Topic: [ANNOUNCE] Tenebrix, a CPU-friendly, GPU-hostile cryptocurrency - page 17. (Read 127188 times)
October 06, 2011, 12:41:41 PM
Awesum, thx
October 06, 2011, 12:32:43 PM
Btw, where can i find a calculator to show how long till i have 25%, 50% and 100% chance of having found a block (or somthing like that)?
1 difficulty in the real world (and here also) represents 2^32 hashes.multiply 2^32 by the difficulty and that gives you an average expected number of hashes per share.
also displays average time for a block at 1kH/s so you can easily scale it
October 06, 2011, 12:14:07 PM
It looks like a rig with a bulldozer CPU and multiple CGN architecture graphics cards will flood the hashrates seen today.
The registers are still 64 bit in the 7xxx series... what people don't understand about the cache is that they take up a significant part of the die. The point of GPUs is to execute parallel instructions from sequentially stored memory because that's what graphics acceleration and scientific process acceleration requires. They won't add giant caches and make the die size 150% larger (making the effective cost per unit 50% higher) if 99% of their consumers won't use it.
http://www.anandtech.com/show/2702/2
For instance, with Phenom II's the L2 cache size is slightly less than the size of an extra two cores and the combined L1, L2, and L3 cache system consumes about as much space as all the cores. The whole point of a GPU/CPU combined system is that the CPU handles instructions that require rapid random access to small amounts of memory on the cache while the GPU handles large parallel instructions that operate with small amounts of cache and sequentially read memory at high rates. If they both could do everything, we wouldn't need both.
October 06, 2011, 11:50:12 AM
Just a bulldozer will decimate everything seen today most likely.
October 06, 2011, 11:40:13 AM
A clarification. The "low end cards" 77xx and 78xx series use same architecture as 6xxx series with a die shrink. So simplisticly double the performance per $ and performance per watt and not much else new. They will continue to use the VLIW4 instruction set.
The 79xx series will use a completely new architecture currently coded named "GCN" (Graphics Core Next) by AMD. Not much is known about it. It could have much better or worse performance per clock than 5xxx and 6xxx series. Wouldn't it be ironic is the 79xx series did well on "CPU friendly" chains and sucked at traditional SHA-256 chains.
I found an article about it http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute
It looks like a rig with a bulldozer CPU and multiple CGN architecture graphics cards will flood the hashrates seen today.
October 06, 2011, 10:59:29 AM
Please can somebody also post a OpenCL miner as I think it could be useful even with scrypt algo. Thanks !
LOL that is seriously funny October 06, 2011, 10:59:03 AM
Btw, where can i find a calculator to show how long till i have 25%, 50% and 100% chance of having found a block (or somthing like that)?
1 difficulty in the real world (and here also) represents 2^32 hashes.multiply 2^32 by the difficulty and that gives you an average expected number of hashes per share.
but if you run 'tbd getinfo' it will tell you the difficulty (among other things) if the daemon is running
Maybe the qt one will do it too? No idea
October 06, 2011, 10:55:32 AM
Please can somebody also post a OpenCL miner as I think it could be useful even with scrypt algo. Thanks !
October 06, 2011, 10:49:16 AM
Btw, where can i find a calculator to show how long till i have 25%, 50% and 100% chance of having found a block (or somthing like that)?
1 difficulty in the real world (and here also) represents 2^32 hashes.multiply 2^32 by the difficulty and that gives you an average expected number of hashes per share.
October 06, 2011, 10:44:34 AM
Btw, where can i find a calculator to show how long till i have 25%, 50% and 100% chance of having found a block (or somthing like that)?
1 difficulty in the real world (and here also) represents 2^32 hashes.multiply 2^32 by the difficulty and that gives you an average expected number of hashes per share.
October 06, 2011, 09:29:34 AM
Anyone know how to squeeze more hashes out of my 2500k? Running at 4.2 ghz with 4 threads, I am getting a little over 2 khash per thread using the default settings for the Windows miner.
October 06, 2011, 08:48:41 AM
When i mine this on my i3, the hash rate of my Gpu's goes down from ~300mh/s to ~20mh/s :|
Use 1 or 2 less threads, the GPU miner needs to have a CPU thread too.Exactly while GPU is doing the heaving lifting it does need CPU time. Not enough CPU time means the GPU becomes starves and is sitting idle for large periods of time (looks like 33% in your case).
Another thing you could try is make the GPU mining process high priority and the CPU mining process low priority. The CPU process will use the "unused" clock cycles from GPU thread without starving it. In otherwords when the GPU needs CPU time it gets it but when it doesn't those cycles can CPU mine.
Right now that might not work if you are suffering from the 100% CPU bug (multiple AMD GPU) but if AMD ever fixes their 100% CPU bug it would be a way to maximize your CPU "time.
October 06, 2011, 08:37:35 AM
When i mine this on my i3, the hash rate of my Gpu's goes down from ~300mh/s to ~20mh/s :|
Use 1 or 2 less threads, the GPU miner needs to have a CPU thread too. October 06, 2011, 08:31:25 AM
When i mine this on my i3, the hash rate of my Gpu's goes down from ~300mh/s to ~20mh/s :|
October 06, 2011, 07:14:51 AM
Aren't the new 7xxx series supposed to be a new CU architecture has a separate access to the LDS with 128 Bytes/clock bandwidth?
Thought I read that a while back somewhere?
Hmm I read they will most likely use half the power with the same performance (28nm)... but other than that ... ?Thought I read that a while back somewhere?
A clarification. The "low end cards" 77xx and 78xx series use same architecture as 6xxx series with a die shrink. So simplisticly double the performance per $ and performance per watt and not much else new. They will continue to use the VLIW4 instruction set.
The 79xx series will use a completely new architecture currently coded named "GCN" (Graphics Core Next) by AMD. Not much is known about it. It could have much better or worse performance per clock than 5xxx and 6xxx series. Wouldn't it be ironic is the 79xx series did well on "CPU friendly" chains and sucked at traditional SHA-256 chains.
October 06, 2011, 07:04:08 AM
ArtForz: do you think this would work?
Nope, haven't played with nv yet, and certainly not with teslas.But looking at the spec and architecture...
14 CUs with 64kB L1/LDS per CU and 768kB global cache (128k per 64-bit memory controller). 32 ALUs/CU, clocked at 1.15GHz
6970 = 24 CUs with 64kB L1/LDS per CU and 512kB global cache. 64 ALUs/CU, clocked at 880MHz
nv L1 can do 2 reads/clock, so... wild-ass guess... might end up at 30kH/s or so.
Of course the lack of a native rotate won't exactly help, scrypt is sha256 (we know that one...) and salsa20, in which *every 2nd op* is a rotate.
so... probably more like 20kH/s.
Thank you for showing us the working. It looks like ATI/AMD's do have native rotate so something in the FirePro range may get a good hash rate. Either that or the new HD7000 series as Bitcoin Express suggested.
October 06, 2011, 04:45:39 AM
Probably because these "issues" aren't issues unless you have a ancient P4 or something.
My PhenomII is happily hashing along at 20kHps while increasing system power usage at the wall by 108W.
2^32 * 0.0926 / 20000 / 3600 ... about 5.5h/block
so ~108tbx/day for 2.6kWh/day
*checks btc-e* 108tbx = about 0.83 btc less fees, let's say 0.8
at 4.82/btc ... $3.85/day less power... about $3.20
Yeah, can't see why people are throwing as many CPUs as they can find at it...
because I have an E8400 so only 2.6khpsMy PhenomII is happily hashing along at 20kHps while increasing system power usage at the wall by 108W.
2^32 * 0.0926 / 20000 / 3600 ... about 5.5h/block
so ~108tbx/day for 2.6kWh/day
*checks btc-e* 108tbx = about 0.83 btc less fees, let's say 0.8
at 4.82/btc ... $3.85/day less power... about $3.20
Yeah, can't see why people are throwing as many CPUs as they can find at it...
also, the fan is loud as hell
actually, hold on, I'll start mining to heat my house
October 06, 2011, 03:26:12 AM
Aren't the new 7xxx series supposed to be a new CU architecture has a separate access to the LDS with 128 Bytes/clock bandwidth?
Thought I read that a while back somewhere?
Hmm I read they will most likely use half the power with the same performance (28nm)... but other than that ... ? Thought I read that a while back somewhere?
October 06, 2011, 03:24:06 AM
Probably because these "issues" aren't issues unless you have a ancient P4 or something.
My PhenomII is happily hashing along at 20kHps while increasing system power usage at the wall by 108W.
2^32 * 0.0926 / 20000 / 3600 ... about 5.5h/block
so ~108tbx/day for 2.6kWh/day
*checks btc-e* 108tbx = about 0.83 btc less fees, let's say 0.8
at 4.82/btc ... $3.85/day less power... about $3.20
Yeah, can't see why people are throwing as many CPUs as they can find at it...
Coz most people have at most 1 fast CPU - then there are those who don't even have that.My PhenomII is happily hashing along at 20kHps while increasing system power usage at the wall by 108W.
2^32 * 0.0926 / 20000 / 3600 ... about 5.5h/block
so ~108tbx/day for 2.6kWh/day
*checks btc-e* 108tbx = about 0.83 btc less fees, let's say 0.8
at 4.82/btc ... $3.85/day less power... about $3.20
Yeah, can't see why people are throwing as many CPUs as they can find at it...
... and, as I said: ... to double the performance you gotta buy another whole computer ...
(or throw away your CPU and buy a new one ... if you can upgrade it)
October 06, 2011, 03:05:35 AM
ArtForz: do you think this would work?
Nope, haven't played with nv yet, and certainly not with teslas.But looking at the spec and architecture...
14 CUs with 64kB L1/LDS per CU and 768kB global cache (128k per 64-bit memory controller). 32 ALUs/CU, clocked at 1.15GHz
6970 = 24 CUs with 64kB L1/LDS per CU and 512kB global cache. 64 ALUs/CU, clocked at 880MHz
nv L1 can do 2 reads/clock, so... wild-ass guess... might end up at 30kH/s or so.
Of course the lack of a native rotate won't exactly help, scrypt is sha256 (we know that one...) and salsa20, in which *every 2nd op* is a rotate.
so... probably more like 20kH/s.
Jump to: