CUDA miner

Genoil

Hey,

I've been working on a CUDA miner. I've just got all bugs resolved. First hashrate measured on a GTX780 is a meagre 0.7MH/s, but it's a start! More updates here when I make improvements...

  D  17:47:23|main  #00004000
Benchmarking on platform: { "platform": "CUDA 7.0", "device": "GeForce GTX 780",
 "version": "Compute 3.5" }
Preparing DAG...
  D  17:47:24|main  pause took 0
  i  17:47:24|main  Spawning cudaminer0
  i  17:47:24|cudaminer0  Entering work loop...
  D  17:47:24|maiUns i nkgi cdkeOvfifc et:o oGke F0o.r0c0e2 GTX 780(3.5)

Warming up...
Trial 1... 611669
Trial 2... 699050
Trial 3... 699050
Trial 4... 699050
Trial 5... 611669
  i  17:47:46|main  Stopping cudaminer0
  i  17:47:46|cudaminer0  Finishing up worker thread...
  i  17:47:46|main  Terminating cudaminer0
min/mean/max: 611669/664097/699050 H/s
inner mean: 699050 H/s
Phoning home to find world ranking...
Error phoning home. ET is sad.
Druk op een toets om door te gaan. . . 

Michael_A

Well Done Genius... lol well done Geniol
Does it mean all Nvidia user need to update and upgrade their ethminer?

Genoil

No not as long as the OpenCL hashrate on NVidia is higher than my CUDA version. And then it's still built on a relatively ancient build of eth.

Genoil

After a bit of tweaking I'm now at 1.7MH/s, still a long way to go...

jzen

@Genoil Nearly a 2.5x performance improvement, way to go!

Genoil

@jzen yes but I'm still way behind the opencl performance. I think I'm missing an important clue somewhere, because ultimately, OpenCL on NVidia runs on CUDA too. I may actually try and apply the changes I made to the CUDA kernel to the OpenCL kernel and see how that affects the OpenCL kernel speed on NVidia.

Genoil

8.1MH/s now. The missing clue was to put the cuda compiler in release mode, which gave me a 4x speed bump. Only 3 short to match opencl performance.

jzen

@Genoil, Good goin! Where do think the 3 short are hiding?

Genoil

@jzen, not sure yet. Before I discovered the debug/release thing I had already improved the speed 3x from the first working kernel. So it's just a matter of more optimization, although at this point I really don't know where to look for any.

Brillopad12

@Genoil Looking awesome man! Keep up the great work

Genoil

10.8MH/s on CUDA vs 12.5MH/s on OpenCL now.

ConradJohnson

Bad ass @genoil just in time for me to switch back to the nvidia machine.

Genoil

I think this is it for now, can't get past 10.8, 11 on occasion. I have also tried porting some of the cuda optimizations over to the opencl kernel, but I'll leave that to somebody actually owning an AMD card, as debugging opencl is impossible on NVidia hardware.

I've attached a win64 release binary for those of you who want to test the difference. Might be that on different hardware, the speed difference is different from what I have on the 780. Other OS can grab and build the source from https://github.com/Genoil/cpp-ethereum/tree/cudaminer.

Use the -U flag instead of -G

Genoil

I've got 12.4MH/s on CUDA now vs. 12.4MH/s on OpenCL. almost done syncing the blockchain, so in a few hours I'm going to see if it's actually able to mine any blocks

PascalVerst

I'll test along with you, I am getting 17 600 000 H/s with GTX 970 on Windows 7.

lefteris

Very nice work Genoil. I will see what can be ported back upstream. I am also using Nvidia (GTX 770), but on Linux with OpenCL.

Genoil

@lefteris, thanks, I could really use some help setting up the CMake scripts. For now, for each build I have to manually set the VC++ build customisation in MSVC (which cmake doesn't support apparently), add the .cu and .cuh files and modify NVCC properties.

Before porting anything upstream, I really want to have mined a block on the testnet . During development, I made sure the CUDA miner kept reporting the same results as the opencl miner, but you never know

lefteris

I am curious to see if Cuda can be used also in Linux directly instead of as an OpenCL backend. Will have to try that. Not officically supported by us ofcourse, but still the Nvidia guys could use a bump in hashrate if that's done.

Genoil

For now I just have equal hashrate on CUDA vs OpenCL, so in a way it's quite pointless . I still have some optimization left to do, but it isn't going to do much, I think.

I'm now mining with CUDA on the blockchain, but with the current hashrate I don't think I'll be finding any block anytime soon.

Genoil

I can now mine a tiny bit faster than opencl @12.6MH/s. I'm down to replacing C with inline PTX assembly now, with some luck I should be able to push a bit harder...

jzen

@Genoil Did you mine one yet?

Genoil

@jzen I don't know. The only machine I can reliably mine on is behind a firewall, so I'm connecting to geth on EC2 with rpc. It's now the second time in a row I can't reach my EC2 node after having run it for a longer period. I hope I'll be able to recover it.

@PascalVerst The build I attached was for Kepler architecture (Compute 3.5). I'll post one that suport Maxwell better, I would expect it to do a little better.

Genoil

Got back into the EC2 node. 0 mined blocks, while statistically speaking, I should have mined about 10. But my balance is 1.6875 ether, while it was 0 when I started. Don't know where I got that from. Anyone?

https://explorer.etherapps.info/address/0x99b645a86bc157ec695f7db8dbe5751260c788ea

klao

Genoil said:

Got back into the EC2 node. 0 mined blocks, while statistically speaking, I should have mined about 10. But my balance is 1.6875 ether, while it was 0 when I started. Don't know where I got that from. Anyone?

1.7 ether looks like exactly one uncle block. (Meaning, that you have mined a block, but someone was a little bit faster. So, your block was not incorporated in the chain, but it was mentioned in the next block, and you've got a little reward for that.)

As for why you only got that and not more, I've no idea. But yeah, with that hashrate you should have got more...

Genoil

@klao thanks for clearing that up. I'm happy with the uncle block. I'll try again overnight.

Genoil

Seems I tried to over-optimise here and there. Reverted some things to how they were originally, and now I'm suddenly doing 14MH/s.

Genoil

Latest build for Windows. Requires a Compute 3.5 Device (GTX 780, Titan) minimally. Compute 2.0 and 3.0 shouldn't be an issue but doesn't compile after the asm optimizations.

Now does 17.8MH/s on the GTX780.

TheoCoyne

@Genoil I'm running a GTX 980. If you throw up a quick and dirty how to get started with your CUDA miner guide, I'd love to do some testing.

o0ragman0o

Benchmarking on platform: { "platform": "CUDA 7.0", "device": "GeForce GTX 750", "version": "Compute 5.0" }
Preparing DAG...
  i  1W9a:r0m2i:n3g3 |ucpu.d.a.m
iner0  workLoop 0 #00000000 #00000000
  i  19:02:33|cudaminer0  Initialising miner...
Using device: GeForce GTX 750(5.0)
Trial 1... 0
Trial 2... 0
Trial 3... 0
Trial 4... 0
Trial 5... 0

Not sure what's going on here with the benchmark. It also hangs before phoning home.

I'm running GTX750 TI on Win 7.
Normal mining seems to work but at only 1MH/s. Given that stats put the card about 1/3 as powerful as 780 I was thinking it would have been around 5MH/s.

i  19:01:49|main  Mining on PoWhash #1b119561 : 1096476 H/s = 13369344 hashes / 12.193 s

Genoil

@o0ragman0o thanks for reporting. It's not good, indeed. I have a 750 Ti in another system as well. Depending on the machine it is installed in, it performs very different. The other machine has a Celeron G1840 with 4GB, opencl hashing around 500KH/s. Once I took it out and installed it side by side with the 780 in a Xeon E5, it suddenly hashed 8Mh/s. After the weekend I'll upgrade the Celeron to an i5 and double the RAM. Then I can look into tuning performance for the 750Ti, because eventually that's a more interesting card for mining than a 780.

o0ragman0o

corrections:
GTX 750 OC (not TI, but no real difference)
Did phone home on benchmark after a while.

The cards power consumption is only 30% when mining so there's still a lot of legs in there somewhere.

  i  09:54:29|cudaminer0  workLoop 1 #c6320e3c #c6320e3c
  i  09:54:29|main  Mining on PoWhash #336a7d6f : 1048576 H/s = 524288 hashes / 0.5 s
  i  09:54:30|main  Got work package:
  i  09:54:31|main    Header-hash: c9b987cc71afeff02082550720d55ae121031cee0199e07a7ff564cfe3e2fdb9
  i  09:54:32|main    Seedhash: c6320e3c1c3456002aa6ccc5b60cf5bf054d7e9392712f8b7764869abe630035
  i  09:54:34|main    Target: 000000001ddf0fe4a8ee03c49bcf8b5057537dfcd8f569923dc16481c969c7b8
 !   09:54:39|main  pause 3.001 s

Also, there is a 3 second pause warning after every work loop. Is that intentional? Sorry I can't be more constructive, I think I'm going to have to knuckle into the material about how mining actually works...