Bitcoin Forum

 ::) ::) ::)If undrervolt
Or under clock
S1 how much will be the power of watt per gh??

any details

underclocking, you could probably get the S1 to be stable in the area of 150GH/250W. any lower and you need to undervolt it as well.

undervolting requires modification to the resisters associated with the TPS53355 chip(s). Increasing the resistance slightly (opposite of a pencil mod) would help drop the voltage from 1.1V to 0.9V which is where you would be closer to 150GH/200W

underclocking, you could probably get the S1 to be stable in the area of 150GH/250W


ok how i can do that any steps by steps details or link/?

Increasing the resistance slightly (opposite of a pencil mod) would help drop the voltage from 1.1V to 0.9V which is where you would be closer to 150GH/200W

Im a bit confused why you simly repeated what i posted as two seperate posts.

underclocking is simple, opposite of how you overclock - SSh into the unit, and change in the hex variable you want. If you go too low, harware errors may start to spit out all over. I think you could do 300MHz (~160GH @ 260W if i had to guess), not sure how much lower than that though

I think he wants to undervolt primarily not underclock.  Some units (if they are like video cards) run fine on less voltage and therefore use less power for the same hashrate. 

Underclocking and undervolting at once could make these units use far less power which would be great for the long term and if you are power limited. 

I am assuming the voltage control is purely in hardware and not available with software so you would have to change SMT resistors. 

So say I buy 6 Ant S1s, which is about the same price as an Ant S2. Underclock to 160GH/s, and pull ~200W each.

6x 160 = 960GH/s
6x 225 = 1200W

Not too different from the 1000GH/s and 1000W the S2 is doing. Cooling would also be a lot easier than normal, as the chips would be running much cooler.

So what's from stopping people from buying the shit out of those S1s and undervolting?

Someone figuring out how to do it is the first hurdle.

BM1380 chip specs:

	Vlotage(V)		Hash Rate(GH/s)		Total power(W)		J/GH
	0.75		1.60		1.07		0.68
	0.85		2.00		1.69		0.85
	1.00		2.50		2.82		1.13
	1.10		2.80		3.70		1.32

It would be great to figure out how to undervolt S1 properly to get the best power efficiency.

Wait, I thought the BM1380 was for the S2? Do the S1 and the S2 use the same chip? Why did I think they were using different chips? I've been sick this week, and a little out of it, so my apologies for any confusion.

Has anyone tried just reducing the clock speed and measuring watts used at the wall?

Since overclocking increases power consumption, it stands to reason that underclocking would reduce power consumption, wouldn't it?

Undervolting is aimed to reduce power consumption while retaining the same hashrate. This doesn't do that.

If Bitmain can't even do that with the S2, I don't see how users would be able to do that with the S1. I believe Bitmain is simply using more underclocked 55 nm chips to obtain 1W/GH with the S2.

Yes. Over/underclocking gives a roughly linear result. If 180 GH/s draws 360 watts, then overclocking to 200 GH/s draws 400 watts, and underclocking to 150 GH/s draws 300 watts. These are just estimates, because things like PSU efficiency changes depending on power draw, and other little details.

any details how i can exactly do these???


 Underclock to 160GH/s, and pull ~200W each.

160 GH/s should draw 320 watts.

The file to edit is "/etc/config/asic-freq". ssh to the unit and edit that file. You can search the forum for more details.

150gh how much 200watt or not?

I'm gonna play armchair engineer for a few minutes, so bear with me.


Looks like DC/DC power conversions are somewhere around 35% inefficient on the S1 boards, and I'm guessing that's because of how hot the VRMs run. You've seen those thermal images that showed the VRMs as the hottest part? Well heat is nothing more than escaping (wasted) energy.  Then you add power for the controller board, fans, sensors, etc, and you're right at the 360W power draw a stock S1 pulls.

If you're undervolting, and pushing less voltage/power through the VRMs, I bet think they would not only run cooler, but also more efficient. Say you went with the slowest, most efficient voltage/clock setting, moved from 35% inefficient to ~20%, and then re-added the fixed power consumption stuff like the controller board and fan (running at a lower setting). That's almost the same 1W/GHs The S2 are doing.

Each S1 has 64 chips, 32 on each side, and they're using the bottom one, the least power efficient. I estimated the frequency of the chips assuming that frequency is directly linear with hashrate. Here's what happens if they were to use the other frequency/voltages in an S1:

Frequency(MHz)(estimate)	Voltage(V)	Hashrate(GH/s)	Power(W)(chip level)	Total Power(W)(estimate)
350	1.10V	179.2	236.8	355
320	1.00V	160	180.5	265
260	0.85V	128	108.2	165
205	0.75V	102.4	68.5	115

In fact, I bet that's what the S2s are doing. They're running hundreds of chips, at a low voltage, with better VRMs, inside a giant 4U case with huge heatsinks, and fewer fans. BAM there's your power improvements.

But none of this accounts for OC room. Lets say we went with the 0.75V @ 205MHz, and OC'd it to 245MHz. Each S1 would be making 125GH/s, and pulling 82W at the chip). That means @20% inefficiencies, and adding the fixed costs = ~130W. 8 ants would be making 1TH/s while only using 1040W (from the PSU, which means 1140W from the wall).

These how I can do at my antminter


320   1.00V   160   180.5   265

Yep the S1 and the S2 use the same chip.


Sounds great  :)
We need to find out somehow where that resistor is and what resistance it should be changed to to get 0.75V!

Due to the design of the unit I would imagine you would need to change eight SMT resistors.  A bit of a pain but interesting.

yeah my eyes are too old but it is doable.

 maybe 100gh at 100 watts could be done.

bump!

I don't see why you couldn't try a pencil mod on an S1. They just use a voltage divider to provide feedback of the output, so you'd pencil mod one to increase the voltage and pencil mod the other to lower it.

On the Antminer S1 that's R3 (8.2k) and R12 (9.76k) for the leftmost VRM, under the inductor. To make the voltage go up pencil mod (lower the resistance) of R12, and to lower the voltage pencil mod R3 instead.

Pencil mods are great for testing, or for minor bumps. According to my math, I would have to swap the 8.2k R3 with a 5.6k resistor, and that should bring the V down to 0.75? Anyone else want to double-check my back-of-the-napkin math?

This weekend I'd like to grab one of our S1s, underclock down to 200, replace all 8 resistors, and then see what the power difference is. After that, I'd like to see how far I can OC on 0.75v before the hardware errors go >1%.

pencil mods work well long-term, but be aware that they 'bake-in' and with age and heat the resistance lowers slightly from when first applied.  In the case of pencil-modding bitfury cards I often had a card run fine for 6-24 hours after a pencil mod before getting to a point that it couldnt stay stable for more than an hour at a time. Keep that in mind and leave a bit of headroom when testing a pencil mod.

Now that the right resistance/voltage is functioning on the bitfury, I am getting about 36-39GH per card

I tested our Ant, and R12 and R3 were both 4.45k?! This seems to be different than what you measured, and means to drop the voltage down to 0.78v, I gotta replace the 4.45k with a 3.07k resistor.

Hmmm. I was just going off the schematic, but it's possible they updated it. The same principle applies though.

how do you get the voltage of the asic with multimeter (on the board) ?

Anyone managed to change resistors. We need results.

I am also anxious to see some results.  This could be a game changer for profitability calculations by extending the life of our miners.  Once the revenue dips below energy cost, we can swap our resistors to improve the efficiency (at a reduced hash rate) and get many more months out of them.

Right.

I have succefully undervolt 3 antminer s1

the result are fucking awsome :

chip are at ~ 0.85 volt, freq is 08A1 (275 MHz)

~ 140 GH/s per antminer s1

500 watts at the wall for the 3 underclocked ant ! (166 per ant for 140 ghs) :)

The PSU is a corsair AX1200i

for comparison at stock volatage it's was drawing 1050 WATTS for 3x180 GHS !

I will make a topic to explain how later (I need time to see if it's working well)

It's running for 1 hour now without problem

the first ant  has 0.95 % of HW
the second ant has 0.3 % of HW
and the last has 0.5% of HW

https://i.imgur.com/Mnnj8Bn.png

Awesome!

What resistors did you use?

Also, compared to standard and overclocked how is the hw percentage rate changed? Better or worse? Or same?




.

The HW has increase (before I had 0.1 ~ 0.2%).

I used pencil mod on R3 (4.47kOhms => ~2.9kOhms)

Since its a pencil mod, just make sure the voltage remains the same, as pencil makes different resistance under heat.

Maybe with lower resistance give lower hw?




.

I made a quick guide https://bitcointalk.org/index.php?topic=526060.0

In this case, under heat you will get a lower reference voltage at the voltage divider. So your HW errors would likely go up if you maintained the same frequency.

This has the opposite effect of pencil modding to increase voltage, where you'd be putting lead on the other resistor.

Very awesome dude!
I wonder what 1v @ 325mhz will yield, it should be doable, I got a U1 @ that setup and its 0% HW

https://i.imgur.com/wOhy80a.png

420GH/s @ 500W even. That's lower than what I was expecting. I like!

140.8GH/s @ 275mhz means each chip is doing ~2.20GHs/chip. They say 0.85W should be making 2.0GHs/chip, which is a stock of 250mhz. You're OCing 25mhz (10%) above the spec for that voltage. 1.993A per chip x 64 chips, plus a 10% OC means the chips are pulling 119W @ 0.85V. I understand the voltage is slightly higher on the higher chips, so they'll pull more (more W, same A), but I'm averaging.

That PSU is 80PlusPlatinum, which means 92% efficient at 50% load. That means each Ant is pulling ~153W from the PSU. Add ~10W for the controller, and another 5W for the fan (cuz it'll be on low), and you're only looking at 14% inefficiencies from the VRMs.

Our OC ants pull 201GH/s @ 393mhz @ 1.11V. From the wall is 405W, on a 90% PSU = 365W from the PSU. 3.359A per chip (x((393/350)x1.11))=268W (from the chips). Add 10W for the controller, 20W for the fans (I'm guessing here). That means that the VRMs are 20% inefficient on our units.

If we were to go down to 0.75V, and do a 10% OC, you're looking at 112.6GH/s. You're looking at 75W chip draw. Add 14% VRM inefficiencies (although it could be lower than that), and you're looking at 87W draw from the boards. Add 10W for the controller, and 3W for a low power replacement fan, and you're looking at 100W even (from the PSU). Add a 92% PSU, and you're looking at a 109W draw from the wall. This gets you 0.96W/GHs from the wall.

Stock for 1.0V is 2.5GHs/chip @ ~313mhz, so you're only OCing by 4%. 325mhz @ 1v should = 188W from the chips. After VRMs, draw goes up to 235. Plus 30W for controller and fan, and draw goes up to 265. On a 90% PSU, draw is 295W from the wall.

325mhz = 166.4GH/s. This gets you 1.77W/GHs from the wall.

*note* Just an estimate. May be higher due to higher voltages on chips closer to the VRMs, or lower due to the VRMs running more efficiently. Fan speed might also save power. Point is there are a lot of variables, so idk. :P

So to be sure then, it looks like I can underclock three S1's at 300mhz and get around 155gh/s each and be fine on say a 850/900w psu?