DiabloMiner GPU Miner

[Cdecker]

Just a quick test reveals that the performance drop is due to commit dc5b12ee55c0dd3971591cfb7e1420de06bb5e59 (ArtForz's Kernel). Too bad I was really looking forward to this one. Is it just on my side or does the kernel not work with ATI Stream SDK 2.1?

[FatherMcGruder]

I noticed a 15 Mhash/s boost with the latest binary on my Nvidia GTX460. Well done!

[DiabloD3]

Cdecker, it might just be you. I'm using 10.12 + 2.1, and its working fine here, and its also working for a few other 5xxx users in #bitcoin-dev. If your -f is set too low, you might be cutting your card off at the knees.

[DiabloD3]

Apparently 11.1 is a boned release. At least one person has come into #bitcoin-dev stating OpenCL performance is shittastic. Stick with 10.11 and 10.12.

[Syke]

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

My 5770 was pulling about 15 MH/s until I set "-w 64", then it goes back to the expected 150 MH/s.

[DiabloD3]

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

My 5770 was pulling about 15 MH/s until I set "-w 64", then it goes back to the expected 150 MH/s.

Thats the problem though. There should be no reason -w 256 (the default on ATI hardware) should be THAT slower. It is slower for OpenCL kernels like the one I use in the miner, but it should be still >100.

[gusti]

Hi :

Running Diablo successfully in Fedora 14, HD 5870, 260/290 Mhash/s with -w 64.
Any place I can find complete reference of tunning switches, sorry cannot not find them in wiki nor in docs.

Many Thanks
Gustavo

[DiabloD3]

Run the miner with --help. The only tuning switches are -w and -f, -w changes worksize (on ATI it has to be a multiple of 64, 64 seems to be the fastest, 256 is the default; on NVidia it has to be a multiple of 32, and all values seem to work exactly alike), -f changes how long the kernel runs (FPS, the lower the value the longer it runs, -f 1 has the least overhead but also destroys desktop interactivity).

[Cdecker]

Strange, with -f 10 it now gives me similar results to what I got before ArtForzz's kernel was introduced. But I have a standard deviation of ~10%! Still not perfect but I'm going to tweak my settings further

[DiabloD3]

Use below -f 5, my miner counts hash performance from start, instead of using a 30 second average. If you're seriously benchmarking it, use -f 1 and let it run for like 15 minutes.

[Cdecker]

Use below -f 5, my miner counts hash performance from start, instead of using a 30 second average. If you're seriously benchmarking it, use -f 1 and let it run for like 15 minutes.

Yeah, that's exactly my problem. It was fluctuating like hell but it stabilized over time right now I have a 5 hash/s standard deviation which is incredible. I had only a 5 minute warmup phase,a 5 minute hot phase and measured the average over the 5 minute hot phase and it didn't look nice. But with a longer warmup phase it looks better. Here's my monitoring graph for the last 24h.

Notice the high fluctuation and then the stabilization to a higher average

[DiabloD3]

Hmm, are you connecting locally or to slush's pool?

[Syke]

With dual cards, my hash/sec are all over the place. It never settles in to a steady rate. (running locally)

[DiabloD3]

With dual cards, my hash/sec are all over the place. It never settles in to a steady rate. (running locally)

As I said to cdecker, if you're benchmarking use -f 1

[Cdecker]

Ohoh, somehow this is worrying me:

Any idea what might be causing this? I'm getting it along with a few connection errors to slush's pool

[DiabloD3]

Connections to the miner when they lag (such as connection failures) confuses the miner. If thats a long count hash, thats completely inaccurate... it could have flatlined completely to 0.

[Syke]

Cdecker, how do you get the data for those graphs?

DiabloD3, "-f 1" definitely improves my hash-rate, but it still swings wildly. Typically between 300M and 350M, but also way below 300M from time to time. I'll try turning off crossfire and see if it improves.

[Syke]

Ok, it's not crossfire related. Even with a single card my hashes are all over the map, unlike my other system. It must be driver versions.

[gusti]

Using Fedora 14, "-f 1"  switch improved my hashrate more than 25% both on 5870 and 5970,
peaks are 350 and 610 Mhash respectively.  f1 is your friend ...   

[DiabloD3]

The miner, even with -f 1, tries to adjust to system activity. I use -f 1000 on my 4850 now, and it swings wildly in all sorts of directions when I'm using my computer.... this is by design. Most people want to mine while having a useful desktop. -f 1 curbs this behavior as much as possible, but obviously does not eliminate it.