new record?
Unfortunately not but that is a damn good hash rate but then 1055 MHz is a very high clock. I feel that it should be possible to get more MH/s at that clock rate but is your card stable at that rate? How long did you run the miner for? I decided to overvolt my good 5850 from stock (1.0875V) to 1.25V and see what it could do. It managed to hold 1110 MHz under load for 3 hours (finally dying when I got adventurous and tried to start mining on my other card) and managed 461.1 MH/s with very few stales. I posted about this here. You can get a screenshot here and I have pictures of my rig here. Admittedly, I have the advantages of Linux without any GUI, very good aftermarket coolers, and lowish ambient temperatures where I am (peaking at 21*C in the middle of the day). I was proud to have beaten the highest claimed 5870 record but it wasn't long until scifimike12 took his water cooled 5870s to an astonishing 501 MH/s (which ran for an hour before he turned it down). Needless to say, I do not run at 461.1 MH/s all the time. I'm not very concerned about the voltage or heat and I'm not convinced it's not profitable but there was simply far to much noise. 428 MH/s at stock volts and 50% fans is ok for now and I'm going to undervolt my cards and turn down the fans once my power meter finally arrives. |
|
|
|
|
Did you decide that the claimed rates of 461 MH/s for a 5850 and 501 MH/s for a 5870 don't qualify for inclusion in your list or are you just unaware of them? If the former, what are the qualifying conditions?
|
|
|
|
69*C and 51*C for my two 5850s using Zalman VF3000A coolers with fans at 60% (they are running at 975MHz and 1025MHz respectively). The hot card is directly behind the cooler one. See http://forum.bitcoin.org/index.php?topic=7216.msg345510#msg345510 for pictures. I hope to undervolt my cards soon which should make them cooler and allow me to turn the fans down. |
|
|
|
With the ideas that Vince gave to us, I was able to lower the ALU OP usage even further. This means next version will speed up things for 69XX and 58XX again  . Thank you Vince, I didn't need all your changes (some seem to reduce kernel speed, even if they look good), but merged the ones I like and verified everything with KernelAnalyzer. Edit: Drawback is, that you will need to replace the Phoenix __init__.py file, so it won't be easy usable for non Phoenix users, sorry for that (some init values changed)! Dia You say that some of Vince's changes seem to reduce kernel speed but it looks like actual speed gain/loss is very much card dependent. That being the case, which cards are you using for testing? |
|
|
|
I currently solo mine but would be interested in joining this. Unfortunately I see that the payout method is proportional which is a deal breaker for me. Any plans to implement something fairer (like score based shares or shares from last 5 minutes only)?
I assume it's a deal breaker because of pool hopping? I don't have any specific plans, but I have been thinking about the different payout methods. I realize it will be tough to prevent pool hopping with proportional payouts. Delaying stats may work for DeepBit, but it would not work as well for a smaller pool (or stats would have to be delayed much longer). I think paying for last N minutes is vulnerable to pool hopping too. The more shares were submitted the last N minutes, the less they are worth. And the higher the pool hash rate the better the chance of creating a block. If many shares were submitted and hash rate then dropped: time to pool hop. You don't want to work alone and then give all the money away. If few shares were submitted compared to the current hash rate: time to come back again. Someone else is likely to find a block and you want to get paid. Maybe "pay per last N shares" is the most fair. If pool hash rate drops there is less chance someone else finds a block that you get paid for. But also you will take over more and more of the N shares at an increased rate, "pushing out" the shares of other users. Interesting charts on different payout methods: http://eligius.st/~luke-jr/samples/800MH/I'm open to suggestions. You're right, there does seem to be a different kind of hopping problem with PPLNminutes but at least it is very much to the pool's benefit (I guess pool hoppers would try to keep the mining volume constant). I'm new to pool mining theory and PPLNS is a new discovery for me; it looks better than PPLNM but would be interested to know more about the weaknesses of these payout methods and why the more obviously faulty proportional is preferred. Also, it strikes me that without any information about the pool at all it is still quite possible to pool hop using the block chain. Simply mine with the pool for 5 minutes after each block discovery and mine solo the rest of the time. You are only making money on the blocks which the pool itself finds but you're doing better than 100% nonetheless. This would be particularly effective against a large pool and the only way I could see them blocking it is penalising you in some way for disconnecting from the pool, or detecting your behaviour and banning you. Both methods defeat the only advantage to proportional that I see: it's simplicity. Are there other advantages to proportional aside from simplicity (versus PPLNS)? I feel there is something very wrong if this is the only advantage to proportional payouts. Another payout method could be to pay an equal chunk to each user reguardless of their hash rate and this would have the advantage of being even more simple than proportional but very little else going for it. Just as with proportional, people would take advantage, others would complain, and pool operators would be trying to block/ban people who took advantage and/or hide information to make their pool more attractive to 'honest' miners, all for some algorithmic simplicity. Sorry, I went a bit far there. Which pools do you know which do PPLNS? To me this is more attractive than +5% to be honest. |
|
|
|
the 11.6 drivers will let you use aticonfig to set clocks outside the normal config range.
I guess I don't understand how. I've used 11.6 plenty, but I've never had success getting passed the 'normal' range. Is there a list of commands? I use things like "aticonfig --od-enable" "aticonfig --odsc=GPU,mem" "aticonfig --odgc" and so on, but I don't seem to be able to get that to work. I guess it's possible that Catalyst 11.6 only breaks BIOS limits for older cards. With my two non-reference 5850s Catalyst 11.5 limits me to 900 MHz core clock and 1.0875V. With Catalyst 11.6 I can go to at least 1140 core clock (not pushed it further) and to a new limit of 1.3V. I can use 'aticonfig --odsc' to push the core clocks past the BIOS limit and didn't have to do anything special (used AMDOverdriveCtrl for the voltage). Perhaps it just doesn't work with the HD6xxx cards. |
|
|
|
|
I currently solo mine but would be interested in joining this. Unfortunately I see that the payout method is proportional which is a deal breaker for me. Any plans to implement something fairer (like score based shares or shares from last 5 minutes only)?
|
|
|
|
|
One should also be aware that, given sufficient data, it is usually easy to find any pattern you are looking for.
|
|
|
|
Broke 500MH/s.  Congrats! A significant milestone! You also beat my highest stable core clock: 1110 MHz (I say stable but in reality I only ran it for 3 hours so who knows). I see that you took your screenshot shortly after starting the miner. How long did you run it for at that speed? Perhaps you are still running it at that speed. I did get my 5850s core clock up to 1140 MHz (473.4 MH/s) in testing but it only lasted for 30 seconds. I'd love to try and compete but with a 5850 and air cooling I'm a little out of my depth. |
|
|
|
|
This does work with SDK 2.1 but it might be a tiny bit slower than your previous version.
HD5850, 1.0875V, 975MHz clock, 360 MHz RAM, aggression=14, worksize=256, Catalyst 11.6 (Linux)
SDK 2.1: 404.6 MH/s -> 404.5 MH/s
SDK 2.4: 401.8 MH/s -> 402.2 MH/s
Note that at aggression=14 my rate can sometimes drop as much as 1 MH/s suddenly before recovering but usually varies by 0.2 MH/s so the apparent decrease with SDK 2.1 could well be statistical noise.
I might also have to play with the RAM frequency again.
|
|
|
|
I'm sure you've seen it but this topic has some very good 5870 rates. The highest I've seen there is 490 MH/s. |
|
|
|
|
Excellent, 490 MH/s from a single GPU is the highest I've seen on this forum by a margin.
Also, thank you rethaw for the highly informative post.
Honestly, the only reason I don't overvolt all the time is due to noise. Profit/Loss are card longevity are practically irrelevant compared to my quality of life.
|
|
|
|
If all goes well you should be able to comfortably climb past 480 MH/s, maybe touch 490 MH/s. You guys are nuts to try running a 5850 or 5870 at nearly 500mhash/s. Those cards cost $150 to 200 bucks at the least & 1.25-1.3 voltages will 100% surely kill the card due to electromigration or VRAM failure within a month or two. It'd last years if you ran it at 420-430mhash and the monetary 'loss' is negligible considering you don't have to buy a new GPU Also the power consumption will skyrocket if you go past 1.25v Don't worry, I'm sure we're all aware of the various issues with overclocking/overvolting. This thread is more about finding and establishing the limits of the cards for fun and to satisfy curiosity. Furthermore, what we learn can be applied to design very efficient setups. My occasional forays into the world of high voltage have all been experimental and my discoveries have allowed me to reach a stable 430 MH/s on stock voltage (5850). Also I've noticed you like to make the assumption that miners are primarily interested in profit but perhaps this is not a safe assumption of the people using a thread with 'voltage mod' in the title. I expect there are some passionate overclockers on this board that are taking great joy in the sudden interest Bitcoin is drawing to overclocking. |
|
|
|
Cant be donde whiout increasing voltage, voltages cant be increased beyond bios limit in Linux, AMDOverdriveCtrl only allow voltages changes in bios limit (so nothing, as the limit is the stock voltage). I just noticed that voltage adjustment with Sapphire 5850 Xtreme works. However, power consumption rises about 80 W/h, even with slightest changes, e.g. 1,088 V -> 1,094 V. Two of these cards are in my PC. Temperatures rise over 80 °C, too. I use Catalyst 11.6 and SDK 2.4 with phoenix. I could be wrong but I believe that the VRMs on the 5850s only handle voltage in increments of 12.5mV so the next level above 1.088 (really 1.0875V) would be 1.1V. I say this because 1.094V is pretty much half-way between 1.0875V and 1.1V so I'm wondering if you know more than I do (this information is hard to find). It's possible that the driver is rounding 1.094V up to 1.1V. [rant]Also, at least for my card, Catalyst 11.6 doesn't remove the limit so much as it raises it. I can't overvolt past 1.3V. Perhaps the VRMs on my card don't know how to do anything with more than 1.3V but, given everything we've seen up until now, I'd wager it's yet another software limit. Does anyone know of a card manufacturer that does not believe in software limits? It might be nice to have a card which I can order to "fall on it's sword" so to speak. I'm tired of using safety scissors and tippy cups. I don't mind limits which force the card to shut down in the event of a surge or the like but at least let me change those limits in software where possible.[/rant] |
|
|
|
I've not seen anyone claim a hash rate that high on a single GPU before but people do occasionally get higher core clocks. Last night I took my 5850 (non-reference, Sapphire 5850 Xtreme) to 1.25 V and managed to get 461.1 MH/s out of it for 3 hours. The card crashed immediately when I started up a second miner on another card in the same system. - 1110 MHz core clock
- 370 MHz RAM clock
- 55*C GPU temperature
This is on Linux with Catalyst 11.6, SDK 2.1, and phoenix (latest phatk kernel). Here's a screenshot of me about an hour into the pooled mining test. This same card went to 1140 MHz before crashing and is perfectly stable at 1030 MHz at stock voltage so you might put this down to luck. I only have experience with 5850s but I'm hoping they are similar enough to 5870s that some of the following can be of help. I'm assuming you're using the very latest phatk kernel; check out http://forum.bitcoin.org/index.php?topic=25860.100 if not. I found that SDK 2.1 is a few MH/s faster than SDK 2.4 on my card so I'm guessing you'll see a similar improvement with SDK 2.1 (yes, that's with phatk). Have you tried increasing your RAM speed a little. I know 300 MHz known to be good but I've found that as core clock rises, the maximal RAM speed rises. You may well find another 3 MH/s or so by bumping up to 355-360 MHz. Another idea I have is to use Catalyst 11.4. On my card this gives a straight up 9 MH/s boost to my 5850 clocked at 900 MHz. Over 1GHz I'd expect 10 MH/s or more. Windows users have claimed 12-13 MH/s over me by using Catalyst 11.4 instead of Catalyst 11.6 (same card, clock, RAM, kernel, SDK, miner settings). Unfortuately, no-one seems to know of a way of overclocking a card past BIOS limits using Catalyst 11.4 in Linux but it seems like you've solved this by flashing your BIOS to increase the limits. May I ask how much variance you get with your hash rate? You may need to look at your command line options (for phoenix '-a 1' - for poclbm '-e 1' maybe, I'm not sure). If your MH/s is varying by more than 1 MH/s then you can surely manage further gains through configuration, especially if you are using Linux. I'm guessing that this is not a dedicated miner given that you are using the '-f 15' option but if it is and it's running Linux then PM me for some further tips. If all goes well you should be able to comfortably climb past 480 MH/s, maybe touch 490 MH/s. |
|
|
|
The system really does lock properly - Ctrl+Alt+F2 has no effect. Removing the line in your 'bare-bones' xorg.conf is the difference between hanging and not, so it really is that second card. Upon further investigation, I think that second card has died at some point during this process - if I only have that physical card in slot 1 present, the system refuses to POST. It boots fine with just the first card in slot one. I might tinker a little more later, but I suspect it's RMA time. Ah, a faulty card! If only my knowledge of Linux were sufficient to help you fix that; alack. |
|
|
|
I decided to see how far I could push my good 5850 (a Sapphire HD5850 Xtreme with a Zalman ZF3000A after-market fan cooler, see http://forum.bitcoin.org/index.php?topic=7216.msg345510#msg345510). I think I've gotten it half-way stable (it's been running for 3 hours now with 0.5% rejects). I've focused a second rather powerful fan on the VRMs for this run, I can't read the temperature of the VRMs in software. 1.25V - 1110 MHz core - 370MHz RAM - 55*C - 461.1 MH/s. This is with Linux, Catalyst 11.6, SDK 2.1, phatk, aggression=14, worksize=256. I've got a screenshot here http://img195.imageshack.us/img195/339/72screenshot2.png. |
|
|
|
I foolishly ordered a pair of Zalman ZF3000A coolers for my non-reference 5850 cards; they didn't really fit but I decided to mount them anyway. The system was mounted on a wooden board courtesy of BlightyCoin http://www.blightycoin.com/.     |
|
|
|
|
I don't know so can only provide some data points:
I'm running Linux. I've never been able to get the CPU to show more than 2% from the mining software, no matter what version of Catalyst or SDK, which miners, kernels, or settings.
A friend of mine is mining in Windows and about 50% of his core i3 processor is in use while mining with a single 6990. He's not tried to reduce the CPU usage.
|
|
|
|
Sorry, you may want to largely ignore my last post for now. I was just reading over your inital post and you claimed that you could start X up with your own xorg.conf. Please try this first before trying to purge the system of fglrx. If you can get to a stage where you are able to run and have X running in Linux without crashes then we're good and can work from there. Probably the most basic xorg.conf which should initialise both cards is: Section "Device"
Identifier "Card 0"
Driver "fglrx"
BusID "PCI:1:0:0"
EndSection
Section "Device"
Identifier "Card 1"
Driver "fglrx"
BusID "PCI:2:0:0"
EndSection
Section "Monitor"
Identifier "Dummy Monitor"
EndSection
Section "Screen"
Identifier "Screen 0"
Device "Card 0"
Monitor "Dummy Monitor"
EndSection
Section "Screen"
Identifier "Screen 1"
Device "Card 1"
Monitor "Dummy Monitor"
EndSection
Section "ServerLayout"
Identifier "Server Layout"
Screen "Screen 0"
Screen "Screen 1"
EndSection
Try to isolate a small simple change to xorg.conf which will cause Linux to crash and look at the log if possible. Make sure Linux really is crashing if possible. Connect a keyboard and monitor if you have them and run "xinit &" on the machine using the console. If Ctrl+Alt+F2 does not take you to a second terminal (enabled by default in Ubuntu) then things really are bad. If Ctrl+Alt+F2 does give you a responsive console then login and look at "top" and "ps aux" for clues. Try killing all "xinit" and see if responsiveness returns to the first terminal (Ctrl+Alt+F1). |
|
|
|
|
Show Posts
