Topic: 1GH/s, 20w, $700 (was $500) — Butterflylabs, is it for real? (Part 2) - page 3. (Read 146880 times)

Who's beating a dead horse? Is it you Syke?

I've tested the wattages (via KillAWatt) for my undervolted 2x5970 rig with 0-4 GPUs mining.

Num GPUs

Total Watts

0

175

1

230

2

285

3

340

4

395

At my current settings, each GPU eats 55 watts and generates 530 MH/s. That's roughly 4.82 MH/W. I could probably increase the clock a little bit, and use a more efficient power supply to reach the quoted performance.

A 7970 is significantly more efficient than a 5970.

Damn president's day holiday! Probably woulda been today otherwise. 

A little birdie said my first single would be flying out to me tomorrow.     

The deadline is end of day Friday (PST). There were whispers in the #bitcoin chatroom that someone had received a device on Sunday, but I haven't heard anything since.

I dunno how much info can be gleaned from a tracking number, but is it OK to PM it to a mod, who can then post and say it's legit? For 13 BTC, I would do that and an unboxing + setup video, if I ever get a tracking number (and a single) in the first place.

Who is going to collect that 13 BTC bounty? I thought we'd have something for sure this last weekend.

To collect, we'll need a tracking number posted into this forum AND an un-boxing video.

As one of those enterprising individuals, I'd like to speak up here. BFL's bait-and-switch certainly did hurt us. The only reason we got anything made was because we decided to go with our gut and ignore the threat of that mythological device actually becoming a reality. We pulled that trigger back in October, well before your test in mid-December showing that they had completely missed the mark. Making the decision to pour tens of thousands of dollars into a production run that could very well go unsold, or need to be sold under cost to recoup some of the investment, was not an easy one. The weeks before your test were very stressful, believe me.

So yeah, they didn't dissuade us, but any rational person really should have been. At the very least, delaying the order until more information was available. Fortunately, the laws of physics held and they didn't reach the power efficiency that is impossible with 65 nm FPGAs (although the fact that they used 65 nm FPGAs was not even clear back then). They did dissuade us in a way though--if BFL had provided honest specs from the beginning, we certainly would have gone ahead with a larger order and been able to reduce the cost.

Fortunately, our product is better. No, we're not planning to copy their design. They used old FPGAs, and get a fraction of the power efficiency. Why would we want to copy that?

For reference, O_Shovah measured 15.64 W at 360 MH/s or 23 MH/J for the X6500 (see this post). I've measured about 18 W at 400 MH/s.

Another numbers run for the 10h39m period ending at 1:45am CST; average hashrate continues to drop:

start: 15:06CST (yesterday), shares 79630
end: 1:45CST (today), shares 85418

So:
1:45am - 15:06pm = 10h39m = 38340s
85418 - 79630 = 5788 shares
5788 shares * 2^32 hashes per share = 2.49e13 hashes
2.49e13 hashes / 38340s = 648Mhps

Hell yeah he's right.  If he can show me his Fluke 375, hooked up in the configuration he describes and reading an accurate power draw, the system hashing away at 97w, you bet your ass he's right and I will stand mother fucking corrected like a bitch.  I fully admit it.  Guess that might require video, though... but that seems a bit unreasonable.  

I may try the experiment myself, but like I said, I'm sure my numbers would be considered biased, which is why the burden of proof is on him.

I also don't have a Fluke 375.  This makes me sad.

So ... you're saying that if he did that then he is right?

"Clamp meters are inaccurate" - Could you compound your stupidity by making a more broad and inaccurate statement?

The meter I used is a Fluke 375 but for high current DC work a much cheaper meter would be more than accurate.

I separated and clamped the six 12VDC conductors on the PCIe power connectors.  I also used a extender cable which I cut to separate the lanes used for 12VDC so they could be clamped.  That is pins #1, #2, #3 on both A & B side of PCIe slot (technically A1 is used for presence detect but it is micro current so a rounding error on card load, trying to slice off just that lane isn't worth the hassle).   I measured current across all 3 power connections multiple times over a 5 minute span.  I measured the voltage on the 12VDC rail using a power cable which I verified was connected to the same rail.  I had gpumon & munin running and made sure card load was 99% over the entire test period.

I then double checked the results by looking at complete system power at the wall using Kill-a-watt and compared that to the tested PSU efficiency.  I also measured the power (at the wall) at idle w/ no GPUs installed (headless boot) to get the non-GPU load for the system.

97w is bullshit, though.  You aren't mining at 550 MH/s at 97w on a 5970.  Sorry.  I already told you why the test Mousepotato did was flawed, therefore your 5 MH/w is also flawed.  Retest with a proper setup and report back... I would do it, but I'm sure my numbers would be called into question as being biased.  Put a 5870 as your primary card and your 5970 as a secondary card.  Measure your idle, then fire up the 5970 and measure power consumption, and leave it running for 24 hours.  I will be glad to retract my statement if you can produce 97w at 550 MH/s for 24 hours.


That's ok, at least I'm joining you at the bottom, since you achieved that pages and pages ago... but I digress.  Please describe how you verified this at the card level with a clamp meter... I wait with baited breath for you to describe the machinations you went through to achieve this marvel of electrical engineering.  I also question how someone with your apparent knowledge of things technical, somehow thinks that using a clamp meter is an accurate measurement of power.  Clamp meters are notoriously inaccurate unless you have a fairly expensive one, coupled with the fact that I wait to hear how you used a clamp meter "at card level," again, complete bullshit.  Though I admit, I've not tried to measure a DC load with a clamp meter... it will be interesting to hear how you achieved this with not one, not two, but THREE separate conductors that need to be measured to check the load "at card level."  In DC no less...  It will be a mother fucking breakthrough of epic proportions... in fact, I think Fluke will buy your design for a tidy sum.

So please, do tell... how did you achieve this?  What clamp meter did you use? When was the last time it was calibrated, assuming it actually is a true RMS meter.   What, exactly, did you measure with this clamp "at card level?"  It's amazing how much bullshit you've been able to pack into a few sentences, you should market that ability.

It's very possible to get 5MH/W out of a GPU. The reason why people don't do it much is that space is usually limited. In that case, you'd rather have 1 rig than 2 more efficient undervolted rigs. I think the main reason these look attractive (to some) is the space/convenience factor. Otherwise, on a $/mh/s or $/mh/w basis they aren't that compelling.

I can't help that you can neither read nor follow a subject.  Undervolting involves lowering voltage however that requires lowering the clock also.  The goal is to increase system MH/W, not keep maximum clock and magically get something from nothing.

It is possible to acheive 5MH/W on a 5970.  I have verified it myself at the wall using kill-a-watt and using clamp meter at the card level.

Nobody (I mean nobody on the entire forum in any thread including this one) claimed you can get 750MH/S @ 150W.    NOBODY.  Since you are easily confused I will say it one more time.  Nobody but you was that stupid.  Nobody.
 
I don't care if you are pissed.  
I don't care if you are too stupid to read a thread properly.
I don't care if you don't believe.

Your ignorance doesn't really affect the efficiency of undervolting does it?

You have progressed past annoying into being downright stupid.

So I just read that whole thread, and now I'm pissed.  Basically, either you just lied to me and I wasted my time, which is why i'm pissed, or I completely missed the post where someone achieved 750 MH/s at 150W on a 5970.  The best I saw was 550 MH/s at 250w... that's not even close.  NOT. EVEN. CLOSE.  So please, point out where in that thread someone achieved 750 MH/s at 150w with an undervolted 5970. I'm all over the ability to run a 5970 at 150w and I will gladly retract my statement.  After you do that, if you can produce the link that I missed, I want to see it run stable for 24 hours at that voltage and hashrate.

To add insult to the injury of that thread, it was performed with a single 5970's to measure the wattage.  This is a false reading - you can't compare mining wattage with idle wattage on the primary card, since the primary card is engaged in the display, thus increasing it's "idle" draw.  You have to perform it on an inactive card.  You need two cards in the system.  Measure idle, crank up the miner on the second card and measure mining wattage to get an accurate power draw per card. Additionally, I would like to see any of those numbers stable for 24 hours.  I very seriously doubt the system would remain stable for long, making the whole experiment moot, other than to show the card will run briefly at those voltages. Hell, I can OC my CPU to 6 GH/s for a benchmark run... doesn't mean I have a magical 6 GHz CPU either.

I call complete bullshit until then.  I call making up numbers to make the argument possible.  I call the same tactics BFL is being accused of used to bolster a counter argument.


I didn't say you did, I quoted Yochdog.

Overnight Icarus was getting
(5s):10.2 (avg):357.7 Mh/s | Q:7671  A:3818  R:11  HW:0  E:50% U:5.36/m
and the U: figure suggests the Hash is either correct or even low (5.36/m if calculated directly gives 383.7MH/s)
So using the number at 19.5W, 357.7 Mh/s is 18.3MH/W so certainly not far from your number there.
(He had it plugged into his MIPS router ... running cgminer 1% CPU ... so not much extra power there either )

Anyway, yeah, 20MH/W +/- 10%

(383.7MH/s is 19.7 MH/W)

Yeah but it may end up doing more than 700 MH/s.  Was just trying to grab a middle of the road figure.

Best case scenario it does 850MH/s sustained @ 85W = 10 MH/W
Worst case scenario it does ~700MH/s sustained @ 90W = 7.8 MH/W


To look at it on a larger scale.

CPU Rig (i5-2600)  ~0.2 MH/W
Casual GPU Gamer rig ~1.5 to 2.0 MH/W (single graphics card, high end CPU, a gaming rig used for mining)
High efficiency GPU rig ~2.5 MH/W  (sempron, minimal build, linux on usb drive, 80Plus-Gold PSU, 3x5970)
Underclocked GPU rig ~5 MH/W  (i.e. 3x 5970 550Mhz @ 0.8 VDDC)
BFL Single Worst Case ~8 MH/W (excluding host system power draw)
BFL Best Case ~10MH/W (excluding host system power draw)
Ztex FPGA Boards ~ 20MH/W (and other Spartan-6 based rigs, excluding host system power draw)
28nm "next gen" FPGA ~40MH/W (guestimate based on die-shrink of 40/45nm designs)
28nm SASIC ~60MH/W (guestimate based on power savings due to reduce gate count going from FGPA -> SASIC)
Custom 65nm ASIC  ~100MH/W (based on "testbed" processor for SHA-2 testing)
Custom 45nm ASIC  ~200MH/W (Moore's law applied to "testbed" processor)
Custom 28nm ASIC  ~400MH/W (Moore's law x2)

Note: each data point is likely upper limit in its category and likely is overly optimistic but provides a rough estimate for SHA-256 efficiency.

There are at least 3 FPGA miners shipping today, with ~2X the MH/W of BFL:

BitcoinFPGA

If BFL takes off and starts driving difficulty - no worries - all of these units will still be profitable.
If these Spartan6 designs begin to drive the difficulty - they will make the BFL units unprofitable.

It doesn't matter if it was cheaper per MH up front, once it's losing money every month.

-rph