CCminer(SP-MOD) Modded GPU kernels.

[joblo]

Also 90-95% of bids was probably already received when the price was still at $240-$250 range. I seriously doubt anyone who tried to bid within the last 24-36 hours probably did not meet the deadline with going through the process.

I'm not trying to pick a fight here but I think you're wrong. If I was to bid in an aution for a security that is
exchanged on an open market I would submit my bid at the last possible moment for fear that the price
could drop and I could be left holding the bag with a bid above market. It's all about minimizing the time
between selecting a bid price and getting delivery. The longer the time the more that could go wrong.

If you don't believe this is the case let me tell you about when I worked for a low tech telecommunications
company (post office). Government tenders closed at 5 PM so every time there was a tender closure everyone
was told to skim through the last minute incoming mail searching for tenders (the envelope were highly visible).
Unlike Bitcoins, the cost of the contracts didn't change from day to day but the tenders still flirted with the deadline
even with the unreliable post office. With this bitcoin auction all you need to do is send a PDF which can be done,
liretally, at the last minute.

[lmaonade80]

I am bored - are there any new or good coins to mine with (3) Nvidia rigs?

you can mine boolberry

lol yaah I suppose he could... Terrible profit margin though

[joblo]

the argument is not whether the the auction is the casue of the price rise or not - it is about the reason WHY its being pumped ...

I'm not sure it is being pumped. Pumping altcoins is relatively easy but bitcoin takes a lot more capital. Not government
size, not megacorp size, but players currently bigger than anyone else currently in the crypto market to be able to
manipulate the price of a USD 6B security by 100%.

Like I said before I'm liking the idea that the Chinese are turning to bitcoin to circumvent currency controls.
If so it's a big step forward for crypto.

[lmaonade80]

Thanks

I am looking at the Etherum miner now. I have some improvements.

Perhaps allowing ethminer to run on the 750ti's?!

[chrysophylax]

the argument is not whether the the auction is the casue of the price rise or not - it is about the reason WHY its being pumped ...

I'm not sure it is being pumped. Pumping altcoins is relatively easy but bitcoin takes a lot more capital. Not government
size, not megacorp size, but players currently bigger than anyone else currently in the crypto market to be able to
manipulate the price of a USD 6B security by 100%.

Like I said before I'm liking the idea that the Chinese are turning to bitcoin to circumvent currency controls.
If so it's a big step forward for crypto.

that could be the case mate ...

which ever way - we are all be speculating anyhow ...

the beauty about it all is - that btc IS moving - and moving very well at the moment ...

if you injected another few million dollars ( like the winklevoss bros did a couple of years back with their $11million ) - then the movement would be very nice upwards again ...

ive never really been hung up about btc / fiat prices ... coins are coins are coins ... you can use them all if the service giver is willing to accept the coin you are willing to trade ... so its no major issue on this end ... i just like to watch - beats the hell out of sports ...

btc however IS a serious confronting notion to the big players in this world - the ones that do control markets ...

so if it ever came out that 'they' were involved in the fiat price fluctuations - i wouldnt bat an eyelid over it ...

give me my partner - good food - my place - my farm - and my sports car - and im a very happy guy which ever way it turns out ...

#crysx

[sp_]

According to  http://coinmarketcap.com. Bitcoins  has been traded for  $259,625,000 in the last 24Hours. But this webpage doesn't include the biggest chinese exchanges like OKCoin, huobi, BTCChina and Bter

The bitcoin USD market is just a small fraction of the world market. In the recent rally  more than 1 billion$ worth of bitcoins has been traded in a couple of days.

Another interesting point is that there is Arbitrage between the CNY exchanges and the USD
   
Huobi (china) $431.73/ (2740.21 CNY)
Bitstamp       $389.98

Bitcoins are 10% more expensive to buy in china than in Europe / America. Perhaps because somebody is buying in china and selling in europe..
Perhaps someone is moving fiat currency trough the bitcoin blockchain between borders without having any governement control...

[pallas]

and the price rebounded ;-)

[sp_]

And the altocoins are up, and the mining profit meassured in BTC is up..

[cbuchner1]

Hi all, could this thread be relevant for Maxwell based mining? Supposedly there is a trick to boost memclock safely using nvidia-smi.

https://devtalk.nvidia.com/default/topic/892842/cuda-programming-and-performance/one-weird-trick-to-get-a-maxwell-v2-gpu-to-reach-its-max-memory-clock-/

Christian

[pallas]

Hi all, could this thread be relevant for Maxwell based mining? Supposedly there is a trick to boost memclock safely using nvidia-smi.

https://devtalk.nvidia.com/default/topic/892842/cuda-programming-and-performance/one-weird-trick-to-get-a-maxwell-v2-gpu-to-reach-its-max-memory-clock-/

Christian

I tried it some months ago on a 970 and it worked, but you loose the auto clock tuning based on temperature and fan speed, which you may or may not like ;-)
I personally like it and ended up with better hashrate using the default settings than changing it with nvidia-smi, but YMMV.

[sp_]

You can boost the core clock as well if you put the gpu in the p1 state.

p2 state: adjust memory
p1 state: adjust gpu boost clock

I think I will make code that does this automaticly with an option.. --Max-Boost --Max-memspeed

1. List all possible modes, parse the information and select the fastest for each gpu in the rig
2. Set p1 mode and set the max possible gpu-boost clock
3. set p2 mode and set the max possible memclock clock

The NVIDIA api is broken in x86 builds, so I will call the commandline from the c++ code in the same way I have done before. (windows only)

For 970 cards with 4pin+3pin connector, boosting is important to reach full speed. (but this costs more power)

Other cards will probobly trottle and perform worse..

(f.ex all the gtx 970 mini cards, and the 970 cards that are made small to fit in a small box.)

[sp_]

Testing sp-mod release 72 opensource

On the GTX 970 OC windforce G1 boost: 1505Mhz core clock I get

19MHASH in the quark algo.

The titanx reference is doing 20.8MHASH

(source http://cryptomining-blog.com/tag/gtx-titan-x-hashrate/)


But the Titanx costs $1000 and the gtx970 1/3rd of that..


So I get 3 gtx970's for the price of one titanx. and almost 3x the hashrate and income of the $1000 spent...

[sp_]

With a $10 000 rig:



Only

188MHASH@quark

But with 30 970 cards (will cost around $10 000)

you get

540 - 570MHASH@quark

[tbearhere]

You can boost the core clock as well if you put the gpu in the p1 state.

p2 state: adjust memory
p1 state: adjust gpu boost clock

I think I will make code that does this automaticly with an option.. --Max-Boost --Max-memspeed

1. List all possible modes, parse the information and select the fastest for each gpu in the rig
2. Set p1 mode and set the max possible gpu-boost clock
3. set p2 mode and set the max possible memclock clock

The NVIDIA api is broken in x86 builds, so I will call the commandline from the c++ code in the same way I have done before. (windows only)

For 970 cards with 4pin+3pin connector, boosting is important to reach full speed. (but this costs more power)

Other cards will probobly trottle and perform worse..

(f.ex all the gtx 970 mini cards, and the 970 cards that are made small to fit in a small box.)

I can't get my 980ti to boost at all on mem....

[tbearhere]

Hi all, could this thread be relevant for Maxwell based mining? Supposedly there is a trick to boost memclock safely using nvidia-smi.

https://devtalk.nvidia.com/default/topic/892842/cuda-programming-and-performance/one-weird-trick-to-get-a-maxwell-v2-gpu-to-reach-its-max-memory-clock-/

Christian

Hi Christian, how you been?.... Thank you for the info.

[tbearhere]

And the altocoins are up, and the mining profit meassured in BTC is up..

I have twice the hash power now and making the same amount. 

[tbearhere]

The EVGA_PrecisionX_16 does support more than 3 cards.....

[Genoil]

Thanks
I am looking at the Etherum miner now. I have some improvements.

Very curious what you come up with. I hope you can challenge me to look at the code once again, Kind of lost interest with the whole TLB trashing thing going on on Windows.

in the dagger.cuh:

__device__ uint4 fnv4(uint4 a, uint4 b)
{
   uint4 c;
   c.x = a.x * FNV_PRIME ^ b.x;
   c.y = a.y * FNV_PRIME ^ b.y;
   c.z = a.z * FNV_PRIME ^ b.z;
   c.w = a.w * FNV_PRIME ^ b.w;
   return c;
}


Since a.x*2^24= a.x<<24


This can be rewritten to:

__device__ uint4 fnv4(uint4 a, uint4 b)
{
   c.x = sharedmemprecalc[a.x&0xff]^ b.x;
   c.y = sharedmemprecalc[a.y&0xff] ^ b.y;
   c.z = sharedmemprecalc[a.z&0xff] ^ b.z;
   c.w = sharedmemprecalc[a.w&0xff] ^ b.w;
   return c;
}

The precalcbuffer must be 32bit  (256*4 bytes) and the  values shifted by 24 bits (shared mem level1cache):

xx000000

Code:

		__shared__ uint32_t sharedmemprecalc[256 * 4];
		for (int i = 0; i<256; i++)
		{
			sharedmemprecalc[i] = (193 * i) << 24;  // Since the FNV_PRIME is a high number the 24 highest bits of the product are ignored. We only need to know the 8 low bits.
		}

since you ony need to read 1 byte and not the whole 4 bytes (32 bits), you might be able to solve it with 1/4th of the memory reads...

__device__ uint4 fnv4(uchar4 a, uint4 b)
{
   c.x = sharedmemprecalc[a.x]^ b.x;
   c.y = sharedmemprecalc[a.y] ^ b.y;
   c.z = sharedmemprecalc[a.z] ^ b.z;
   c.w = sharedmemprecalc[a.w] ^ b.w;
   return c;
}

But you might have to reorganize /scramble the memory. and read 32 bit lineary in one read to fill the uchar4.

I looked into this a for a bit. I rewrote it like this:

block size == 128
lower byte of FNV_PRIME is 147, not 193. (0x01000193 & 0xFF = 0x93 == 147)

Code:

__shared__ uint32_t sharedmemprecalc[256];

In compute_hash_shuffle:

sharedmemprecalc[threadIdx.x] = (147 * threadIdx.x) << 24;  
sharedmemprecalc[threadIdx.x + 128] = (147 * (threadIdx.x + 128)) << 24;
__syncthreads();

And this unmodified:

__device__ uint4 fnv4s(uint4 a, uint4 b)
{
	uint4 c;
	c.x = sharedmemprecalc[a.x & 0xff] ^ b.x;
	c.y = sharedmemprecalc[a.y & 0xff] ^ b.y;
	c.z = sharedmemprecalc[a.z & 0xff] ^ b.z;
	c.w = sharedmemprecalc[a.w & 0xff] ^ b.w;
	return c;
}

It doesn't work. But more importantly, your suggestion of only requiring the lower byte from the DAG entry seems wrong, since that's param b in the fnv function...

[scryptr]

The EVGA_PrecisionX_16 does support more than 3 cards.....

MAXIMUM OF FOUR CARDS--

It only supports four at the most.  I had to remove EVGA PrecisionX 16 from my rig when I added a fifth card, it complained on boot-up.       --scryptr

[tbearhere]

The EVGA_PrecisionX_16 does support more than 3 cards.....

MAXIMUM OF FOUR CARDS--

It only supports four at the most.  I had to remove EVGA PrecisionX 16 from my rig when I added a fifth card, it complained on boot-up.       --scryptr

Ok ...scryptr what are you using in place of it.   Thx
EDIT: It was complaining about the 4th card though.... that's what I have in there.