leak" that i already have fixed but not have pushed
Its only supposed to create the 2 buffers once, and reuse the, in my old code I was creating it every time and not deleting....:P

[auroracoin]

I tried it on a 750ti comp 5.0 and I get this massage. Missing file MSVCP140.dll for win 8.1

[maztheman]

oh, you need to install the redist file:
https://www.microsoft.com/en-ca/download/details.aspx?id=48145

[maztheman]

@chronosek ive checked in the fixes

[auroracoin]

That worked...thx :)

[chronosek]

@maztheman thx, not crashing now, but always get 0 sol/s no matter what -cb, -ct will set

[maztheman]

Yeah some internal issue, ill have to fully debug it again...

[dtawom]

Getting @ 17-18 sol/s with a 1060 3gb card with 0.4h. I found the best rate for this card is obtained with -cb 128 -ct 32 (autodect puts it at -cb 63 ct 64 and only yields 15 to 16 sol/s.

[krnlx]

removing packed atomic counters = bad, I tested it in opencl, it is fastest. I tested packed 64-bit atomics too, 1-2% slower

[chronosek]

tpruvot did good job porting to cuda, specially with atomic part (working stable at 58 sol/s but still eqm doing 66 sol/s), i think maztheman problems was from some hardcoded values or other code what was not in silentarmy

[maztheman]

I added a test program that will help me debug this cuda issue. Please post your log files.

[drigger]

log.txt

[maztheman]

I created a new build that should "work" on 1080 and 1070's again. Probably wont break any records though...

[maztheman]

I uploaded a 16 thread version

[maztheman]

Thanks for your testing! Not exactly what I was expecting but seems each card has its own sweet spot.

[jddebug]

t16
16.5 sols on gtx760
So far seems the 32 is the best.

Whenever I start the miner it always shows blocks 42 and threads 64. Is that when the mining software thinks I should be using?

[maztheman]

The original miner used that information but I dont use it.

42 is 7 * your sm count, which I guess is 6. It seems silent army uses a different block dimension so that it can use the thread id as an index into the hash table.

[drigger]

My results with latest beta:

GTX 580:

• t16 22.41
• t32 27.5
• t64 32.3
• t128 29.5
• t256 26

GTX 550 Ti:

• t16 5.7
• t32 7.4
• t64 7.5
• t128 6.5
• t256 5.4

GTX 650:

• t16 5.3
• t32 6.6
• t64 8.8
• t128 8.6
• t256 8.0

[maztheman]

@drigger. Thanks for the testing.

[Kahana82]

I'm getting around 100 sols with GTX980 with version L now
best result (102 sols) with -cs -cb 512 -ct 64 -cd 0

with cpu (7 out of 8 threads) it gives around 120 sols
-t 7 -cs -cb 512 -ct 64 -cd 0
I found keeping 1 thread free (7 of 8 total) gives better result on gpu and same if not better final sols.

On previous versions (i) best (42 sols) was -cs -cb 256 -ct 32 -cd 0
around 66 sols with -t7 -cs -cb 256 -ct 32 -cd 0

This is a huge improvement ! keep up the good work Maz :)

[maztheman]

Tnx I'll keep porting any enhancements I find.

[dtawom]

@jddebug Kenai/Soldotna area. Thanks for the updates @maztheman, I'll post updates for all my stuff tomorrow.

[ceozero]

Can you please send me a private generation absenteeism?

[maztheman]

@ceozero Sorry I dont understand what you mean...

[ceozero]

@maztheman I need a new miner. You need to modify the code to re compile. can you do it?

[maztheman]

Yeah I can compile anything for windows.

[ceozero]

May I know your email address? Or do you want to send a mail to me. [email protected]

[maztheman]

I sent you an email

[dtawom]

Getting @ 75-85 sol/sec off my GTX 1060 3GB with 0.4l, for some reason that number doesn't improve much when I have all 4 CPU's running too. I was getting better results using your miner for CPU and Silentarmy for GPU at a combined 100 sol/s. Default detection for the 1060 is 63 blocks & 64 threads, I cap out at about 75 with this setting, 128b and 32t seems to give me the best capping out at about 85 sol/sec. Is there a calculator out there that will tell you what the best block and thread count is for any given model and ram size video cuda card?

[ceozero]

I can't run, I'm still working on that.

…

在 2016年11月29日，上午4:08，dtawom ***@***.***> 写道： Getting @ 75-85 sol/sec off my GTX 1060 3GB with 0.4l, for some reason that number doesn't improve much when I have all 4 CPU's running too. I was getting better results using your miner for CPU and Silentarmy for GPU at a combined 100 sol/s. Default detection for the 1060 is 63 blocks & 64 threads, I cap out at about 75 with this setting, 128b and 32t seems to give me the best capping out at about 85 sol/sec. Is there a calculator out there that will tell you what the best block and thread count is for any given model and ram size video cuda card? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#5 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AVeH-ZwE_v68V9-n2Rw-fCDUF63s7Phvks5rCzSzgaJpZM4KxkRY>.

[maztheman]

@dtawom , the silentarmy does not include some of the changes I added, so it maybe more efficient for that specific card. unless you are using the https://github.com/zawawawa/silentarmy version. I ported those changes and some other nvidia specific parameters and stuff. If I had all the nvidia cards I would be testing and testing before I created builds. I just have to trust that the changes im porting "work" and "work well" with all nvidia 10xx series cards. There is a "Version 6" of silentarmy coming VERY soon. I will be porting those changes ASAP.

Stay tuned...

[dtawom]

@maztheman yeah the zawawawa one is the one I was using. The difference is so negligible (5-10 sol/sec per machine) that it is easier just to use your latest build only. The new nvidia one you made for older cards is getting me @ 40 sol/s on my 580 gtx which is @ 15 more sol/sec than I was getting before. Thanks soo much for the updates, they've got me hitting peaks of 650 sol/s now between all my rigs. Looking forward to silentarmy V6, hope the improvements are as significant as they were from 4 to 5.

[maztheman]

I'm just glad the work I'm doing is useful to someone :-)

[ceozero]

@maztheman what‘s time update. to Faster.,

[maztheman]

A few more days

[dtawom]

I dunno if this guy is running off silentarmy v6 or something else, but these numbers are soooo sexy. Gonna give um a try on my 1060 rigs. https://forum.z.cash/t/ewbfs-nvidia-cuda-zcash-miner-1060-170-h-s-gtx-1070-250-h-s/12523

[maztheman]

Oh it's closed source so I can't see what he is doing. I can profile the exe probably and see what I can see

[dtawom]

Confirmed 175 sol/sec on 1060 with no CPU. You should do what this guy does and take a couple percent for yourself. I'd rather support you than him.

[dtawom]

Holy moly, getting triple speed off my R9 Fury's now with claymore. 325 H/s per card with https://github.com/nanopool/ClaymoreZECMiner
Gonna break the gigahash mark per sec tomorrow :)

[maztheman]

:) wow, looks like claymore is on top again!

[maztheman]

Looks like an update will happen pretty soon...

[maztheman]

new build new thread

#8