Ethereum GPU mining?

[Madpacket]

Looks like the bubble's bursting and people are panic selling. Good opportunity to buy. Hope everyone's going long

Maybe this will slow some of the card sales down.

https://poloniex.com/exchange#btc_eth

Of course I just completed my ghetto milk crate rig lol.

http://imgur.com/a/9e6FJ

 

[metalliax]

Over at https://forum.ethereum.org/discussion/5594/rig-porn-post-your-pics/p5, someone posted some details about his 96 nano mining farm. Each nano running at ~70 W and generating 22 Mh/sec. Pretty cool: https://15254b2dcaab7f5478ab-24461f....vanillaforums.com/editor/ls/vzfdoc04k4og.jpg

link to his details on bitcointalk: https://bitcointalk.org/index.php?topic=1424132.0

 

[poofyhairguy]

My 390s test at 30/each and average closer to 29 over long runs. I would expect 33ish out of the 390x.

My 390x isn't the best overclocker, I learned I can't push it past 1150 while mining or the 2D display has corruption. At that speed it does 29.5 after 6 hours.

What driver version are you using? I wonder if I have the wrong drivers.

 

[DrMrLordX]

Don't use any driver newer than 16.2.1

 

[DrMrLordX]

Looks like the bubble's bursting and people are panic selling. Good opportunity to buy. Hope everyone's going long

Maybe this will slow some of the card sales down.

https://poloniex.com/exchange#btc_eth

Silly, really. All the things that made Ethereum interesting from the start are still in play. Bitcoin is still having its blockchain problems etc etc. Let them panic sell. If it brings coins back down to $1 apiece, then it becomes easier to buy them up.

Of course I just completed my ghetto milk crate rig lol.

http://imgur.com/a/9e6FJ

Nice. You using anything as spacers to keep the cards from leaning on one another? And did you have to run any ground wires?

edit: oh I see, you have the mobo down in there too. Thought you might have been using riser cables. I use that fan on my main rig. Those things move some air for cheap!

 

[Madpacket]

Over at https://forum.ethereum.org/discussion/5594/rig-porn-post-your-pics/p5, someone posted some details about his 96 nano mining farm. Each nano running at ~70 W and generating 22 Mh/sec. Pretty cool: https://15254b2dcaab7f5478ab-24461f....vanillaforums.com/editor/ls/vzfdoc04k4og.jpg

link to his details on bitcointalk: https://bitcointalk.org/index.php?topic=1424132.0

That's one way to spend around 75K. Wow!

One of his Nano rigs would slightly outperform my 390 build but uses roughly half the energy. Too bad Nano's cost so much here in Canada compared to 390's. Even if Polaris is amazing the Nano rig will still likely be useful to mine on after release.

 

[Madpacket]

Silly, really. All the things that made Ethereum interesting from the start are still in play. Bitcoin is still having its blockchain problems etc etc. Let them panic sell. If it brings coins back down to $1 apiece, then it becomes easier to buy them up.



Nice. You using anything as spacers to keep the cards from leaning on one another? And did you have to run any ground wires?

edit: oh I see, you have the mobo down in there too. Thought you might have been using riser cables. I use that fan on my main rig. Those things move some air for cheap!

I'm using riser cables. The cards are simply zip tied in place, no spacers. I guess I could come up with them but they only move a little bit. I took an old bar from a dumbbell set and rammed it through the crate holes. It acts as a support platform for the bottom end of the cards. No grounding cable. Just zip tied the motherboard to the bottom of the crate but I put some spacers so the board isn't directly touching the plastic. I also ripped out the power and reset plus leds from an old case use that to power on the miner.

Kind of a funny story. It was really late last night and I accidently put one of the 1X PCIe risers in backwards. Thought I had everything set right and when I powered on the unit I saw instant smoke. Thought it was the power supply at first but it was smoke coming off the riser and 16x port closest to the power supply. Fried the riser and melted the port lol.

Anyway luckily this board has 6 PCIe ports and I had one riser left so I rewired it to a working port and everything works fine phew. The smell of burnt electronics still emanates from the board though hehe. Got lucky. Those cheap fans work really well. I have it set to the lowest speed and it keeps the cards very cool. Once I move the miner to the garage that will help temps as well.

Edit:

I was able to build the the unit without cutting or hacking any of the plastic on the milk crate.

 

[MrTeal]

I was able to build the the unit without cutting or hacking any of the plastic on the milk crate.

That's important, since you don't want to mess with the resale value.

 

[DrMrLordX]

I'm using riser cables. The cards are simply zip tied in place, no spacers. I guess I could come up with them but they only move a little bit. I took an old bar from a dumbbell set and rammed it through the crate holes. It acts as a support platform for the bottom end of the cards. No grounding cable. Just zip tied the motherboard to the bottom of the crate but I put some spacers so the board isn't directly touching the plastic. I also ripped out the power and reset plus leds from an old case use that to power on the miner.

Kind of a funny story. It was really late last night and I accidently put one of the 1X PCIe risers in backwards. Thought I had everything set right and when I powered on the unit I saw instant smoke. Thought it was the power supply at first but it was smoke coming off the riser and 16x port closest to the power supply. Fried the riser and melted the port lol.

Anyway luckily this board has 6 PCIe ports and I had one riser left so I rewired it to a working port and everything works fine phew. The smell of burnt electronics still emanates from the board though hehe. Got lucky. Those cheap fans work really well. I have it set to the lowest speed and it keeps the cards very cool. Once I move the miner to the garage that will help temps as well.

Edit:

I was able to build the the unit without cutting or hacking any of the plastic on the milk crate.

Nice work. Hmm I had been wondering how people powered on their systems minus power/reset buttons and all that. Some mobos have power buttons on there (the A88x-Pro does, after a fashion).

 

[Wickedcool]

Nice work. Hmm I had been wondering how people powered on their systems minus power/reset buttons and all that. Some mobos have power buttons on there (the A88x-Pro does, after a fashion).

You use a flat head screw driver or a car key to jump the two pins on the MB. Since you start the mining PC once, and never turn it off... You don't need anything else other than that...

 

[Wickedcool]

I've now read through 22 pages of Ethereum mining info. Good stuff.

I do want to add in that one of my 5x270X 2GB rigs stopped mining the other afternoon. I got the dreaded DAG file (-38) error. I worked for hours on it trying all the setx commands and everything I could find on the internet about it... I finally decided to try removing one card, and the other 4 started immediately mining. So that got me thinking that adding 2GB more system memory might fix it and so I bumped from 8GB to 10GB and voila, all 5 cards back to mining.

As the DAG file gets bigger, and depending on the number of cards you're running, everyone may need to bump up their system ram in order to get their 2GB cards to continue to run.

 

[EightySix Four]

Quite the ETH-USD roller coaster right now. Will be interesting how this plays out over the next few months, a $12 -> $8 drop will certainly impact mining profitability as mining effort goes up.

 

[DrMrLordX]

Quite the ETH-USD roller coaster right now. Will be interesting how this plays out over the next few months, a $12 -> $8 drop will certainly impact mining profitability as mining effort goes up.

I've run a few numbers, and it looks to me that people in lower-price power areas will be able to mine profitably all the way down to an eth price of ~$5. After that things get ugly. But if it goes that low, it becomes more sensible to buy eth directly, provided you aren't panic-selling yourself.

 

[DrMrLordX]

Side question: has anyone here with a Fury X/Fury/Nano tried overclocking the memory and seeing what that does to hashrate?

 

[zagitta]

Side question: has anyone here with a Fury X/Fury/Nano tried overclocking the memory and seeing what that does to hashrate?

My fully unlocked fury didn't gain anything from overclocking the memory which makes sense because the OpenCL kernel is quite poorly written (low amount of wavefronts, lack of various small optimizations and a pretty terrible outer loop) and I'm suprised no one has bothered optimizing it yet like the fierce competition in writing the best bitcoin miners back in the days. This is also backed up by HWInfo reporting about 50% memory controller usage on fury while a 290 is reporting around 75%.
In conclusion Fiji is quite poorly utilized at the moment.

 

[RussianSensation]

My fully unlocked fury didn't gain anything from overclocking the memory which makes sense because the OpenCL kernel is quite poorly written (low amount of wavefronts, lack of various small optimizations and a pretty terrible outer loop) and I'm suprised no one has bothered optimizing it yet like the fierce competition in writing the best bitcoin miners back in the days. This is also backed up by HWInfo reporting about 50% memory controller usage on fury while a 290 is reporting around 75%.
In conclusion Fiji is quite poorly utilized at the moment.

I tested 390 at 1400/1500/1600mhz memory and I get no improvement at all. I haven't tried going above 1600mhz as I have Elpida memory. That's 410GB/sec vs. 290X's 320GB/sec but even in games the impact is almost none. I am kinda stumbled why there was so much hype behind Hynix based Hawaii cards.

Computerbase confirms - there is barely a 4% difference in gaming performance between a 320GB and 390GB/sec 290X:
http://www.computerbase.de/2013-10/amd-radeon-r9-290x-test/15/

I am starting to think the Hynix vs. Elpida arguments for the last 2.5 years was just a giant synthetic benchmark e-peen contest. Anyone have data that shows otherwise?

My 7970 hit 1750mhz @ 336GB/sec but again just looks good on paper. These cards aren't fast enough to benefit from that much bandwidth.

 

[EightySix Four]

I tested 390 at 1400/1500/1600mhz memory and I get no improvement at all. I haven't tried going above 1600mhz as I have Elpida memory. That's 410GB/sec vs. 290X's 320GB/sec but even in games the impact is almost none. I am kinda stumbled why there was so much hype behind Hynix based Hawaii cards.

Computerbase confirms - there is barely a 4% difference in gaming performance between a 320GB and 390GB/sec 290X:
http://www.computerbase.de/2013-10/amd-radeon-r9-290x-test/15/

I am starting to think the Hynix vs. Elpida arguments for the last 2.5 years was just a giant synthetic benchmark e-peen contest. Anyone have data that shows otherwise?

My 7970 hit 1750mhz @ 336GB/sec but again just looks good on paper. These cards aren't fast enough to benefit from that much bandwidth.

I don't get any mining improvements from core overclocking but I do from VRAM overclocking.

 

[RussianSensation]

I don't get any mining improvements from core overclocking but I do from VRAM overclocking.

Ok I literally just went to my computer and tested it.

R9 390
1001 mhz = 25.3-28.2 MH/sec (mininum - maximum range)
1165 mhz = 30.7-33.8 MH/sec
1220 mhz = 33.5-36.5 MH/sec

The flag I am using is
"--cl-local-work 256 --cl-global-work 8192"

Changing memory speed from 1500mhz to 1600mhz has no material impact on performance at any of those GPU clocks.

Are you sure your R9 390/390X card's overclock is actually being applied?

 

[Madpacket]

Ok I literally just went to my computer and tested it.

R9 390
1001 mhz = 25.3-28.2 MH/sec (mininum - maximum range)
1165 mhz = 30.7-33.8 MH/sec
1220 mhz = 33.5-36.5 MH/sec

The flag I am using is
"--cl-local-work 256 --cl-global-work 8192"

Changing memory speed from 1500mhz to 1600mhz has no material impact on performance at any of those GPU clocks.

Are you sure your R9 390/390X card's overclock is actually being applied?

I know you have Elpedia but my Hynix cards perform in the same manner. Hynix vs Elpedia and higher memory speeds made a difference for Bitcoin though. I set all my memory speeds to 1275Mhz and get almost the same performance as setting up to 1600. The nice thing is that power consumption drops quite a bit allowing you to spend your electrical budget on GPU speeds.

Also IIRC 290's and likely 390's have "memory ramps". It would be nice to get a list of those ramps so we can adjust speeds accordingly. Need to look closer at Hawaii BIOS editor.

 

[Madpacket]

I don't get any mining improvements from core overclocking but I do from VRAM overclocking.

Interesting. What card do you have again?

 

[Madpacket]

My fully unlocked fury didn't gain anything from overclocking the memory which makes sense because the OpenCL kernel is quite poorly written (low amount of wavefronts, lack of various small optimizations and a pretty terrible outer loop) and I'm suprised no one has bothered optimizing it yet like the fierce competition in writing the best bitcoin miners back in the days. This is also backed up by HWInfo reporting about 50% memory controller usage on fury while a 290 is reporting around 75%.
In conclusion Fiji is quite poorly utilized at the moment.

Good info. Thanks for sharing. My Fury x's are only a few Mh faster than my 390's. I believe all GPU development was done on a 7850 so that makes sense. No way 4096 cores shouldn't be a good 30-40 percent faster. One nice side benefit of HBM is you can set it to 300Mhz and still achieve almost the same speeds as 545. Saves some energy too.

 

[Madpacket]

Side question: has anyone here with a Fury X/Fury/Nano tried overclocking the memory and seeing what that does to hashrate?

As stated in my last post. 300Mhz seems to be the optimal speed per watt. Clearly the DAG algorithm isn't optimized past Hawaii.

 

[Madpacket]

I've now read through 22 pages of Ethereum mining info. Good stuff.

I do want to add in that one of my 5x270X 2GB rigs stopped mining the other afternoon. I got the dreaded DAG file (-38) error. I worked for hours on it trying all the setx commands and everything I could find on the internet about it... I finally decided to try removing one card, and the other 4 started immediately mining. So that got me thinking that adding 2GB more system memory might fix it and so I bumped from 8GB to 10GB and voila, all 5 cards back to mining.

As the DAG file gets bigger, and depending on the number of cards you're running, everyone may need to bump up their system ram in order to get their 2GB cards to continue to run.

Great info thanks. What's interesting is I'm only using 4GB of RAM to mine with four 390's. I haven't run into any issues with DAG sizes so I wonder if in your setup part of the DAG needs to stored in system RAM due to the lack of VRAM?

 

[Madpacket]

That's important, since you don't want to mess with the resale value.

My intention was to Dremel the crap out of the crate but I didn't have the proper bits on hand. Oh and finding a milk crate in good condition wasn't that easy

 

[RussianSensation]

Good info. Thanks for sharing. My Fury x's are only a few Mh faster than my 390's. I believe all GPU development was done on a 7850 so that makes sense. No way 4096 cores shouldn't be a good 30-40 percent faster. One nice side benefit of HBM is you can set it to 300Mhz and still achieve almost the same speeds as 545. Saves some energy too.

It's a shame no one has been able to optimize the development over time for faster GPUs. As difficulty kept increasing with BTC, we upgraded from HD4800->5800/6900->7900, etc. but if no further optimizations happen, even if we get 5000-6000 shader Vega cards, it might hardly make an impact unless something changes in AMD's GCN 4.0 architecture.

As a point of reference for BTC:

VLIW-4 6970
1536 stream processors @ 996mhz = 3.056 Tflops = 422 Mhash/sec

GCN 1.0 7870
1280 stream processors @ 1165mhz = 2.982 Tflops = 423 Mhash/sec
https://en.bitcoin.it/wiki/Non-specialized_hardware_comparison

GCN became far more efficient for games but it hardly made an improvement for BTC mining. I fear a similar result for GCN 4.0 since I presume modern optimizations will be made towards graphics workloads. I mean some of these are not even related to hashing.

Geometry Processor
Multi-media cores
Display engine
and we both noted that memory speeds don't have much influence.

That leaves Command processor, L2 cache and GCN 4.0 as any remaining parts that could have a significant impact.

But then the increased IPC could mean AMD would limit Polaris 10 to just 2304-2560 stream processors, which wouldn't translate into a linear increase in hashing uplift as opposed to if they straight up replaced 2560 GCN 1.1 R9 390 cores with a new 4096 GCN 4.0 core part.

I am not too optimistic.