DiabloMiner GPU Miner - page 29. | Bitcointalksearch.org

Joshwaa

hero member

Activity: 497

Merit: 500

No biggie I will stay on the old version till you get round to it. Would love to see you squeeze more out of GCN though.

DiabloD3

legendary

Activity: 1162

Merit: 1000

DiabloMiner author

Quote from: Joshwaa on February 16, 2012, 11:09:49 AM

Diablo what's up with GCN on the new miner. Do I need different flags only getting 30Mhash?

SDK bug that I've rammed into. Haven't quite figured out why its sucking, either.

Joshwaa

hero member

Activity: 497

Merit: 500

Diablo what's up with GCN on the new miner. Do I need different flags only getting 30Mhash?

DiabloD3

legendary

Activity: 1162

Merit: 1000

DiabloMiner author

Quote from: ShadesOfMarble on February 16, 2012, 02:51:59 AM

Quote from: DiabloD3 on February 15, 2012, 07:00:59 PM

But you shouldn't be using -v 4 -w 64. That is not appropriate for the new kernel on vliw5. For me, -v 2 -w 256 is fastest, -v 2,1 -w 256 is slower, -v 2 -w 128 isn't faster (but -w is very sensitive to memory timings, so -w 128 being faster for others does not surprise me).

What kernel do you mean by "the new kernel"? I'm using the zip-package linked in the first post of this thread and this is still the "old kernel", isn't it?! I did not compile DiabloMiner from source so I guess I'm still using your "not optimized for SDK 2.6"-version. Maybe we can clear that confusion first Wink

Quote from: DiabloD3 on February 15, 2012, 07:09:02 PM

Okay, I figured out how to do this on Windows. In the .bat file or in cmd.exe or whatever you use, before running DiabloMiner, do SET GPU_DUMP_DEVICE_KERNEL=3. It should cause it to dump the .isa files.

-v4 http://pastebin.de/23309
-v1,2 http://pastebin.de/23310
-v2 http://pastebin.de/23311

Nope, its the new kernel as of the post that you quoted originally that says "Update: Committed the new kernel."

Er, 1503 on -v 2? Dude, you're not using the newest copy of DM. 1503 is what the old kernel got on -v 2 on 2.6.

And it explains your weird results with -v 2,1... the hash mater was broken at the time of that version for non-power of two -v, I told you the mhash meter was fine because I fixed it in above referenced commit.

Joshwaa

hero member

Activity: 497

Merit: 500

The new binaries for windows he put up yesterday. 20MHash increase on the 6970's.

So yes new kernel is up give it a try, unless you have a 7970. Don't know what happened there.

ShadesOfMarble

donator

Activity: 543

Merit: 500

Quote from: DiabloD3 on February 15, 2012, 07:00:59 PM

But you shouldn't be using -v 4 -w 64. That is not appropriate for the new kernel on vliw5. For me, -v 2 -w 256 is fastest, -v 2,1 -w 256 is slower, -v 2 -w 128 isn't faster (but -w is very sensitive to memory timings, so -w 128 being faster for others does not surprise me).

What kernel do you mean by "the new kernel"? I'm using the zip-package linked in the first post of this thread and this is still the "old kernel", isn't it?! I did not compile DiabloMiner from source so I guess I'm still using your "not optimized for SDK 2.6"-version. Maybe we can clear that confusion first Wink

Quote from: DiabloD3 on February 15, 2012, 07:09:02 PM

Okay, I figured out how to do this on Windows. In the .bat file or in cmd.exe or whatever you use, before running DiabloMiner, do SET GPU_DUMP_DEVICE_KERNEL=3. It should cause it to dump the .isa files.

-v4 http://pastebin.de/23309
-v1,2 http://pastebin.de/23310
-v2 http://pastebin.de/23311

DiabloD3

legendary

Activity: 1162

Merit: 1000

DiabloMiner author

Okay, I figured out how to do this on Windows. In the .bat file or in cmd.exe or whatever you use, before running DiabloMiner, do SET GPU_DUMP_DEVICE_KERNEL=3. It should cause it to dump the .isa files.

DiabloD3

legendary

Activity: 1162

Merit: 1000

DiabloMiner author

Quote from: ShadesOfMarble on February 15, 2012, 06:49:44 PM

Quote from: DiabloD3 on February 15, 2012, 06:28:36 PM

Lets say you're getting similar numbers to mine. I get 393 or so on my 5850 at 960/320, 393/960*850 is 348. You have two of those, so 696. Now, your 5870, would be 393/960*950/1440*1600 = 432. You should be getting around 1128.

precisely... and that's about that what I got using SDK 2.5. Now with SDK 2.6 and using -v4 -w64 I get 1060 MH/s, which is a ~10% decrease as expected from the change SDK 2.5->2.6. Using layered vectors gives those ridiculous high number of 1370 MH/s, which just cannot be explained.

So if there is anything I can do/check, tell me... until then I can only ask you to double-check your meter code even if you are sure everything is correct - but something isn't ;o

But you shouldn't be using -v 4 -w 64. That is not appropriate for the new kernel on vliw5. For me, -v 2 -w 256 is fastest, -v 2,1 -w 256 is slower, -v 2 -w 128 isn't faster (but -w is very sensitive to memory timings, so -w 128 being faster for others does not surprise me).

ShadesOfMarble

donator

Activity: 543

Merit: 500

Quote from: DiabloD3 on February 15, 2012, 06:28:36 PM

Lets say you're getting similar numbers to mine. I get 393 or so on my 5850 at 960/320, 393/960*850 is 348. You have two of those, so 696. Now, your 5870, would be 393/960*950/1440*1600 = 432. You should be getting around 1128.

precisely... and that's about that what I got using SDK 2.5. Now with SDK 2.6 and using -v4 -w64 I get 1060 MH/s, which is a ~10% decrease as expected from the change SDK 2.5->2.6. Using layered vectors gives those ridiculous high number of 1370 MH/s, which just cannot be explained.

So if there is anything I can do/check, tell me... until then I can only ask you to double-check your meter code even if you are sure everything is correct - but something isn't ;o

DiabloD3

legendary

Activity: 1162

Merit: 1000

DiabloMiner author

Quote from: ShadesOfMarble on February 15, 2012, 06:22:46 PM

I'm on Windows 7 32-Bit... As you can read in post #1201 I also averaged over 3 hours (2500 shares) and I'm still getting the same results. And beside these "hard numbers"... don't you think it's very strange I can achieve 1370 MH/s with 2x5850@850/1000 + 1x5870@950/1200, SDK 2.6 and your old "SDK 2.1 miner"? It does not make any sense at all.

And again: If I'm not using layered vectors, miner mh/s and pool mh/s are matching very good. Also if I calculate the mh/s by hand using the submitted shares in a specific timeframe (like I did in post #1201), the numbers match. Only when using layered vectors the miner mh/s are WAY higher than everything else suggests (pool mh/s, hand-calculated mh/s, 24h earnings)

Lets say you're getting similar numbers to mine. I get 393 or so on my 5850 at 960/320, 393/960*850 is 348. You have two of those, so 696. Now, your 5870, would be 393/960*950/1440*1600 = 432. You should be getting around 1128.

So yeah, 1370 seems a smidgen high. But I don't believe its a fault of the meter, something else strange is going on.

ShadesOfMarble

donator

Activity: 543

Merit: 500

I'm on Windows 7 32-Bit... As you can read in post #1201 I also averaged over 3 hours (2500 shares) and I'm still getting the same results. And beside these "hard numbers"... don't you think it's very strange I can achieve 1370 MH/s with 2x5850@850/1000 + 1x5870@950/1200, SDK 2.6 and your old "SDK 2.1 miner"? It does not make any sense at all.

And again: If I'm not using layered vectors, miner mh/s and pool mh/s are matching very good. Also if I calculate the mh/s by hand using the submitted shares in a specific timeframe (like I did in post #1201), the numbers match. Only when using layered vectors the miner mh/s are WAY higher than everything else suggests (pool mh/s, hand-calculated mh/s, 24h earnings)

DiabloD3

legendary

Activity: 1162

Merit: 1000

DiabloMiner author

Quote from: ShadesOfMarble on February 15, 2012, 11:10:42 AM

Yes, I know the pool only "guesses" the hashrate, but btc guild is averaging over some time period (something like 10 minutes) and in the past miner mh/s and pool mh/s did not differ by more than 5%. While I was using -v2 the numbers where very close to each other, and now that I'm using -v2,1 (or -v1,2) the numbers differ by a large amount (1260 vs. 920 MH/s). I never saw such a big difference before... that's why I'm asking if you can say for sure (!) that the numbers displayed by your miner are 100% accurate.

I just checked... a 5850 with 850/1000 clocks, SDK 2.6 and the version of your miner linked in op gives 390 MH/s. Is that realistic given the fact you did not update the package in op and as such I'm using a miner not optimizied for SDK 2.6?! Using poclbm and SDK 2.5 I got something around 340 MH/s...

Using -w 256 is a bit slower than -w 128 (1230 vs. 1260 MH/s).

Edit: One more sign that something is wrong with the mh/s display: Let's say I have a nominal rate of 1000 mh/s. The pool sometimes shows less, sometimes more... like 950 mh/s and then 1050 mh/s - I saw that before (the 5% I was talking about). But now, using -v2,1, I *only* see less, but *never* more... And it's not "a bit less" but the pool always shows at least 20% less than the miner.

10 minutes isn't nearly enough to average that out unless you're producing 1000 shares in 10 minutes on average.

-v 2,1 and 1,2 are identical, the miner sorts them for size first. The numbers displayed by my miner are accurate, I just find it strange you're getting more on -v 2,1 and I'm not. Maybe there is a bug, but I'm not triggering it here.

I'm not sure if you said, but are you on Linux? If you do, do export GPU_DUMP_DEVICE_KERNEL=3 and then pastebin the .isa file that gets produced for -v 2, and then do it again for -v 2,1 (it overwrites it, so make sure you run, pastebin, run, and then pastebin). I can tell for sure if the numbers are legitimate from those.

ShadesOfMarble

donator

Activity: 543

Merit: 500

I'm now almost sure that the miner's mhash display is wrong when using -v1,2.

In 3 hours one rig created 2530 shares, which translates to 1006 MH/s. This is a statistical process, so we can assume an error of magnitude sqrt(2530) -> 50. So with very high propability, the hashrate is in a range of 986-1026 MH/s. The miner says 1370 MH/s which is nowhere near that value. 1000 MH/s is also much more realistic because I'm now using SDK 2.6 and a miner optimized for SDK 2.1...

As much as I would like to believe in those 1370 MH/s, no matter how I look at it, there is no support for this value.
I think you should really double-check the mhash display... If there is anything I can do, let me know.

macrum

newbie

Activity: 24

Merit: 0

Hello Diablo/Forum!

I installed a new machine with CentOS 6.2 and installed the Stream SDK and so on, as I always do.
Now when i try to start the miner i get this error:

Code:

# A fatal error has been detected by the Java Runtime Environment:
#
#  SIGSEGV (0xb) at pc=0x00007f5481ff84b6, pid=28094, tid=140000934340352
#
# JRE version: 6.0_22-b22
# Java VM: OpenJDK 64-Bit Server VM (20.0-b11 mixed mode linux-amd64 compressed oops)
# Derivative: IcedTea6 1.10.6
# Distribution: CentOS release 6.2 (Final), package rhel-1.43.1.10.6.el6_2-x86_64
# Problematic frame:
# C  [libstdc++.so.6+0x9d4b6]  std::basic_string, std::allocator >::~basic_string()+0x26

Any Ideas what the Problem could be? I am using the newest Binary of your miner.
I don't want to reinstall the machine....

A full copy of the log is available here: upload.robert-hager.at/uploads/temp/hs_err_pid28094.log (ignore the certifcate error)

Thanks in Advance Wink

ShadesOfMarble

donator

Activity: 543

Merit: 500

Yes, I know the pool only "guesses" the hashrate, but btc guild is averaging over some time period (something like 10 minutes) and in the past miner mh/s and pool mh/s did not differ by more than 5%. While I was using -v2 the numbers where very close to each other, and now that I'm using -v2,1 (or -v1,2) the numbers differ by a large amount (1260 vs. 920 MH/s). I never saw such a big difference before... that's why I'm asking if you can say for sure (!) that the numbers displayed by your miner are 100% accurate.

I just checked... a 5850 with 850/1000 clocks, SDK 2.6 and the version of your miner linked in op gives 390 MH/s. Is that realistic given the fact you did not update the package in op and as such I'm using a miner not optimizied for SDK 2.6?! Using poclbm and SDK 2.5 I got something around 340 MH/s...

Using -w 256 is a bit slower than -w 128 (1230 vs. 1260 MH/s).

Edit: One more sign that something is wrong with the mh/s display: Let's say I have a nominal rate of 1000 mh/s. The pool sometimes shows less, sometimes more... like 950 mh/s and then 1050 mh/s - I saw that before (the 5% I was talking about). But now, using -v2,1, I *only* see less, but *never* more... And it's not "a bit less" but the pool always shows at least 20% less than the miner.

DiabloD3

legendary

Activity: 1162

Merit: 1000

DiabloMiner author

Quote from: ShadesOfMarble on February 15, 2012, 09:33:08 AM

So the linked zip-package is still the same? I thought it was updated because you said "you can now upgrade to 2.6"... and that's what I did Wink

But maybe there is no 16% increase at all - are you sure the MH/s display in your miner is always correct? Taking the numbers from the console I should be around 2,6 GH/s - but BTC Guild says I just have about 2 GH/s.

In one PC I now have 2x5850 (@850/1000) and 1x5870 (@950/1200). The console says 1260 MH/s, BTC Guild says 920 MH/s... I have to say that 1260 GH/s is also somewhat unrealistic because that would be 420 MH/s per card... which looks ok for the 5870 but is too much for the 5850, isn't it?!

My commandline:
start "DiabloMiner" DiabloMiner-Windows.exe -u xxx -p bitcoin -o btcguild.com -r 8332 -f 2 -v 2,1 -w 128

Windows 7 32-bit, AMD driver 12.1

(when using -v2 the console says ~1000 MH/s and BTC Guild says ~950-1000 MH/s... so both numbers are almost matching)

Oh, and the SDK version is not displayed when starting your miner. Is this normal when 2.6 is installed? Before it showed "...using SDK 2.5", when 11.8 was installed.

Don't use pool hash meters, they're useless. They try to guess your hashrate by counting the number of shares you submit (which is a 100% random process) and multiplying by the average hashes it takes to make a share (2^32, or over 4 billion). I've had pools read as low as 200 some and as high as over 500, and I mine at what is now 393 mhash on my 5850.

It does still display SDK version, but 2.6 has a different version than the rest, it says 851.4 instead.

Also, try -w 256 instead. Its probably faster.

ShadesOfMarble

donator

Activity: 543

Merit: 500

So the linked zip-package is still the same? I thought it was updated because you said "you can now upgrade to 2.6"... and that's what I did Wink

But maybe there is no 16% increase at all - are you sure the MH/s display in your miner is always correct? Taking the numbers from the console I should be around 2,6 GH/s - but BTC Guild says I just have about 2 GH/s.

In one PC I now have 2x5850 (@850/1000) and 1x5870 (@950/1200). The console says 1260 MH/s, BTC Guild says 920 MH/s... I have to say that 1260 GH/s is also somewhat unrealistic because that would be 420 MH/s per card... which looks ok for the 5870 but is too much for the 5850, isn't it?!

My commandline:
start "DiabloMiner" DiabloMiner-Windows.exe -u xxx -p bitcoin -o btcguild.com -r 8332 -f 2 -v 2,1 -w 128

Windows 7 32-bit, AMD driver 12.1

(when using -v2 the console says ~1000 MH/s and BTC Guild says ~950-1000 MH/s... so both numbers are almost matching)

Oh, and the SDK version is not displayed when starting your miner. Is this normal when 2.6 is installed? Before it showed "...using SDK 2.5", when 11.8 was installed.

DiabloD3

legendary

Activity: 1162

Merit: 1000

DiabloMiner author

Quote from: ShadesOfMarble on February 15, 2012, 07:30:58 AM

Quote from: DiabloD3 on February 14, 2012, 11:53:03 PM

Update: Committed the new kernel.

-v 1: 803/2.1 and 883/2.6 -> 738/2.6
-v 2: 1362/2.1 and 1503/2.6 -> 1380/2.6

398 peak hashrate on my 5850 drops to only 394, or 1%. I think I've won. BTW, 2.1 performance is now dead. Everyone, you can now upgrade to 2.6.

Using -v1,2 I get a 16% boost compared to -v2. -v3 and -v4 are even slower than -v2. (Driver 12.1, 5850/5870.)
Is this expected behaviour? In the first post you recommend -v4 and say "Layered vectors, probably not faster"... but my findings are just the opposite.

You seem to have a slight misunderstanding: I haven't updated the op for the new kernel. Also, no where in op do I recommend -v 4, I say "try the above and then try -v 4 -w 64", where the implication the above (-v 2 and -w 128 or 256) is still going to win.

-v 2,1 isn't unexpected. While developing the new kernel -v 2,1 was actually winning for awhile, and now its not for me. The interesting thing is -v 2,1 actually fits in all the registers now, something that has never been done before. So, on your cards, what clock rate are they at now and what mhash are you getting?

ShadesOfMarble

donator

Activity: 543

Merit: 500

Quote from: DiabloD3 on February 14, 2012, 11:53:03 PM

Update: Committed the new kernel.

-v 1: 803/2.1 and 883/2.6 -> 738/2.6
-v 2: 1362/2.1 and 1503/2.6 -> 1380/2.6

398 peak hashrate on my 5850 drops to only 394, or 1%. I think I've won. BTW, 2.1 performance is now dead. Everyone, you can now upgrade to 2.6.

Using -v1,2 I get a 16% boost compared to -v2. -v3 and -v4 are even slower than -v2. (Driver 12.1, 5850/5870.)
Is this expected behaviour? In the first post you recommend -v4 and say "Layered vectors, probably not faster"... but my findings are just the opposite.

DiabloD3

legendary

Activity: 1162

Merit: 1000

DiabloMiner author

It seems the drivers currently have a nice wonderful bug. GCN results in about 30 mhash.

Thanks AMD.

Topic: DiabloMiner GPU Miner - page 29. (Read 866596 times)