= conventional =
* amazon ec2+GPU
* high-end ATI cards 5770,5870,5970
= wildcards =
* map/reduce miner in javascript.
* ASIC miner
* BTC/USD fluctuations
* bitcoin difficulty setting
= performance talk from irc =
30-Nov-2010
01:11 <+Keefe> tylergillies: 760 mhps, 5770 + 5970
09:13 < donpdonp> Keefe: that 760 megahashes, thats from two video cards?
09:16 <+brocktice> donpdonp: yeah
09:16 <+brocktice> donpdonp: the 5970 will turn around 600 depending on how much you
overclock it
09:17 <+brocktice> donpdonp: the 5770 will do about 150-180 depending on overclocking
09:17 <+brocktice> so I'm guessing that's probably 600 (5970) + 160 (5770) = 760
09:17 < donpdonp> brocktice: neat. im surprised that those are ATI cards. I thought CODA
was used
09:17 <+brocktice> donpdonp: nope, OpenCL
09:17 <+brocktice> nvidia cards suck for mining
09:18 <+brocktice> the integer performance per dollar and/or watt is much better on ATI
cards
09:18 <+brocktice> heck, these days even the floating point is better
09:18 < donpdonp> brocktice: is the opencl miner available?
09:18 <+brocktice> http://www.bitcoin.org/smf/index.php?topic=1334.0;all
09:18 < donpdonp> not that i even own an ATI card, just curious :)
09:18 <+brocktice> I'm currently turning about 1.7 GHash/s
09:18 <+brocktice> 2 5970s and 2 5770s
09:19 < donpdonp> is that in one box?
09:19 <+brocktice> all overclocked to various degrees, the 5970s are water cooled
09:19 <+brocktice> no, two boxes
09:19 <+brocktice> the 5770s are in my workstation
09:19 <+brocktice> the 5970s are in a dedicated miner
09:19 <+brocktice> but with two PSUs you can put them in one box
09:19 <+brocktice> ArtForz does.
09:19 <+brocktice> he uses extra fans.
09:19 < donpdonp> interesting. have you calculated the bitcoin value vs. electricity?
09:19 <+brocktice> http://www.flickr.com/photos/brocktice/5195604987/
09:20 <+brocktice> yep, it's very profitable vs electricity
09:20 <+brocktice> the hardware is getting harder to pay off though
09:22 < donpdonp> brocktice: how long until a block is only 25 coins? wont that make
mining unprofitable?
09:22 <+brocktice> another 2-3 years
09:22 <+brocktice> and no, probably not
09:22 <+brocktice> depends on difficulty really, and future gpu hardware
09:22 * donpdonp nods
09:23 <+brocktice> and btc price
09:23 <+brocktice> a lot of variables
09:23 <+brocktice> I did some quick calculations a month ago and figured I could pay off
the miner at current difficulty/prices in less than a month.
09:23 <+brocktice> with increases in difficulty and price fluctuations it's looking more
like a month and a half.
09:25 <+brocktice> ArtForz is running something like 24 5970s
09:25 < donpdonp> zomg.
09:25 <+brocktice> he has really cheap electricity
09:25 <+brocktice> and he's using them to heat his house this winter, basically
09:25 <+brocktice> he started very early with a custom miner.
09:25 <+brocktice> and all his equipment is already paid for from selling his btc
09:25 <+brocktice> he's spent around $12000 now I think
09:25 <+brocktice> maybe more
09:26 <+brocktice> he has about 20-30% of the hashes on the network
= multiple miners =
2010-Dec-1
11:10 < JudStephenson> I have another question: If I have two computers generating hashes,
what is to stop them from generating the same hashes? Since there
is no "progress" on a block, two computers generating a total
22,000 kh/sec isn't the same as one computer generating the same,
correct?
11:11 < OneFixt> JudStephenson: As long as you don't copy your wallet/addresses from one
computer to the other, you should be ok.
11:12 < OneFixt> Each computer will add a different address of its own before hashing.
11:13 < JudStephenson> ah, so they aren't generating hashes in the same space?
11:13 < JudStephenson> \
11:13 < [Noodles]> depends on how you connect them to a client, you can remotly connect
multiple "miners" to a single client to make them all "share the same
work", if you use multiple clients/wallets though, theyll all have
their own work todo
11:14 < JudStephenson> Would you be more likely to generate blocks using multiple remote
miners or multiple clients / wallets?
11:14 < OneFixt> Equally likely.
= miner network =
11:15 < gp5st> i'm sure this has been thought of: working with folding or seti or w/e
@home and having something come out of the expenditure of cycles?
11:15 < [Noodles]> we've got something coming out
11:16 < [Noodles]> a secure network
= ATI cards =
2010-Dec-5
13:15 < xelister_> http://en.wikipedia.org/wiki/Comparison_of_AMD_graphics_processing_units
13:17 < ArtForz> 6870 ~232Mh/s, 5850 ~240
13:19 < lolcat> ArtForz: Isnt the 6870 much faster?
13:20 < [Noodles]> one might think so, but it's not
13:20 < ArtForz> for 3D, yes, for mining, no
13:21 < lolcat> Then I will get first one 5850, and then if it pays off, Ill buy another
5850 to double the speed
13:22 < ArtForz> 6870 has more ROPs and fixed-function hardware than 5850, which improves 3D
perf quite a bit, but raw shader perf is a bit < 5850
13:23 < lolcat> What would be the best one to buy for mining? price vs capacity, and wich
would have the fastest speed possible in a single card?
13:23 < ArtForz> fastest affordable single card is still HD5970
13:24 < ArtForz> fastest card period is the various AIB 5870x2 cards
13:24 < [Noodles]> i think i'm ordering a 2nd 5850 tomorrow, i'll need another PSU too though
13:24 < ArtForz> if you have to pay for the rest of the hardware 5970 is a good bet, for
card alone best Mh/$ is still 5770
13:25 < ArtForz> I dont have any probs
13:26 < ArtForz> just stay away from SDK2.2
13:27 < xelister> on catalyst 10.09/.10 (ubuntu), it generates at 50% speed (it reports 530,
but half of blocks are invalid)
13:27 < ArtForz> that sounds like crossfire is enabled
13:28 < ArtForz> I run cat 10.9 and 10.10, no invalid blocks
13:28 < ArtForz> sdk 2.1
13:28 < xelister> ArtForz: I run some shitty 10.10-beta from ubuntu
13:29 < ArtForz> and I'm pretty sure I'd notice losing half hashrate
13:30 < ArtForz> 5770 = about 150Mh/s
13:31 < lolcat> 5850 = about 240Mh/s ?
13:31 < ArtForz> yep
13:31 < Lysacor> Sure beats my Nvidia GTS 250 at 32Mh/s...
13:31 < ArtForz> 5870 310, 5970 530
13:31 < ArtForz> GTX580 ... 120
13:32 < ArtForz> 5750 115
13:32 < ArtForz> kinda funny, nvidias biggest card is a tad faster than a 5750 ...
13:32 < Lysacor> and too expensive :P
13:33 < ArtForz> and 250W... 5750 is 85W
13:35 < lolcat> Does the mining require a lot of transfer speed between the gpu and mobo?
13:35 < ArtForz> nope
13:36 < lolcat> Would x8 work?
13:36 < ArtForz> I got a 0.4% drop from PCIe 2.0 x16 to PCIe 1.0 x4
13:36 < [Noodles]> Lysacor: i'm running m0's on windows, cc10.10 sdk2.2, nothing special, no
probs
13:36 < lolcat> Woho, then I could stick 4 of them in my new computer xD
13:36 < ArtForz> most of my 5970s are running at x8
13:37 < ArtForz> AMD [7|8]90FX chipset can only do x16/x16, x16/x8/x8 or x8/x8/x8/x8
13:38 < ArtForz> = all my quad 5970 boxes are running at x8
13:41 < xelister> 5970 + ubuntu 10.10 x64 + catalyst/fglrx from ubuntu + sdk = FAIL, just
215M (generates 50% of coins wrong out of reproted speed 530M)
13:41 < ArtForz> = tried setting DISPLAY and running a seperate miner for each GPU?
13:41 < xelister> 5970 + ubuntu 10.10 x64 + catalyst/fglrx from ubuntu + sdk2.1 + daiblo
miner(git) = FAIL, just 215M (generates 50% of coins wrong out of reproted
speed 530M)
13:41 < xelister> same on sdk2.2, just it easts 400% cpu time
13:41 < ArtForz> 2.2 = FAIL with any multiGPU setup
13:42 < ArtForz> I don't rememebr how diablos selects GPUs for mining
13:44 < ArtForz> does it allow you to specify a OCL device index to use?
13:44 < xelister> it runs on all gpu
13:45 < ArtForz> yeah, thats asking for trouble
13:45 < xelister> why
13:45 < ArtForz> beacuse ATIs openCL runtime is retarded
13:46 < ArtForz> run a seperate miner process for each GPU
= ASIC =
2010-Dec-10
15:09 < ArtForz> if difficulty keeps rising like this even payoff of a 5970 @ $0.10/kWh is uncertain
15:12 < ArtForz> and at $.10/kWh that 5970 costs about $25/month in power
15:12 < ArtForz> yeah, I'm down to 1250btc/day
15:12 < ArtForz> at least until my next 12 5970s arrive
15:12 < appamatto> that's $250 a day
15:13 < ArtForz> then I should be back at ~ 2kBTC/day
15:13 < ArtForz> well... it's also >$20k in hardware
15:17 < ArtForz> I'm hoping total hashrate keeps rising
15:18 < MT`AwAy> it should
15:18 < MT`AwAy> especially if services around bitcoins grow more
15:18 < appamatto> ArtForz, why?
15:19 < ArtForz> because my next design is pretty largescale
15:20 < ArtForz> ASICs
15:21 < appamatto> ArtForz, not an FPGA?
15:21 < ArtForz> nope, FPGA is too inefficient
15:21 < ArtForz> barely beats GPUs on hash/W and sucks for hash/$
15:21 < appamatto> So an ASIC will need some kind of chip fab?
15:22 < ArtForz> yep
15:23 < ArtForz> well.. if you want to see a single miner with >66% of total network hashrate ...
15:24 < appamatto> do you have the ability to do that?
15:24 < ArtForz> right now? no
15:24 < ArtForz> I already got a design that works in simulation
15:28 < ArtForz> deciding if it's a good idea to invest about $200k in it
15:29 < ArtForz> fixed costs for a production run are pretty high
15:33 < appamatto> ArtForz, 50% of bitcoins over 3 months would be ...
15:33 < ArtForz> 324kBTC or so
15:34 < appamatto> seems like you'd need about half a year averaging 50% to recoup
15:35 < appamatto> But you could probably sell some units as well
15:35 < ArtForz> more like 9 months, it still uses quite a bit of power
15:35 < ArtForz> but should beat even 5970s by a factor of 5 or so on hash/W
15:35 < appamatto> you could sell units at a price that would offset the cost to make and the cost
they would exact on your mining operation
15:36 < Kiba`> so a gradual distribution of GPU machinery to the population at large?
15:36 < appamatto> on the other hand, you might raise the price of btc because you are making the
system more secure
15:37 < ArtForz> the problem is, it beats GPU mining by a large margin for hash/W
15:37 < ArtForz> so by throwing something like this at the current network it'd make GPU mining
(while paying for power) unprofitable pretty much over night
15:42 < ArtForz> well... a single board would be about 2Gh/s and ~200W
15:42 < tcatm> How much would such a board cost?
15:43 < ArtForz> err... $2k
15:43 < ArtForz> I scaled it until I ended up at about same hash/$ as a 5970
15:45 < ArtForz> it's a fully pipelined SHA256 core
15:48 < ArtForz> completely custom, somewhat inspired by deep crack and COPACOBANA
= more ASIC =
13-Dec-2010
15:49 < ArtForz> 300W per 5970
15:50 < fabianhjr> ArtForz: you got 33 5970s? 0_o
15:50 < ArtForz> no, I only have 24
15:50 <@theymos> Are you going to demand extra fees? With 65% of the network, you'd
be able to force everyone to pay at least 0.01.
15:50 < ArtForz> I was talking about machines with same power draw as my planned
$200k miner
15:51 < ArtForz> for the same money cou could buy 300 5970s, which would do about
160Gh/s and need 90kW
15:52 < ArtForz> erm, with a mjority of the network I could simply hog generation
15:52 < ArtForz> = 7200btc/day
15:53 < Kiba> what if your machine got confiscated by the government and used to
hijack the network?
15:53 < ArtForz> bad luck I guess
15:53 < da2ce7> lol with an majority, you could exclude anyone who generates bolcks
other than yourself!
15:53 < ArtForz> yep
15:54 < ArtForz> why do you think I've suddenly stopped adding GPUs when I was ~33% [Noodles]
of total hashrate?
15:54 < nanotube__> ArtForz: are you really about to invest in 200ghps of processing achristi~
power?
16:02 < ArtForz> ASICs
16:03 < ArtForz> ever heard of deep crack?
16:03 < ArtForz> EFFs DES bruteforce cracker, about 1800 custom ICs
16:04 < ArtForz> I'm pretty much intending to do the same thing with bitcoin-sha2
16:04 < ArtForz> about same hash/$, 8-10x better hash/W
16:05 < ArtForz> = a machine matching the current network would need about 5kW
16:05 < Kiba> ArtForz: so you're implementing much more efficent form of mining? altamic
16:06 < ArtForz> pretty much
16:06 < ArtForz> onyl problem is it doesn't scale down
16:06 < ArtForz> even at 1k chips fixed costs dominate
16:07 < Kiba> are you going to buy a wind mill? [Noodles]
16:07 < Phoebus> ArtForz, did you study/are studying engineering somewhere? AAA_awri~
16:07 < nanotube__> ArtForz: how about you collect investment from other community achristi~
members, then distribute the chips pro-rata?
16:07 < Kiba> he's a 37 years old fart
16:07 < ArtForz> Phoebus: yup
16:07 < nanotube__> ArtForz: that way you get more volume, lower cost. and as a side
benefit... you don't end up with >50% of the network
= new ati card=
16:23 < fabianhjr> ArtForz: Are the new 6000 GPUs worth it?
16:24 < ArtForz> not really
16:24 < ArtForz> and 69xx will probably suck even worse
16:26 < ArtForz> I got a 6870, not worth it
16:27 < ArtForz> 235Mh/s, 150W, doesnt have much OC headroom
16:28 < ArtForz> and you need sdk 2.2 for it
16:28 < ArtForz> = say goodbye to decent multiGPU performance
= nvidia telsa =
13:54 < ArtForz> DP: tesla C2050 = 257.6 GFLOPs, HD6970 = 337.9GFLOPs, HD5970 =
464GFLOPs
13:55 < ArtForz> and the 5970 can do another 1856GFLOPs SP on the XYZW units in
parallel to DP on the T unit
= artforz summary =
09:11 < ArtForz> 16.5Ghash here
09:13 < ArtForz> 5 HD5770, 24 HD5970, 1 HD6870, 1 HD6970, for a total of 55 GPUs
09:13 < Sultan_> you lot have 5 computers?
09:13 < ArtForz> 9
09:15 < ArtForz> with a total power draw of ~9kW
= network hash rate =
< donpdonp> is it possible to estimate the gigahash rate of the entire network?
< ArtForz> estnextdiff 17571.2
< ArtForz> so thats 17571.2 * 2**32 / 600 hps or... 125.78 ghash
< donpdonp> whats the / 600 hps for?
< ArtForz> 600 seconds nominal per block
(one block generated every 10 minutes)
= PS3 =
08-Jan-2011
< ArtForz> rooted PS3 should have peak ~24.5Mhps
< ArtForz> = all 7 SPEs + PPE
= artforz bank account =
08-Jan-2011
< newsham> also are you keeping a large stock of btc or liquidating most
< ArtForz> I kept about 15%
< ArtForz> ~40kBTC currently
< ArtForz> I'll probably keep on selling 80% of daily production or so
< molecular> who is buying all these? people speculating? I don't understand where the demand stems from?
< ArtForz> I'd guess so
= cartel / majority =
15-Jan-2011
< ArtForz> once a mienr (or a bucnh of collaborating miners) get > 50%, they can claim all generation and block transactions from getting into the chain
< molecular> on the other hand: if all the gamers start mining (some hype might happen), artforz's asics dont look so good any more
< molecular> ArtForz: is that your plan? take over btc and therefor kill it? can't be, right?
< ArtForz> I prefer to add hashrate slowly to stay at 15-25% of total network
< ArtForz> it simply doesnt make much economic sense to grow beyond 33% or so
< ArtForz> you're effectively competing against yourself
< davex__> artforz: why does that matter? it would be like having a monopoly on all gold mines.
< ArtForz> except you dont
< ArtForz> lets say rest of network is 100Gh/s, you add 10, you're 9.1% of total, you add 50, you're 33.3% of total
< newsham> what artforz is not counting on is that once he has large monopoly, he will be able to extract nice transaction fees if he so desires :)
< ArtForz> so ~ 3.5x payout for 5x the hardware
< davex__> artforz: yeah true... i guess it's also not reasonable to assume when you get to 50% everyone else will just give up, also.
< ArtForz> yep
< ArtForz> and a mining cartel is really really obvious
< davex__> because you get lots of simultaneous solutions submitted?
< ArtForz> that and nobody else suddenly gets blocks
= ATI Driver rev =
18-jan-2011
ArtForz> for 5xxx, anythong >= 10.10 seems to work well
sipa> but it's not like the stream sdk where you shouldn't use the latest version
ArtForz> nah
ArtForz> I run 4*5970 on 10.12 no problem
ArtForz> seems to be even a bit faster than 10.10
ArtForz> btw, the multiGPU slowdown and CPU hogging is limited to OpenCL in sdk 2.2/2.3, CAL works fine
15-feb-2011
I mean, 6970 is pretty exactly as fast as 5870 for me
ArtForz: 6950 .. would you recommend it? instead of 5870
well... for 3D, yes, for mining.. no
retsilex: 6xxx is useless for mining
you need sdk2.3 for 69xx, which has broken OpenCL multigpu support
retsilex: you're stuck with sdk 2.2 or 2.3, which is shit
so unless you have a CAL miner, 69xx sucks ...
retsilex: 6970 should be 15% faster than 5870
but its not
because the sdk fucks you over
ArtForz: your desktop uses which sdk, which drier version? and what linux?
debian sid, 2.6.32, fglrx 10.9, sdk 2.1
ArtForz: i heard you were the guy to ask about the reason for the drop in stream processors between 5970 and the cards that come after :P
foucist: pretyt simple really, 5xxx had pretty bad shader ALU use % in games on average
so 6xxx got more fixed-function ahdrware, beefed up frontends, better mem controller, ... but shader ALU * MHz went down