Topic: Antminer S6 speculation (Read 19270 times)

I will start a new thread on S5 Undervolting, will put a link here when I've done it.

https://bitcointalksearch.org/topic/m.12124622

Rich

I wonder if perhaps the chip cost of the BM1384 has declined enough that maybe it's more practical to use more "downclocked" parts to get an overall improvement. Give how long the BM1384 has been out, you would think the yields at TSMC (or wherever) would be pretty good by now.

The arrival of the S5+ makes me wonder about how likely a "soon to be released" S7 is, presumably with a BM1386.

I think MrTeal reported the 11V reliability floor, and I trust him to know what he's doing. What board version is that? I only have access to some partly-working V1.3 PCBs with six level-shifter diodes, but I've seen some others (not sure what PCB rev) with only five. I haven't checked what the difference is.

Also, I'm really surprised you got 0.25W/GH out of it with the extra fan and controller current.


I would also be surprised to see a BM1384 S7, but given the S4+ and S5+ they don't seem to mind doing things I think are kinda stupid and still turn a profit on it. I figured 3x17 to be a practical max for BM1384 on a S[odd] chassis but that doesn't give you much benefit over the 2x15 string currently in use - about 10% more hashrate at about 20% less power for 70% more chip cost - not like a new-gen chip would. If they're prepping for someone else to put out a new gen in the next month or so it makes sense to get a leg up by moving better gear, but Bitmain has already proven they have no interest in competing with themselves as long as their machines are the most efficient already. Unless someone else steps up with a gamechanger efficiency (rumors for LK's machine are December shipping), they'll continue to push approximately-status-quo miners like the S4+ and S5+ even though they could have been selling better equipment than that six months ago if they wanted to. I'd be surprised to see the S5 and S5+ as the only BM1384 miners released, since they're not even touching the low range of a really good chip like the S2 did for BM1380 and the S4 and S4+ did for BM1382.

 They'd have to go 20-string to even get close to the numbers Lketc has already posted for it's upcomming miner at the miner level. I don't think they have enough space on the board to put 3x20 strings on it, though with the power reduction per chip cooling would be NO issue with either an 18-string or a 20-string.

 
 I don't see any chance the S7 will be BM1384 based, given previous comments out of Bitmain. I can see a VERY FAINT possibility of them putting out a "S5++" with more chips per string, but it would have to show up Real Soon Now and would have to be priced lower per TH than the current S5+ to have a prayer of ROI.

Interesting, folks have previously said that the S5 couldn't be reliably unvervolted past about 11 volts....

Probably not the right thread but a quick comment on Undervolting an S5.

Have just taken delivery of a 2nd hand S5 and after some quick initial checks that everything was ok had a quick go at dropping the volts.

First, very early impression, is that things work as expected...

Tested at 125MHz with 9.0V at the connectors.

(Not wanting to tell people how to suck eggs, but perhaps those that are bragging about how much current they can push through a Molex connector, you need to measure the voltage that is actually getting to the board)

Only measured over a 30 Minute period the Hash was 440GH/s (a bit lower than expected, should be 470GH/s . Over the short time the rate was fluctuating quite a lot and needs further investigation)

Current from the supply (not the wall) was 12.1 Amps so 109 Watts

So that gives 109/440 = J/GH 0.248 

No xxx, HW error a bit higher than I would like to see @ 0.05%

Boots up and starts hashing at 9V, although to be fair it's 10V with no load because of the rubbish leads I am using on the adjustable PSU.

So a lot more tests to run but a better than expected start. Is it my S5, am I lucky, am I doing something wrong? Time and more measurements will tell.

Rich

That's funny, because they sold out of the S4+ and there was no good reason to buy one of those. I wonder how long these S5+ have been run before they were pulled for sale as "new" units? If they're willing to go 48 chips to a blade, it's possible an S7 would still have BM1384 and something like an 18-string would get you right around 0.3W/GH chip-level. Say 3x18 per board runs you about the same hashrate as an S5 and around 385W machine-level.

I haven't tested undervolting S5 boards directly myself. I don't have an adjustable PSU with that kind of current output, and designing one was stalled when we started working on a miner instead. I have run BM1384 chips successfully from 0.58 to 0.80V and the posted efficiencies are about right from what I've seen. Really the only other parts to an S5 board are the inter-node level shifters and per-node LDOs for 0.9V (PLL) and 1.8V (IO) power. The level shifters use a diode drop to translate logic-level signals about 0.8V uphill, which when your node voltages drop to under 0.8 that can affect the received logic levels referenced to local ground, possibly enough that the ASIC might not detect a logic "0" properly. The LDOs also have a min and max input voltage range, and lower-node LDOs are typically powered from core voltage a few nodes uphill. I don't recall the numbers exactly, but possibly if node voltages drop far enough the input voltage for, say, the 1.8V LDO could drop to below the minimum to maintain regulation (probably at least 2.1V) and start starving the chips. The upper few nodes require >12V for their LDOs, so there's a small ~14V boost on the board for them which might also be failing at decreased input voltages for some reason.

I haven't tested either of these things directly to see what's coming into play, but the miner boards we're designing have taken both into account. Our level shifting will use a different method to guarantee signal integrity, and all LDO input voltages will be in-range across the full operating range of the boards.

Think a stacked pair of DPS2K would run an S5+ pretty well?

Thanks!!!

If it's that loud, you have lots more options for power supplies (e.g. noisy and efficient used server supplies).

The sound is 90 Decibell minumum.

I'm guessing Bitmain will sell it at very close to the $US equivilent to what their chinese site quites it for in yuan - and then translate that to bitcoin.

 Given that this S5+ has *6* fans vs. the single on a S5, and guessing they're probably the same fan (the pics aren't clear on that point), it should run about 7-8db louder than a single S5. Possibly a little more as the air path is a lot more crowded than on the S5, but possibly a bit less as it's fully enclosed.

They have to be twice as loud as its predecessor but I don't mind.  My little mining closet keeps the noise down anyway.  However, I'm not buying many more rigs anyway unless I move to Washington State.  I still have not made up my mind as to the move yet.

https://i.imgur.com/HfwPkni.jpg

Sound levels seem a lot higher on the S5+ than any of the previous machines, right?

Totally agree...  The price at iTop is hundreds higher than what it will be at BITMAIN.  I have no idea what BITMAIN will sell it for and when they will sell it.  I need to see if they have started a thread on it.  It's possible I may purchase 25 units if they are not too pricey.

Even if the ITop price includes shipping, that's $hundreds higher than ordering it direct from Bitmain - and IMO the BITMAIN price is too high for this unit.

$3,299 for the S5+ from iTop  http://www.itopshop.net/world-biggest-miner-antminer-s5-bitcoin-miner-77-t-3436w-p-271.html  I'm not sure if that includes shipping.

Will there be coupons for S5+ ?

I don't see any sign of an included power supply.

 Given previous antminers, I don't see any reason you can't use multiple power supplies to power one of these - just don't use different power supplies on one specific board.

 3x15A 120v circuits should work fine with 3 PS at 1100+ watts (that 34xx watt rating is "at the wall", actual DC usage should be around 3100 watts - though I'd go with at least 1200+ per power supply to leave SOME headroom).

 2 of the Bitmain 1600 Watt PS in theory would handle this monster but you'd have to dig up more PCI-E cabling to add to what they include and it would be very marginal and NO room for overclock.
 2 "someone else" standard 1600 Watt PS would probably also work, though the PCI-E connections might be a limiting factor, and not many folks MAKE that big of unit (and they're pri$ey).

It looks like you could use 3 separate 120V power supplies, though they would need to be on 3 separate circuits (i.e. a breakers) to operate safely. Obviously a 240V setup would be much cleaner, although you will likely need more than 1 PSU since even a 2880W server PSU won't do it alone.

http://shop.bitmain.cn/goods.php?id=47

I can't tell from the picture, but would this come with a PSU or we'd need to find one to handle 3,500W? From what I can tell, they are suggesting 3 of these?

https://bitmaintech.com/productDetail.htm?pid=000201505040743496917U7kGsCm0694

I guess it could also technically run on 3x EVGA 1300W G2. However, at 120v, it would be pulling over 30amps. Is 120v even possible for this thing?

Also of note: