DiabloMiner GPU Miner

[h_corey]

When using multiple GPUs on a machine.  What is the best way to turn off one and only allow diablo to run on the other?

Setup includes two gpus, nvidia 460 and 9600. I want to just use the 460 for mining and the 9600 for my displays.

Is it best if I try to use nvidia-smi to turn one off?  or is there a setting i can use to restrict diablo to just the 460 gpu.

The rest of the computer is AMD, Ubuntu linux 10.04

[1714194364]

1714194364

[1714194364]

1714194364

[DiabloD3]

My miner doesn't support this, and I don't think the Nvidia drivers do either. Try screwing with DISPLAY=:0.0 or :0.1

[h_corey]

What are the chances of adding a

--device n

option to the command line options?

[mizerydearia]

I installed ati-stream-sdk-bin-2.3

Code:

[1/17/11 12:32:14 AM] Started
Error: Compilation from LLVMIR binary to IL text failed!
[1/17/11 12:32:14 AM] ERROR: Failed to build program on ATI RV770 (#1)

Code:

# lspci
01:00.0 VGA compatible controller: ATI Technologies Inc RV770 [Radeon HD 4850]

This is my first time trying DiabloMiner, so I'm uncertain how to proceed.

For those wondering, miz is not using SDK 2.1, and 2.3 may or may not require Catalyst 10.12, and he has 10.11 installed.

Particularly, as Diablo-D3 pointed out, do not use the Gentoo ebuild and instead

untar the ati-stream-sdk tarball
take the icd-registration.tgz, untar that in / as root
then export LD_LIBRARY_PATH=/path/to/ati-stream-sdk-v2.1-lnx64/lib/x86_64/:$LD_LIBRARY_PATH
thats _all_ you have to do

Also, as per Diablo-D3's suggestion, ati-stream-sdk v2.1 is faster than v2.3 and he suggested to pass switch "-w 64" to DiabloMiner

[DiabloD3]

I installed ati-stream-sdk-bin-2.3

Code:

[1/17/11 12:32:14 AM] Started
Error: Compilation from LLVMIR binary to IL text failed!
[1/17/11 12:32:14 AM] ERROR: Failed to build program on ATI RV770 (#1)

Code:

# lspci
01:00.0 VGA compatible controller: ATI Technologies Inc RV770 [Radeon HD 4850]

This is my first time trying DiabloMiner, so I'm uncertain how to proceed.

For those wondering, miz is not using SDK 2.1, and 2.3 may or may not require Catalyst 10.12, and he has 10.11 installed.

[hacim]

I'm curious what different people have found to be optimized flags for their cards. It seems like the tunables are the work size (larger for faster cards), and the -f flag can be set to higher values to increase responsiveness on the desktop, lower to get more hash/sec. I've got a 5870 with the 2.3 stream SDK and 10.12 driver, and I can't seem to get more than 292 khash/sec using -w 256. I can't go up to 512, and setting -f lower doesn't seem to help.

What are others finding?

[Cdecker]

I automated the benchmarking as far as I could and this is what my ATI 5850 looks like:


Setup:

• Moving average over 5 minutes
• DiabloMiner with JMX patch to read Hashrate and Std Deviation

Not sure why it behaves like this, and the outliers on the left are not reproducible in the measurements (gnuplot just doesn't like 3D graph at their boundaries).

[DiabloD3]

Im having trouble reading that graph.... you're only using power of 2 values for -w right? 64 seems to be the fastest, 256 (== hardware max) is the default. This applies to 4xxx and 5xxx Radeons. Not sure what Nvidia does, though.

And yes, higher -f values slows it slightly due to kernel setup overhead. Doing -f 1 screws your desktop performance, but it gets about 5-15% faster over -f 1000 (maximum interactivity).

Also, timing is screwed on -f 1, it will often skew the timer badly for tasks that measure sub-second times (such as the miner itself, or performance benchmarks).

Oh, and btw, maximum performance for a 5850 exceeds 240 mhash/sec. Oh, and as another thing, don't bother benchmarking on 2.3. Use 2.1, its a fair bit faster for miners.

[sturle]

Im having trouble reading that graph.... you're only using power of 2 values for -w right? 64 seems to be the fastest, 256 (== hardware max) is the default. This applies to 4xxx and 5xxx Radeons. Not sure what Nvidia does, though.

IMHO you should change the default to 64 on ATI.

I have two computers mining.  One with one 5450 for mining only, and another ATI card for video.  The other computer has one 5970.  Both machines run Ubuntu (different versions) with the latest ATI driver.  This is the only thing they have in common.

Both experience ASIC hangs every now and then, and require a reboot to start working again when using DiabloMiner with default worksize.  On the machine with two cards, I get white stripes across the screen when mining on the other card.  This only happens when mining and using the computer for normal desktop work at the same time.  Nothing fancy.  When I set worksize to 64, the ASIC hang bug (probably a driver bug) goes away completely, and the problem with white stripes on the second card only show occasionally.

[DiabloD3]

64 being the most optimal may change. And some people with smaller chips report 128 being faster. Theres no easy way to handle this.

Also, 53/54xx Radeons may suffer from a driver bug. Many people report white stripes while running OpenCL apps, not just mine.

Also, if it periodically hangs, your card might be too hot. Make sure it remains under 85c.

[Cdecker]

Oh, and btw, maximum performance for a 5850 exceeds 240 mhash/sec. Oh, and as another thing, don't bother benchmarking on 2.3. Use 2.1, its a fair bit faster for miners.

That got me an additional 15-20 MHash/s, thanks for the tip. I'll redo my measurements and share them with you. Would someone else be interested in sharing their Cards benchmarks? I could create a driver that will collect all the measurements needed.

[DiabloD3]

See https://en.bitcoin.it/wiki/Mining_Hardware_Comparison

[Cdecker]

See https://en.bitcoin.it/wiki/Mining_Hardware_Comparison

Yes I know but I really love to see what happens when varying the parameters, I was surprised that with odd multiples of 32 for the worksize the hashrate dropped for example
It's not really for the sake of getting more out of the cards, as it is for curiosity

[FatherMcGruder]

See https://en.bitcoin.it/wiki/Mining_Hardware_Comparison

Apparently, one can get 68.31 Mhash/s out of an Nvidia GTX460. I'm running about 20 Mhash/s less. Do I just have to mess with the -f and -w modifiers to improve my hash rate or is there any logic or known parameters I can use?

[DiabloD3]

Yes I know but I really love to see what happens when varying the parameters, I was surprised that with odd multiples of 32 for the worksize the hashrate dropped for example
It's not really for the sake of getting more out of the cards, as it is for curiosity

The native hardware worksize (AMD calls these wavefronts) on Radeon 5xxx is 64. Anything less than 64, or anything that isn't a multiple of 64, will run very badly (as half the pipes aren't running anything half or all the time). AMD recommends power of two (subtly implying 192 isn't worth it either).

The OpenCL spec also recommends a power of two and it also implies that implementations should rotate around multiples of 64.

Apparently, one can get 68.31 Mhash/s out of an Nvidia GTX460. I'm running about 20 Mhash/s less. Do I just have to mess with the -f and -w modifiers to improve my hash rate or is there any logic or known parameters I can use?

I'm not sure what a correct value is. Some Nvidia hardware has a minimum worksize of 32 (due to it's CUDA legacy where 32 was acceptable), some seem to have one of 64. Nvidia hardware isn't nearly as parallel as AMD hardware (its 16 wide SIMD in 16 independent pipes and requires 2 batches to pipeline effectively; Radeon 5xxx is 64 wide SIMD in 4+1 units of 4 plus ALU instruction batching), so -w may not have as much power as it does on AMD hardware.

So, try a bunch of values and see what they do. Worst case is 64 runs just as well as 256 or 512 does.

Also, as I said earlier in this thread, higher -f values increase desktop interactivity but effectively slow hashing. -f 1 maxes out performance but makes your desktop useless. Nvidia drivers are badly designed enough that you could lose 20 mhash by not running at low -f values.

[humble]

I'm curious what different people have found to be optimized flags for their cards. It seems like the tunables are the work size (larger for faster cards), and the -f flag can be set to higher values to increase responsiveness on the desktop, lower to get more hash/sec. I've got a 5870 with the 2.3 stream SDK and 10.12 driver, and I can't seem to get more than 292 khash/sec using -w 256. I can't go up to 512, and setting -f lower doesn't seem to help.

What are others finding?

Have your overclocked the card? I was getting the same (identical setup) until I maxed out the overclocking. Now I'm getting slightly over 310.

[sturle]

Also, if it periodically hangs, your card might be too hot. Make sure it remains under 85c.

The cheap card is about 50°C.  The HD5970 runs with forced fan to about 70°C.  I've tried underclocking and running at even lower temperatures, and it hangs just as often.  -w 64 does not hang any of the cards.

[DiabloD3]

Update: Now using ArtForz's new kernel, about a 3% increase in speed. My 4850 does about 76.6 now.

Also, I fixed the slight bug where the hash meter would say absurdly high values because I accidentally queued kernels with a run length of 0.

[DiabloD3]

Update: I've added a mode to make a long count hash meter enable when using below -f 5, so no more 30 second average when using low -f values. Also, the run length of the kernel changes much more smoother and has a much lower minimum, so people that use higher -f values should get more responsive desktops; but people with low -f values may take longer to get up to optimum kernel length.

[Cdecker]

Wow, just pulled the latest commit, added my JMX patch and started the worker again and it maxes out at 14 Mhash/s on my 5850, whereas I got 254 Mhash/s before. Somewhere I got a problem ^^ I'll try to track it down next week