Hi there,
I've seen a lot of interested people on stats on the new pascal GPU's from Nvidia. I am actually setting up a new farm and did some testing the last few days wither i need pascal or old-style AMD's.
So out of the box the GTX1080 really sucks, so bad i did not even consider trying. So yes them story's are true.
however (!) the GTX1070 is a totally different story, and the reason is different memory. Perhaps its the driver that really doesn't support the 1080 in performance and it might take a while ... but with the price difference its not interesting anyways.
So back to 1070's. If you open the box and leave it stock you can forget running it in windows. Windows 7 just crashes and Windows 10 will run it, but very very poor performance with maybe 1MH/s tops (my i7 internal GPU performs better with 2MH/s)
so now its Linux's time. Out of the box with Genoil's 1.1.4 compiled with Cuda8 (which you need to install + latest beta driver) then you will get around 22 MH/s flat out, no tuning at less then 100Watts
With some basic tuning with SMI (which uses the outdated API and is pretty limited) you can bork it up to 26MH/s at the 118W usage. If you look the base power limit is 150W and can be extended to 170Watts so loads of room to play with. But as said with SMI you really cant and you need the real GPU tuning tools from Cuda. Once going there (and with a big help with a AsI tool trying all the options) i was able to find the highest possible overclocking settings.
pom pom pom, with fully maxed out you can get: 31,92 MH/s using 145 watts (so you can keep the 100% power limit) and with the FAN at max only gets 53C.
I have Asus GTX OC 970's to compare them and those are based with 18MH/s and can be tuned to their TDP max with 22-22,5 MH/s but is using 160Watts at that point. So the 1070FE is 50% faster and still uses 10% less power. You can even get the power down a few watts by tuning the FAN's and set it to like 70% perhaps.
This is all with single cards, my experience with 970's is that combined in pairs you can get a little more (perhaps 1MH/s) and in addition its not a lot of time spend with manual tuning. At some point i was getting 32,96 MH/s but after a few minutes it got unstable. With a lot more time tuning perhaps you can get that extra 1+ MH/s and if you ever get it stable, do post your settings.
Photo's with 36 GTX1070's following soon
2 ·
Comments
i've recompiled the opencl miner with the cuda patch and that seems to be really stable now (so far no crashes)
More to follow
Can you provide the link to the tools?
What clocks for mem and gpu you've used to get 31mhs?
i will be creating a nice long blog about it somewhere including pictures and suchs. With a bit of luck it would be somewhere this weekend.
I think i found a driver (or ethminer) bug which hangs ethminer after 4-5 hours even @stock and that is my focus now so stuff is stable and also some Cuda testing which might get a little more power as Genoil said.
I am working with 6 GPU's but its only seeing 5, the ones in the X1 slots ... for some reason the GPU in the X16 slot with riser is not even detected. Did you do something for that?
It would ve nice if you shared your settings so i can check my setup
seems one of the risers was bad, replaced it and got 6 cards now.
At this point: graphics offset -200 to +275 does really do much for Ethereum. Max i got without crashing was +285, but not sure how stable that is. for RAM i am trying 1775/1800 to check stability (gone up to 1885 with hours or good mining)
I am currently going slower to check for crashes but i am getting over 150MH/s atm on 6 cards, but still a long way to go to get the single card (32) x6 = 192MH/s. A lot of tuning with stuff like --cl-local-work (global doenst do anything for me?) is required else you'd be getting only 70MH/s. I changed it to 256 but been playing with 512 and 1024 too.
@Genoil : with all the problems with X and headless virtual monitors i finally got the base working and didn't got around to install the other box. i will do that tomorrow as its nearly 2AM here now
Latest test... 2 480's faster than the 1080 FE...
Dt say much about hashing power but it says a lot about price / wattage
At this point i am getting a poor 130-140MH/s on the entire rig, instead of the 156MH/s at base clocks. With overclocking i seem to get it semi-stable at 150 but not even close to the 6x 32MH/s = 192MH/s based on the single card. Still the power limit is not fully loaded so there is still a lot of wiggle room.
@Genoil i am starting the 14.04 installation with CUDA8 and see if i can compile it and get it running. I had some troubles with getting all the cards online as a few risers seem to be broken
https://smile.amazon.com/PCIE-Adapter-Extender-Flexible-Extension/dp/B00HW1O6OG/ref=sr_1_4
As opposed to the sata or molex-powered board extenders like these:
https://smile.amazon.com/Electop-Express-Riser-Extender-Supply/dp/B014ZTVZB4/ref=sr_1_8
Also, I've noticed strange behavior with two different GTX 1080's, one MSI FE with 5 stage power, the other Gigabyte's Gaming G1 triple fan cooler with 8 stage power. After hours, sometimes 24+, sometimes less, the hash rate jumps up +2-4% (but remains contstantly lifted by that fixed amount thereafter) with no changes made whatsoever, and the power draw goes from its usual slightly noisy band (I have an 81% limiter) to a total flatline, sometimes right at the set value of 81%, sometimes lower (right now it's reporting 71% flatline and a hash rate is +4%). On it's own a few hours or days later it goes away, comes back. It's only a problem for cooling, but I'm running at 69 degrees default, so it only hits 74 max, which is fine. It's just unnerving because it's unpredicatble, even with two different manufacturers and totally different power systems. Have you seen that behavior? You're in linux, so if not, it's probably a win 10 thing. I guess we get the anniversary upgrade for free next week, then new drivers should normalize geth rates. I'm hashing 52mhs/card on lyra2v2 and selling for btc, then buying eth, which is working great.
That you are a total n00b does not mean everyone tries to rip you off. Was pretty easy to create diffs, check bash history, etc. Also if you don't trust people then don't give them access on your system dude. I have loads of others that are very happy with my help and you were just wasting my time because you kept 'demanding' stuff you knew you weren't getting. so good luck getting your stuff to work!
background: this guy is having issues with mining and couldnt get it working. He PM't me and i said i wasnt about to share all my scripts, tools, etc which i run my farm with but i could do a basic install to get him started on mining. That didnt include scripting, management or monitoring. Then he kept asking to get the scripts, which up front he knew he wasnt getting. Happy to share screenshots, skype log on this chat.
For 7.5 ETH's (to be clear, not ETC) ill provide a custom installation for you and you only need to install a clean 'server' Ubuntu 14.04 installation all ill arrange the rest using SSH. After that you just change your password and update the mining address and its all yours. Also i need the exact 1070 models and an IRC/Skype for direct chat so i can tune / overclock the system to its max (i might need a few reboots from your end if i overclock to far and the system hangs).
Just a tip: Mining ETC is not that profitable at this moment. It needs to be above 0.00343 with current mining diff, else ETH would be more profit. Unless you want to support the ETC community of course. The setup i can build for you allows you to easily switch between ETH and ETC within seconds if you want."
this is what he write me... after he contact me on skype he told me that only provide basic setup and not showing me how he setup and what script he used on my OWN server...
everybody should make there own judgement to trust him or not.
for me this is really suspicious that paying $100 and i dont even know what he is doing on my own server.... isn't it?
btw i dodnt asked you to give me all your script for $100 i only ask you to share the stuff what you entering on my system and using it. this is for security and for learning also. not sure why is that a big deal. only thing i can imagine that you doing some shady stuff thats why you dont want to be this transparent. again watch out for this guy
to answer your question: There is no clear answer for this, its weird and certainly something you should figure out. I've been mining for some time now and i never seen sudden bumbs. If you have them bumb its something you should be getting in the first place.
Troubleshooting this will take you weeks, perhaps months ( i had similair problems with stability which took me ages ). Nvidia monitors voltages on stability so i wouldn't think its the PSU. My first (best) guess is either a kernel driver (timing) change or a state change. Look in DMESG and SYSLOG at the moment you get the bumb. Look for anything, even stuff like:
[ 8632.948282] perf interrupt took too long (2507 > 2500), lowering kernel.perf_event_max_sample_rate to 50000
also check your BIOS for anything that 'tunes' performance automatically.
Finally dump the nvidia-smi status into a log and at the moment of the bumb check for power consumptions and P(erformance) state
but like i said, its kinda hard to tell you wat to look for. And i am not the 'free support guy' on this forum
Edit: after typing i did got an idea, is your monitor connected to one of the cards? Perhaps your monitor (X-windows) went to standby / not the screensaver and giving you a little extra on there? Also run a ethminer for each card and see if its just 1 card, or the whole rig ... if its just 1 card it would support this idea or at least gives some direction to look.
good luck
Also if you have a GTX1080, try using Genoil's 1.1.7 ethminer which fully supports CUDA8 ... perhaps you get more stable mining. You must run it with -U and not -G ... elso you get OPENCL. If you are now mining with -G / OpenCL you might get a nice bumb in performance as well.
Some cards have a boost clock, write done both base and boost clock speeds, but i think if you follow that SMI stuff i send ya you will max out the card way above current stuff and get more stable / without bumbs.
If you get it working do lemme know what it was
30MH/s with stock 1080 is not bad actually. please note: the 1080 has less good memory then a 1070, but i got the 1070 to a stable 36,1MH/s ... but windows might be a little better and with stable i mean unattended mining. If you dont care about a reboot (you need a SSD for fast boots) you can get a lot better.
Windows will allow you go push a little further, however its more if you wanna game in between and need windows. If you want to do unattended mining with autostarts, etc ... linux is better. But if Windows bumbs the hashrates for me with a farm that little higher rates on all them nodes really increases profits a lot