Topic: The fastest HD 69xx miner. 250 BTC. - page 2. (Read 70768 times)

I don't know what speeds they're getting at what temps but I'm able to do 2.43GH/s with 3 6990s at 1300w (10.83A) after 15 mins when the fans have been running at 100% for some time and temps are higher.  Temps are all below 80C on air cooling on all cores except for a defective one that's always 12-20C higher than the other core on the same card (I limit the overclocking on that core so temps remain in the low 80s; perhaps I only need to reapply some thermal paste on it but I'm RMAing it to avoid any warranty issues).  At times (probably when the AC kicks on), temps will be in the high 60s to low 70s.  The thermostat is in another room and set to 76F in a very hot climate (currently over 100F outside) so it'll go on and off at various times.  This is with an enclosed HAF 932 case with one 200mm fan removed and 5 120mm fans added.

What seems to makes a big difference is the PSU used.  10.83A on a 15A breaker (it's actually on a 20A breaker but I'm treating it as 15A for safety reasons) means I still have 1A to overclock it to possibly 2.50GH/s.  I've also got plenty of room to add more cooling if necessary to bring the temps down.  I'm comfortable with them running in the 70s and spiking into the low 80s for a short period of time.  I've used a cheaper PSU and could only do 2.38GH/s while pulling over 1380w.  I use two PSUs but only swapped out the main one.  Mind you, the better main PSU isn't high end at all so if you're willing to pay a hefty premium, perhaps it'll be able to do better.  The secondary PSU is on the cheaper side so that could be improved as well.  The extra fans I'm using are cheap Yate Loons and not something expensive like the jet propelling Deltas that can pull in over 2x as much air flow.  I don't even have a big external fan blowing on it.

I'm also running Linux, which means I'm more limited on the things I'm able to tweak (but with the custom tools I've made, I think I'm very close to what someone on Windows is able to do).

If your customers are getting the same rates or less than what I am with those constraints, they're spending a lot of money for things they can tweak on their own.  At 1300w for 2.43GH, that's an efficiency of almost 1.87MH/J for the entire machine.  When the time comes, I can make it go slower and be above 2.1MH/J (likely more) for the entire machine or above 2.4MH/J per card.

Also, assuming that hdminer is 6.3% faster, that'd mean they need to be doing at least 76.2GH/s to do better than spending the same amount of money on extra 6990s.  I agree with Jack of Diamonds that you're pricing yourself out.  If people do the proper research, they wouldn't be paying 250 BTC, let alone 100 BTC for something that is 1.2-2% faster. At 50 BTC and 2%, you'd want to do at least 49.7GH/s to do better than adding the same amount in extra hardware.  I also suspect that it being faster means it will probably use a little more power, though not necessarily in the same proportion as the increase.

From comparing the speeds I'm getting to hdminer, the miner isn't 6+% faster, it's only 1.2-2% faster.  I'm using Catalyst 11.4 without downgrading and underclocking the memory so the difference could be even less than that (11.5 is supposed to give it a few Mh boost and not underclocking it would increase hashing rates by 0.5-1.5Mh/s).

Here are the rates (per core) I get at various speeds in the overclocked position using the phoenix 1.48 miner with the poclbm kernel:

880 - 379.8
900 - 387.0
905 - 389.6
910 - 391.9
915 - 393.4
920 - 396.0
925 - 398.2
950 - 408.5
955 - 410.8
961 - 412.6
966 - 414.6

Those numbers are near the higher end of the spectrum but not the very top.  The numbers are sustainable and not just one time blips that timing/thread switching might inflate.  At 880, hdminer was actually 1.8% slower (759 vs 746).  For the 915 and 960 settings, hdminer was faster (@915, 786 vs 802 = ~2.0%, @960, 824 vs 840 = ~1.94%).  These are using numbers that would favor hdminer more (lower hashing rates than optimal and rounding my rates down).  More optimal numbers would probably put it closer to a 1.2% difference.

Using $17 as the value of a BTC, 250 BTC = $4,250.  If the cost of a 6990 were $740, that'd be 5.74 6990s.  Assuming hdminer was 2.0% faster, you'd need over 295.6 6990s to make a profit.  So unless you're running an operation that exceeds 248GH/s, you shouldn't buy hdminer.  I'd be surprised if there are several people with that much capacity, let alone even one.  It'd be even worse if hdminer were only 1.2% faster.

Do you want to make money by adjusting price, or stubbornly price a 6.3% performance increasing miner at BTC worth $2,000 dollars (or two full rigs with 2x 6990) ?

I operate ~40 GPUs in datacenter and wouldn't pay even 20BTC for this software based on current exchange ratio to fiat.

Your software is good.
Don't ruin your chances to profit from it by setting a ridiculous price or refusing to accept current economical realities; It will take years to pay off the investment even for huge clusters.

THIS

A while back, I recall seeing an example CAL program for free Linux...

memory speed and core voltage are both stock (default bios switch)
settings are -v -w128, the standard for 6xxx series

Settings (switches like -v -w 128...), memory speed, core voltage?

I would be more interested in a comparison with hashkill though...

Using poclbm (guiminer) and the newest catalyst (11.5) and app sdk (2.4), I am getting 711 mhash on stock settings - 830mhz, bios switch in the default position
if hdminer only gets 708 mhash at stock, wouldn't it actually be a performance regression to switch to it?

My guess is you have no interest in trying to reproduce the 680 MH/s (also verified by jgraham), because they significantly reduce your claimed benefits.

My thoughts exactly.

AFAIK it is not yet possible to mine without fglrx (until someone works on this bounty).

Does this support free Linux (ie, no fglrx)?

Yep you are missing something.

hdminer's target market is not you the home user, but cluster owners who operate 20+ GPUs and are power or cooling constrained (eg. power circuits at 80% utilization per the US National Electric Code), yet need that +6.3% perf improvement without increasing power consumption or cooling needs.

I get 810 MH/s out of a water-cooled 6990 running at 950mhz using poclbm.  For 250 BTC I can buy a nice 2 X 6990 fully water cooled rig that'll give me 1620 Mh/s.  I suspect I must be missing something here and would appreciate some enlightenment

It shouldn't. It should go up as the block size increases.   This was part of the design to account for increases in performance.

Couple of things about this super snazzy software.   I'm using vanilla poclbm and each of my 6990 GPUs (stock clocks) seem to be registering 339 and 340 MH/s depending on when I look at them (flags are -w 64 -a 5 -f 1).   This is with 11-5 drivers and SDK 2.4 - so assuming that these have the same problem that 11-4 has this makes hdminer a less than 1% advantage.   Hard to cost justify the $4000 it would cost me today to get it.  This is assuming it gives any real advantage at all. (cue shock and awe)

Consider for a moment that all we appear to be doing here is comparing the MH/s statistics output by these two programs.  
Anybody else see how we could just as easily be seeing some kind of measurement error?  I'm not trying to be a jerk but really small gains should always be viewed with the utmost skepticism.  So if you're a fan of Ioannidis the corollary to "Most published research is false" would be "Most small gains are noise" ;-)

Aleged +6.3% improvement is a better term.   I find it amazing that people who do software development for a living still are kind of retarded when it comes to math.


Sorry for the continued edits.  I'm still a noob by board standards.  I just tried your settings and now I'm within 2% of the hdminer's scores for clocks of 915/1260.  A very similar gap to what I came up with under stock clocks.   Your observation is correct and it's contrast with mrb's numbers illustrates a kind of a pet peeve of mine. mrb made a reasonable assumption.  Poclbm and other programs have overhead (python interpreter, OpenCL) before they can even get the code executing on the GPU.  So writing the whole app "closer to the metal" seems reasonable.  So he re-wrote it and saw an increase and....assumed that ALL of that - down to one tenth of one percent - was due to his optimizations.  In other words he is assuming that beyond that ALL VARIANCE is at 1/100 of a percent or less.   Which is the kind of thinking that strains if not breaks a sanity check.  Generally speaking it's hard to get a general case which is that accurate.

Not to mention things like Ahmdal's law which IMHO is just a way of saying that optimizations are rarely as fast as you think they are. ;-)

Why would difficulty stay constant?  

I don't have the time to crunch the numbers, but you should assume anywhere from a 10% - 50% increase in difficulty which is compounded roughly every 2 weeks. Those kinds of diminishing returns should certainly be factored in. A 6990 may not be worth running in the near-to-distant future due to difficulty increases, thus rendering the miner moot, and perhaps making the investment a net loss.

Yes. As you guessed, my buyers are large miners. (I wouldn't trust that 680 Mhash/s number -- couldn't reproduce it.)

grue: $200, not $20. As I explained earlier, it makes sense for my price to be indexed on BTC, not USD.

At stock:
Your miner numbers: 30 days 12 hours=732 hours per block
Your Phoenix numbers: 32 days 11 hours=779 hours per block
According to another guy's 6990 at stock, 680 MH/s: 31 days 18 hours=762 hours per block

People buying/using your miner will need 250 BTC=5 blocks of extra work to pay for the investment in your miner.

With your Phoenix numbers:
1/732 blocks per hour*x-5 blocks=1/779 blocks per hour*x
x=60662.55 hours=6.925 years

With 680 MH/s:
1/732 blocks per hour*x-5 blocks=1/762 blocks per hour*x
x=92964 hours=10.61 years

Congratulations, you just made a nice investment that takes more than 6 years (or 10 years) to make back, if you're going with single 6990 (assuming difficulty stays constant)!
To be able to recuperate the costs under 1 year, you'll need to be running 7x 6990s (or 11, according to another guy's 680 MH/s stock 6990).
That's not counting electricity bills.

your original price was 400 btc, which was less than $20(February prices). now you're selling it at 250, which is approximately $2100.

I am slashing the price down to 250 BTC.

I measured my miner at 802 Mhash/sec on a properly overclocked Radeon HD 6990 (BIOS switch at "overclocked" position 1; with "aticonfig --odsc=915,1260" to further overclock the GPU to 915 MHz and mem to 1260 MHz). And 840 Mhash/sec at 960 MHz, although the hardware wasn't very stable at this speed.