Donations to open source a GPU Protoshares miner (PTS)

[dga]

dga, do you have any idea why GTX680 card is not faster than 580? And there is any improvements of sm3.5 over sm3.0? In my test 3.5 card (GTX780) loses 20% cpm when using sm3.5 over sm3.0.
Thanks in advance if anyone could explain it for me.

Because the biggest bottleneck with my current design is memory latency and the number of memory references that can be in flight at a time.  This isn't something that improves nearly as fast with different generations of cards (and sometimes gets worse).  Though I'm surprised at your 780 slowing down with sm_35.  I'll have to try that.

Lesee.  K20c I get 1104 c/m with sm_35.  Results are nearly unchanged with sm_30.

The biggest crypto-related improvement in sm_35 is the funnel shifter, but the design of Momentum is less crypto-intensive than a lot of PoW functions because it generates 8 memory references for every one execution of SHA512.  This is in part why I don't expect there to be as big an nvidia-AMD gap (except that by being so much faster, the AMD crew can try to exploit space/time tradeoffs in different ways.).

I'm playing around with some very different design options for faster cards, but haven't decided on anything I really like yet.

Someone over 1gh pool did something similar like you, dga.
https://bitsharestalk.org/index.php?topic=1784.0

I just gave it a short test, and so far I can say that this GPU miner has some similar performance with your miner ( around 10% more depends on cards but it uses at least 2 threads of CPU with 100% load). Temperature of cards stay the same but more memory is required (1.2GB).

Since he doesn't public his code I can say for sure but it maybe use your code. And there some reports that it works with AMD cards too.

PTS mining is exploding in a few days.

Well, let's not let the AMD folks win, shall we?

Obviously, they can keep cloning my software, but let's let the folks using the open version have some fun for a while longer.

I've committed another optimization to the repository.  My GTX690 is now getting about 1850 c/m - not a huge boost, but a little.

The more important thing about this commit is that it removes the dependency upon yasm, which I know was causing some folks a headache trying to compile it on Windows.  Let's see if this makes it easier to use the open source version instead of a who-knows-what-you're-getting closed clone. :-)

  -Dave

[AizenSou]

dga, do you have any idea why GTX680 card is not faster than 580? And there is any improvements of sm3.5 over sm3.0? In my test 3.5 card (GTX780) loses 20% cpm when using sm3.5 over sm3.0.
Thanks in advance if anyone could explain it for me.

Because the biggest bottleneck with my current design is memory latency and the number of memory references that can be in flight at a time.  This isn't something that improves nearly as fast with different generations of cards (and sometimes gets worse).  Though I'm surprised at your 780 slowing down with sm_35.  I'll have to try that.

Lesee.  K20c I get 1104 c/m with sm_35.  Results are nearly unchanged with sm_30.

The biggest crypto-related improvement in sm_35 is the funnel shifter, but the design of Momentum is less crypto-intensive than a lot of PoW functions because it generates 8 memory references for every one execution of SHA512.  This is in part why I don't expect there to be as big an nvidia-AMD gap (except that by being so much faster, the AMD crew can try to exploit space/time tradeoffs in different ways.).

I'm playing around with some very different design options for faster cards, but haven't decided on anything I really like yet.

Someone over 1gh pool did something similar like you, dga.
https://bitsharestalk.org/index.php?topic=1784.0

I just gave it a short test, and so far I can say that this GPU miner has some similar performance with your miner ( around 10% more depends on cards but it uses at least 2 threads of CPU with 100% load). Temperature of cards stay the same but more memory is required (1.2GB).

Since he doesn't public his code I can say for sure but it maybe use your code. And there some reports that it works with AMD cards too.

PTS mining is exploding in a few days.

[earthbound]

2014-01-01 update:

I've committed some new changes to the repository.  Some are cosmetic, but three are important:

  1)  Memory use on the host side is reduced by about 500MB.  This may or may not matter for you.

  2)  Speed is boosted by 10-20% on a lot of platforms.  I have another speed boost patch coming next week once I've made it not horrible, but this one gets a decent chunk of the gains.

  3)  There's now a developer fee that goes to me.  Kinda. 

I'm doing an experiment with this code release in the developer fee:  It's easy to disable.  It's not hidden.  But it's also just a list of addresses that share the dev fee equally.

So here's my proposal:  If you port this software to another platform or release a binary, don't remove my address.  Instead, add yours to the list -- I've tried to make it super easy for you to get your own share.  If this works out, I'll continue to release improvements and try to make it even easier for other developers who improve upon the code, because we'll all have a reason to make software that remains open source and which is user-friendly and high performance.

If you think this is horrible, let me know and let's try to find a way to make it work better.

If you're a user who hates the idea of a dev fee, the source is yours and you can delete the addresses listed there and/or add your own.

  -Dave

Sorry I missed this. I think this is a good idea. Maybe

[dagi]

Hi, test miner on GTX 660.
arCUDAminer - 470c/m (win build from post Reply #91)
Nice!

[archit]

I'm sorry to say it increases pain. YASM is not a big problem I can deal with it

[dga]

Hey, @Archit, @crz, and anyone else who's been porting the software to other platforms:

Does it cause you increased pain if I include any of the Thrust header files?

You could try, e.g., uncommenting thrust/sort.h in gpuhash.cu and see if it breaks anything.  (I'll warn that I've actually had the most problems with it at link time when functions are used, though, not just from using the headers).

I'm trying to get rid of the need for yasm by moving a bit more functionality onto the GPU, and using Thrust will let me be lazy about it.  *grin*

  -Dave

[dga]

dga, do you have any idea why GTX680 card is not faster than 580? And there is any improvements of sm3.5 over sm3.0? In my test 3.5 card (GTX780) loses 20% cpm when using sm3.5 over sm3.0.
Thanks in advance if anyone could explain it for me.

Because the biggest bottleneck with my current design is memory latency and the number of memory references that can be in flight at a time.  This isn't something that improves nearly as fast with different generations of cards (and sometimes gets worse).  Though I'm surprised at your 780 slowing down with sm_35.  I'll have to try that.

Lesee.  K20c I get 1104 c/m with sm_35.  Results are nearly unchanged with sm_30.

The biggest crypto-related improvement in sm_35 is the funnel shifter, but the design of Momentum is less crypto-intensive than a lot of PoW functions because it generates 8 memory references for every one execution of SHA512.  This is in part why I don't expect there to be as big an nvidia-AMD gap (except that by being so much faster, the AMD crew can try to exploit space/time tradeoffs in different ways.).

I'm playing around with some very different design options for faster cards, but haven't decided on anything I really like yet.

[AizenSou]

GTX 780 - 1800cpm

Jesus Christ!

* donschoe went off shopping!

grin.  Now you're seeing which platforms I'm able to develop for.  Gorgeous result.

I have a dual GTX690 rig that's currently doing about 3600 c/m.  I like it.

dga, do you have any idea why GTX680 card is not faster than 580? And there is any improvements of sm3.5 over sm3.0? In my test 3.5 card (GTX780) loses 20% cpm when using sm3.5 over sm3.0.
Thanks in advance if anyone could explain it for me.

[AizenSou]

Any one who can test my miner as well and see how the speeds compare?

I would like too, but without source I can't compile for linux. I have a quite broad range of nvidia cards lying around, so it's very interesting to test, but my computer is all linux so windows binary is impossible for me .

[archit]

Any one who can test my miner as well and see how the speeds compare?

[dga]

GTX 780 - 1800cpm

Jesus Christ!

* donschoe went off shopping!

grin.  Now you're seeing which platforms I'm able to develop for.  Gorgeous result.

I have a dual GTX690 rig that's currently doing about 3600 c/m.  I like it.

[5chdn]

GTX 780 - 1800cpm

Jesus Christ!

* donschoe went off shopping!

[shai]

Tried it with my old 9800GT
it didn't print out any statistics (col/min ?)
and i almost burned the card... (it have a broken fan).. when funny smell started i stoped the cudapts

[AizenSou]

2014-01-01 update:

I've committed some new changes to the repository.  Some are cosmetic, but three are important:

  1)  Memory use on the host side is reduced by about 500MB.  This may or may not matter for you.

  2)  Speed is boosted by 10-20% on a lot of platforms.  I have another speed boost patch coming next week once I've made it not horrible, but this one gets a decent chunk of the gains.

  3)  There's now a developer fee that goes to me.  Kinda. 

I'm doing an experiment with this code release in the developer fee:  It's easy to disable.  It's not hidden.  But it's also just a list of addresses that share the dev fee equally.

So here's my proposal:  If you port this software to another platform or release a binary, don't remove my address.  Instead, add yours to the list -- I've tried to make it super easy for you to get your own share.  If this works out, I'll continue to release improvements and try to make it even easier for other developers who improve upon the code, because we'll all have a reason to make software that remains open source and which is user-friendly and high performance.

If you think this is horrible, let me know and let's try to find a way to make it work better.

If you're a user who hates the idea of a dev fee, the source is yours and you can delete the addresses listed there and/or add your own.

  -Dave

You mean 10-01-2013 update? 

Superb work dga. Since I have a few Nvidia cards lying around, I report some of my test since yesterday.

cudarts version 08-01-2013 (v7 if I not mistaken)
GTX 780 - 1450cpm
GTX 680 - 650 cpm
GTX 580 - 850 cpm (3GB memory)
GTX 580 - 920 cpm (1.5GB memory)
GTX 570 - 750 cpm
GTX 260 - 290 cpm

cudarts version 10-01-2013 (v8)
GTX 780 - 1800cpm
GTX 680 - 950 cpm
GTX 580 - 820 cpm - 930 cpm (3GB memory) (the value varies depends on card manufactures)
GTX 580 - 960 cpm (1.5GB memory)
GTX 570 - 770 cpm
GTX 260 - 240 cpm

Yeah most of cards got very nice bump, but I notice some reduction too. But the most nice thing of V8 is my card running at least 3grad C lower. No change on memory consumption.
There some points I still don't understand:
1. Why GTX680 card not much faster than 580. The GTX680 runs with v7 even slower than GTX580.
2. I tried to compile with sm_35 for GTX780 cards. But it got around 10-15% slower than with sm_30.

Anyway I'm very happy with this. Thanks dga.

[ryen123]

Performance feedback for Speed optimizations #1

GTX 570: Previously 760+ c/m > Now 790+ c/m
GTX 560 Ti: No performance difference 530+ c/m