ATTN Litecoin GPU Miners - Scrypt support for cgminer - page 23.

Rubberduckie

legendary

Activity: 1442

Merit: 1000

Quote from: smoothie on July 21, 2012, 02:15:33 AM

Quote from: Luke-Jr on July 21, 2012, 01:56:02 AM

Quote from: smoothie on July 21, 2012, 01:49:29 AM

well if u start another chain that does make mining only feasible for cpus im sure you will have many interested parties wanting to help the project

That is a technical impossibility.

And of course you are the "know it all". So please elaborate...

I was under the impression Luke Jr didn't like Alt chains

smoothie

legendary

Activity: 2492

Merit: 1473

LEALANA Bitcoin Grim Reaper

Quote from: Luke-Jr on July 21, 2012, 01:56:02 AM

Quote from: smoothie on July 21, 2012, 01:49:29 AM

well if u start another chain that does make mining only feasible for cpus im sure you will have many interested parties wanting to help the project

That is a technical impossibility.

And of course you are the "know it all". So please elaborate...

Luke-Jr

legendary

Activity: 2576

Merit: 1186

Quote from: smoothie on July 21, 2012, 01:49:29 AM

well if u start another chain that does make mining only feasible for cpus im sure you will have many interested parties wanting to help the project

That is a technical impossibility.

smoothie

legendary

Activity: 2492

Merit: 1473

LEALANA Bitcoin Grim Reaper

Quote from: tacotime on July 21, 2012, 12:48:23 AM

Quote from: smoothie on July 21, 2012, 12:46:05 AM

lets say they are "easy" to make......it will cost millions for production to start.

I answered my own question in the android miner thread. even running the optimized neon code from pooler an a9 isn't power competitive.

still, settings for scrypt using 32-128mb would have been better I think -- gpus have faster ram than most any other consumer or industry device.

edit: okay, sorry to be OT but if you wanted to make it CPU mining only, why wouldn't you just set the memory requirement to be 1 gigabyte per thread? Most computers right now have 4-16GB of RAM --> 4-16 threads, which would be terrible on a GPU with only 1-2GB of RAM. You could make the blockchain halving occur every 18 months and then with the blockchain halving increase the amount of RAM required by 2 eg (1GB -> 2GB). In accordance with Moore's law, that would keep the quantity of RAM being used on target for a CPU and most likely always above that provided for by a GPU.

So why didn't artforz do this in the first place?

well if u start another chain that does make mining only feasible for cpus im sure you will have many interested parties wanting to help the project

tacotime

legendary

Activity: 1484

Merit: 1005

Quote from: smoothie on July 21, 2012, 12:46:05 AM

lets say they are "easy" to make......it will cost millions for production to start.

I answered my own question in the android miner thread. even running the optimized neon code from pooler an a9 isn't power competitive.

still, settings for scrypt using 32-128mb would have been better I think -- gpus have faster ram than most any other consumer or industry device.

edit: okay, sorry to be OT but if you wanted to make it CPU mining only, why wouldn't you just set the memory requirement to be 1 gigabyte per thread? Most computers right now have 4-16GB of RAM --> 4-16 threads, which would be terrible on a GPU with only 1-2GB of RAM. You could make the blockchain halving occur every 18 months and then with the blockchain halving increase the amount of RAM required by 2 eg (1GB -> 2GB). In accordance with Moore's law, that would keep the quantity of RAM being used on target for a CPU and most likely always above that provided for by a GPU.

So why didn't artforz do this in the first place?

smoothie

legendary

Activity: 2492

Merit: 1473

LEALANA Bitcoin Grim Reaper

lets say they are "easy" to make......it will cost millions for production to start.

tacotime

legendary

Activity: 1484

Merit: 1005

Quote from: kano on July 20, 2012, 09:16:05 PM

Quote from: tacotime on July 20, 2012, 09:09:37 PM

Quote from: Tittiez on July 20, 2012, 08:00:04 PM

Quote from: tacotime on July 20, 2012, 07:08:41 PM

Are you running the cards all on the same rig? LTC mining is GPU memory bandwidth intensive, if all 4 cards are running at only
1x or 4x or even 8x you'll probably be bottlenecked.

Try using only one GPU in running at 16x.

I have my 5770 on an x1 and it's mining ~200kh/s just fine.

So memory bandwidth has nothing to do with it, at least over the southbridge... Someone should really post a chart that plots performance in kh/s with the speed of the memory for the GPU. I'm really wondering if it's linear or not. This implementation of scrypt is small enough to be done in the L2 cache, so I have always been wondering how the GPU implementation required megabytes. I wish someone would break down the algorithm for us or something, but I guess it's in the source code.

N=1024,p=1,r=1 -> 128KiB
i.e. each thread requires 128K or RAM
e.g. in a GPU with 128 parallel processes running at the same time that 128KiB each becomes 16MB

shouldn't an asic be easy to make then? I mean arm cortex a9s can carry up to 1m level 2 cache... they're slow but they're also cheap and energy efficient. I think they only use about 500mW per cpu

ThiagoCMC

legendary

Activity: 1204

Merit: 1000

฿itcoin: Currency of Resistance!

kH or Kh... It does matter?!

Tittiez

hero member

Activity: 686

Merit: 500

Quote from: -ck on July 20, 2012, 11:20:25 PM

Okay, how's this looking now Wink

Note, mining at higher difficulty 20 on ozcoin.

Code:

cgminer version 2.5.0 - Started: [2012-07-21 14:15:53]
--------------------------------------------------------------------------------
(5s):2213.6 (avg):2076.3 Kh/s | Q:78 A:470 R:5 HW:0 E:603% U:111.9/m
TQ: 2 ST: 3 SS: 0 DW: 61 NB: 5 LW: 3253 GF: 0 RF: 0
Connected to http://lc.ozco.in with LP as user ckolivas.0
Block: a8a8f3d4efcb35dfeac6906cf49b15a3... Started: [14:19:34]
--------------------------------------------------------------------------------
[P]ool management [G]PU management [S]ettings [D]isplay options [Q]uit
GPU 0: 63.0C 5356RPM | 553.6/535.1Kh/s | A:131 R:3 HW:0 U: 31.20/m I:13
GPU 1: 61.0C 5429RPM | 537.1/519.3Kh/s | A:120 R:2 HW:0 U: 28.58/m I:13
GPU 2: 59.0C 5325RPM | 540.1/522.9Kh/s | A:112 R:0 HW:0 U: 26.68/m I:13
GPU 3: 55.0C 5308RPM | 573.5/529.9Kh/s | A:110 R:0 HW:0 U: 26.20/m I:13
--------------------------------------------------------------------------------

Much better!

foggyb

legendary

Activity: 1736

Merit: 1006

Now we're talking. Good work. Hopefully there's more headroom left.

To the haters: nothing wrong with setting the bar high. You'd all be unhappy if ckolivas released a miner that was slower or the same speed as reaper. Don't be hypocrites.

-ck

legendary

Activity: 4088

Merit: 1631

Ruu \o/

Okay, how's this looking now Wink

Note, mining at higher difficulty 20 on ozcoin.

Code:

cgminer version 2.5.0 - Started: [2012-07-21 14:15:53]
--------------------------------------------------------------------------------
(5s):2213.6 (avg):2076.3 Kh/s | Q:78 A:470 R:5 HW:0 E:603% U:111.9/m
TQ: 2 ST: 3 SS: 0 DW: 61 NB: 5 LW: 3253 GF: 0 RF: 0
Connected to http://lc.ozco.in with LP as user ckolivas.0
Block: a8a8f3d4efcb35dfeac6906cf49b15a3... Started: [14:19:34]
--------------------------------------------------------------------------------
[P]ool management [G]PU management [S]ettings [D]isplay options [Q]uit
GPU 0: 63.0C 5356RPM | 553.6/535.1Kh/s | A:131 R:3 HW:0 U: 31.20/m I:13
GPU 1: 61.0C 5429RPM | 537.1/519.3Kh/s | A:120 R:2 HW:0 U: 28.58/m I:13
GPU 2: 59.0C 5325RPM | 540.1/522.9Kh/s | A:112 R:0 HW:0 U: 26.68/m I:13
GPU 3: 55.0C 5308RPM | 573.5/529.9Kh/s | A:110 R:0 HW:0 U: 26.20/m I:13
--------------------------------------------------------------------------------

smoothie

legendary

Activity: 2492

Merit: 1473

LEALANA Bitcoin Grim Reaper

Quote from: Graet on July 20, 2012, 11:08:16 PM

Quote from: wknight on July 20, 2012, 09:41:36 PM

Quote from: -ck on July 20, 2012, 09:39:58 PM

Oh I'm not afraid of putting the code up public. I'm just wary of putting broken code up there and getting lots of people trying it and starting to file meaningless bug reports when the code is in massive flux. I was planning on putting it up shortly anyway.

Take your time.. no need to rush code when you are making sooooo many changes.

+50hundred

none of the people that have said how "easy" it is have done anything (but I guess 99.9% of them never tried Cheesy

) - conman is actually working hard at this
a little patience people

lmao @ ppl demanding btc back when code isnt even finished being written . oh sorry "threatening". way to encourage work....NOT

+1....."i'll pledge x btc.......wait........gimme my btc back u f*g since you didn't finish the code faster!"

Graet

vip

Activity: 980

Merit: 1001

Quote from: wknight on July 20, 2012, 09:41:36 PM

Quote from: -ck on July 20, 2012, 09:39:58 PM

Oh I'm not afraid of putting the code up public. I'm just wary of putting broken code up there and getting lots of people trying it and starting to file meaningless bug reports when the code is in massive flux. I was planning on putting it up shortly anyway.

Take your time.. no need to rush code when you are making sooooo many changes.

+50hundred

none of the people that have said how "easy" it is have done anything (but I guess 99.9% of them never tried Cheesy

) - conman is actually working hard at this
a little patience people

lmao @ ppl demanding btc back when code isnt even finished being written . oh sorry "threatening". way to encourage work....NOT

wknight

legendary

Activity: 889

Merit: 1000

Bitcoin calls me an Orphan

Quote from: -ck on July 20, 2012, 09:39:58 PM

Oh I'm not afraid of putting the code up public. I'm just wary of putting broken code up there and getting lots of people trying it and starting to file meaningless bug reports when the code is in massive flux. I was planning on putting it up shortly anyway.

Take your time.. no need to rush code when you are making sooooo many changes.

-ck

legendary

Activity: 4088

Merit: 1631

Ruu \o/

Oh I'm not afraid of putting the code up public. I'm just wary of putting broken code up there and getting lots of people trying it and starting to file meaningless bug reports when the code is in massive flux. I was planning on putting it up shortly anyway.

kano

legendary

Activity: 4592

Merit: 1851

Linux since 1997 RedHat 4

Quote from: tacotime on July 20, 2012, 09:09:37 PM

Quote from: Tittiez on July 20, 2012, 08:00:04 PM

Quote from: tacotime on July 20, 2012, 07:08:41 PM

Are you running the cards all on the same rig? LTC mining is GPU memory bandwidth intensive, if all 4 cards are running at only 1x or 4x or even 8x you'll probably be bottlenecked.

Try using only one GPU in running at 16x.

I have my 5770 on an x1 and it's mining ~200kh/s just fine.

So memory bandwidth has nothing to do with it, at least over the southbridge... Someone should really post a chart that plots performance in kh/s with the speed of the memory for the GPU. I'm really wondering if it's linear or not. This implementation of scrypt is small enough to be done in the L2 cache, so I have always been wondering how the GPU implementation required megabytes. I wish someone would break down the algorithm for us or something, but I guess it's in the source code.

N=1024,p=1,r=1 -> 128KiB
i.e. each thread requires 128K or RAM
e.g. in a GPU with 128 parallel processes running at the same time that 128KiB each becomes 16MB

tacotime

legendary

Activity: 1484

Merit: 1005

Quote from: Tittiez on July 20, 2012, 08:00:04 PM

Quote from: tacotime on July 20, 2012, 07:08:41 PM

Are you running the cards all on the same rig? LTC mining is GPU memory bandwidth intensive, if all 4 cards are running at only 1x or 4x or even 8x you'll probably be bottlenecked.

Try using only one GPU in running at 16x.

I have my 5770 on an x1 and it's mining ~200kh/s just fine.

So memory bandwidth has nothing to do with it, at least over the southbridge... Someone should really post a chart that plots performance in kh/s with the speed of the memory for the GPU. I'm really wondering if it's linear or not. This implementation of scrypt is small enough to be done in the L2 cache, so I have always been wondering how the GPU implementation required megabytes. I wish someone would break down the algorithm for us or something, but I guess it's in the source code.

Tittiez

hero member

Activity: 686

Merit: 500

Quote from: bitlane on July 20, 2012, 07:40:22 PM

Lots of configs

Hey, try setting all the ones that are at 6144 to 7144, it increases my hashrate. Wink

Tittiez

hero member

Activity: 686

Merit: 500

Quote from: tacotime on July 20, 2012, 07:08:41 PM

Are you running the cards all on the same rig? LTC mining is GPU memory bandwidth intensive, if all 4 cards are running at only 1x or 4x or even 8x you'll probably be bottlenecked.

Try using only one GPU in running at 16x.

I have my 5770 on an x1 and it's mining ~200kh/s just fine.

kano

legendary

Activity: 4592

Merit: 1851

Linux since 1997 RedHat 4

Quote from: jjiimm_64 on July 20, 2012, 03:50:14 PM

Quote from: dizzy1 on July 20, 2012, 03:20:40 PM

Quote from: -ck on July 20, 2012, 12:02:51 PM

Quote from: dizzy1 on July 20, 2012, 11:43:01 AM

Atm the reaper opencl code sucks. My 5870 generates about 420kh/s, to use as a reference. So it should be getting like 6-700 or more from a 7970.

But how many accepted shares does it generate at that alleged hashrate?

According to the pool, about 55k accepted since the night of July 18th ( almost 3 days now) but I also game on this box. I also have a tri 5970 rig that has 250k+ accepted shares. I've found that the thread concurrency switch in reaper has a lot to do with stales/invalids. Also the calculated pool speed is usually +/- 10% of 420, right now at 436.

keep in mind, scrypt was intentionally gpu hostile. you cannot extrapolate what one gpu 'might' or 'should' get because another is getting x this is not btc mining.
...

Yes well as everyone knows ... it was supposedly impossible to GPU mine Scrypt faster than CPU with N=1024,p=1,r=1 -> 128KiB

September 28, 2011
https://bitcointalksearch.org/topic/m.548778

Edit: of course that link above - 3 posts later - shows ArtForz saying he wrote a scrypt GPU miner back in September last year ...........

Though I had a few pages of discussion on the subject of generic GPU vs CPU mining even before that ...
September 17, 2011 starting here:
https://bitcointalksearch.org/topic/m.530426
That all turned out to be a load of crap also

Topic: ATTN Litecoin GPU Miners - Scrypt support for cgminer - page 23. (Read 175855 times)