Topic: hsrminer - Nvidia mining software for various algos by palgin&alexkap - page 50. (Read 30774 times)

It's not the money question, when I was building my farms I choosed 1060 3G setups, just before this spring havoc around GPU shortage I've bought some 1070's and 1080's , after market disaster I didn't even look at hardware anymore 

don`t be like sp_ he earned 100+ BTC for other ppl work and he still don`t have 1080ti nor 1080 I think! 

Seems this time has come when I should buy myself 1080Ti for developement

P.S.: preparing little tweak for 1080 kernel, will update binary after tests.

I updated thread 17.8 is 7870MB use like it can not pass 8GB and card is defo 11GB  

-i 17.9 well its 8296Mb still works
-18 alocate for a sec 8720mb and exit with error Cuda error in func 'neoscrypt_cpu_hash_1080' at line 1447 : an illegal memory access was encountered

klaus ccminer -21 works but use low memory, i23 try just ~3600MB and does not start too

UPDATE:
5x 1070 on dual core
one instance does not slow down compared with 5x, but I need to drop intensity from 17 to 16.x but its looks like system is limted by just 8GB of ram and 16GB swap

so only 1080Ti have problems with multy GPU with hash drop on ine instance all GPU

OK, thank you for tests. 18.5 and 19 is way too high, no such memory amount on consumer-grade GPUs nowadays 

drv 388.59 win10 (8gb ram, swap file 32GB)
one GTX 1080ti
 -i 17.8 works memory usage: 7870MB (msi afterburner) like it can not pass 8GB? BUG?! card is 11GB
 -i 18 Cuda error in func 'neoscrypt_cpu_hash_1080' at line 1447 : an illegal memory access was encountered.
 -i 18.5 Cuda error in func 'neoscrypt_cpu_init' at line 1304 : out of memory.
 -i 19  Cuda error in func 'neoscrypt_cpu_init' at line 1303 : out of memory.

ccminer works on

Will test later

Try tweaking intensity higher,  17.5 allocates little more than half of 1080Ti memory. Try 18 if you can.
Think I'll add some kind of scheduler in next release to tweak work distribution manually.

Strange, I'll need to collect more feedback before any explanation to this behaviour, will add your config to my list.

Thank you for participation!

Palit gamerock 1080 - they are all micron GDDR5x
real GPU clock 1960Mhz (power limit 90%) manual clock

memory real 5022, +504 in MSI afterburner

1070 in same PC with less clock ~1850Mhz is faster by 10% and does not ask for memory OC

Also 4x 1080Ti now works on one instance but hash rate drop 6-7% per GPU
4x instances is faster, same settings, -i 17.5
maybe you need to tie each GPU to different CPU core in one instance?

Can you clarify who is manufacturer of your 1080? And which memory is onboard (Micron, Hynix or Samsung)? All my 1080's hash 1280+ without overclock, manufacturer is KFA2, memory is Micron.

Any intensity tweaks, OC, TDP limiting, or running on defaults?

GTX 1070 + 1080 now hash with one instance fine... gtx 1080 speed match ccminer speed I think there is room for it (pool side is now fine)
1080 like need more memory OC, with +500mhz it give 1140khs


will test 1080ti and 1070 rigs later

Download updated, you can check latest version in the first post of this thread.

GTX 1080 (not Ti),1070 and 1060 3Gb users should run without any problems without intensity tweaks.
All 1060 6Gb, 1080Ti and higher grade card owners should tweak intensity first to fit their hardware. 9xx family and lower not supported in this release.

Please, send feedback, I need to be sure problem is fixed permanently.

UPDATE: found source of that nasty "one binary - one card" bug, now I have to remap all memory pins in kernels to fix hashing. Still far from flawless run, but it's progress  



P.S.: on a screenshot you can see thr0 and thr1 are 1080's, thr2 is 1070 and they run together without conflicts (but now hashing is broken)

UPDATE2: done, nearly 0 CPU usage and different configs start falwlessly (if intensity's defined right, of course). Will re-upload to GitHub in 10-15 minutes.

Little explanation, it won't work at -i 10 because it needs exact number of threads running. Each thread requires some amount of memory allocated on Host and on Device, high speed is gained by rolling only those operations that needed by algorithm without idling. For 1060 3G kernel reaches ~95% efficiency, 1070 ~90% efficiency, 1080(not Ti) ~47-50% efficiency, so for GDDR5X memory it's still space for improvement, but it can only be gained by another memory dealing technique which I have ~0 practice at.

For single 1070 intensity should be 16.5 or 17.5 to start correctly. Seems multiple card rigs won't start correctly after CPU fix, but you're free to launch 1 instance per card, without CPU load it's the way to avoid launch failures.

Regarding illegal memory accesses, C-part of hsrminer code, including correct work distribution among GPUs, is alexkap's work, but I haven't received any info from him, somewhere in code multiple theads are treated as one (GPU0 thread), you can see it when launch single card with number other than 0.
I can spend hours digging alexkap's code and find this nasty bug, but it would be better to get answer from him directly first. Maybe he's dissapointed because nobody's using HSR version and decided to quit this project, I dunno, I haven't been informed about that. I'll solve problems in the order they appear.

I can't even start 1070 at -i 10.
I have the latest 388.71 driver installed.

I have, first version worked fir 1  card, this one does not on gtx1080ti
maybe -i does not work

also tried -i 14 in gtx1070 it start but all shares are invalid, -i 17 works

Play a bit with intensity, put -i 17.5 or -i 18 or -i 18.5, default one won't work with 1080ti

GTX 1080ti now no way to start...

Cuda error in func 'neoscrypt_cpu_hash_1080' at line 1427 : an illegal memory access was encountered.

UPDATE: made fix for high CPU load, now it's barely loading CPU, added ccminer-style intensity option to play with, but work distribution bug is still there, so work continues...

Please, re-download binary from github, I've uploaded CPU-fix + intensity feature version. Tested intensities are 17.5 for 1080/1070, 16 for 1060 3G.

Thank you!